diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_0/index_tensor([4241, 4241])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_0/index_tensor([4241, 4241])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0ee1742cc4b7488db09bebb9276af273ddc13c63 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_0/index_tensor([4241, 4241])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, panoramic shot, slowly advancing from a low angle, depicts the heroic stance of the mythical figure Nezha in a distinctively Guocha style. He stands firmly on a pair of rapidly spinning wind-fire wheels, designed as futuristic, suspended pulleys, continuously emitting fiery orange-yellow flame particles and dazzling arcs of light. Nezha himself, sporting his signature black twin buns, bound tightly with red ribbons, possesses a youthful handsomeness, yet his eyes reveal a resolute and hint of rebelliousness. He wears a sleeveless, high-collared red vest, a fusion of modern tailoring and traditional elements. The fabric is intricately woven with auspicious cloud patterns, creating a striking contrast with the black bodysuit underneath. Behind him, the famous Hun Tian Ling (Hun Tian Ling) is no simple fabric; it is a vast ribbon of crimson, translucent silk, its edges ablaze with flowing flames. It swirls and dances majestically and dynamically behind him, like a living being. The entire scene is set against an abstract, dark ink painting of a sea of swirling clouds, creating an epic and mysterious atmosphere. The strong light source comes from the Wind-Fire Wheel and the Hun Tian Ling themselves. The blazing light they emit casts sharp highlights and deep shadows on Nezha's body and clothing, highlighting the lines of his muscles and the texture of his clothes. The whole picture is full of power, rebellious spirit and the unique charm of oriental aesthetics under the strong color contrast of red, gold and black.", + "seed": 33057, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_0/index_4241_seed_33057_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.046373744308948514, + "vlm_logprob": -0.046373744308948514, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.046373744308948514, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0820583924651146 + ], + "opposite_logprob_tokens": [ + -2.5820584297180176 + ], + "logprob_reward": -0.0820583924651146, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0820583924651146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14382223784923553, + 0.041814323514699936, + 0.14968857169151306, + 0.15133725106716156, + 0.11410630494356155, + 0.03778024762868881, + 0.039712127298116684, + 0.17004798352718353, + 0.12759092450141907, + 0.024100016802549362 + ], + "credit_map": [ + 0.015714339911937714, + 0.014887767843902111, + 0.019127849489450455, + 0.019171802327036858, + 0.014249244704842567, + 0.02600577473640442, + 0.016606532037258148, + 0.018058935180306435, + 0.004495661240071058, + 0.007860505022108555, + 0.0048463596031069756, + 0.003577213501557708, + 0.008270896971225739, + 0.0033508087508380413, + 0.00643578264862299, + 0.002977092284709215, + 0.017617009580135345, + 0.029103513807058334, + 0.01242569088935852, + 0.004605064634233713, + 0.05278392508625984, + 0.012661605142056942, + 0.009964938275516033, + 0.01052682101726532, + 0.011363435536623001, + 0.03710581362247467, + 0.04416792467236519, + 0.00852187443524599, + 0.00807785615324974, + 0.013904878869652748, + 0.0045834858901798725, + 0.023611992597579956, + 0.013536873273551464, + 0.0184010099619627, + 0.02204125002026558, + 0.010052189230918884, + 0.011939045041799545, + 0.014137934893369675, + 0.009787099435925484, + 0.014210904948413372, + 0.0033058314584195614, + 0.009643114171922207, + 0.0033867880702018738, + 0.004082854371517897, + 0.0038979395758360624, + 0.004015986807644367, + 0.005746701266616583, + 0.0037010342348366976, + 0.0035301530733704567, + 0.005780900362879038, + 0.005773861426860094, + 0.00440374156460166, + 0.006393734831362963, + 0.00581678282469511, + 0.004936137702316046, + 0.0030768129508942366, + 0.015428283251821995, + 0.03464160114526749, + 0.02525866962969303, + 0.01639309711754322, + 0.01726568676531315, + 0.02390744909644127, + 0.01942247338593006, + 0.017730727791786194, + 0.025553464889526367, + 0.008391891606152058, + 0.016903460025787354, + 0.006954393815249205, + 0.007238045800477266, + 0.030451737344264984, + 0.012512732297182083, + 0.019585199654102325, + 0.0032917973585426807, + 0.004855487961322069, + 0.0022209954913705587, + 0.0033326761331409216, + 0.003785183886066079, + 0.0022543242666870356, + 0.0027464316226541996, + 0.001613121828995645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.0820583924651146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026356296613812447 + ], + "opposite_logprob_tokens": [ + -3.6513562202453613 + ], + "logprob_reward": -0.026356296613812447, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026356296613812447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10426627099514008, + 0.0535711832344532, + 0.06093473359942436, + 0.0837998241186142, + 0.15882036089897156, + 0.04219013825058937, + 0.13177670538425446, + 0.23019199073314667, + 0.11285240948200226, + 0.021596338599920273 + ], + "credit_map": [ + 0.008264990523457527, + 0.008839488960802555, + 0.020685141906142235, + 0.015316703356802464, + 0.008622581139206886, + 0.015576050616800785, + 0.01405832078307867, + 0.012902988120913506, + 0.007560637779533863, + 0.011545829474925995, + 0.005048089660704136, + 0.0042494735680520535, + 0.01226777397096157, + 0.00457110395655036, + 0.0053226519376039505, + 0.003005622187629342, + 0.003590938402339816, + 0.005647679790854454, + 0.017433997243642807, + 0.002923911204561591, + 0.012055537663400173, + 0.003950497135519981, + 0.004050063900649548, + 0.011282108724117279, + 0.007590421009808779, + 0.008882206864655018, + 0.013276828452944756, + 0.01257170457392931, + 0.01153885293751955, + 0.008943800814449787, + 0.010540653020143509, + 0.010455351322889328, + 0.016818402335047722, + 0.01664481870830059, + 0.03021084889769554, + 0.013227062299847603, + 0.018745414912700653, + 0.032961972057819366, + 0.015791339799761772, + 0.014420506544411182, + 0.004623664543032646, + 0.00845950935035944, + 0.004315956495702267, + 0.004590265918523073, + 0.004310940857976675, + 0.004253659397363663, + 0.005752909928560257, + 0.005883229896426201, + 0.010508086532354355, + 0.021279482170939445, + 0.015662608668208122, + 0.01538956817239523, + 0.021006736904382706, + 0.01966039277613163, + 0.017051497474312782, + 0.01121833361685276, + 0.015301291830837727, + 0.05604160204529762, + 0.02126159518957138, + 0.02514931745827198, + 0.025334909558296204, + 0.03399820253252983, + 0.033203739672899246, + 0.01990133710205555, + 0.017848262563347816, + 0.009608483873307705, + 0.01667039841413498, + 0.00811733864247799, + 0.007109790109097958, + 0.025369657203555107, + 0.00914794486016035, + 0.01898052729666233, + 0.002796935848891735, + 0.0028341987635940313, + 0.002245031064376235, + 0.0031635432969778776, + 0.0020746998488903046, + 0.002773888874799013, + 0.0028982176445424557, + 0.0028098246548324823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.026356296613812447 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634591285139322 + ], + "opposite_logprob_tokens": [ + -5.379634380340576 + ], + "logprob_reward": -0.004634591285139322, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634591285139322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18607477843761444, + 0.06341022253036499, + 0.0930120050907135, + 0.0941585823893547, + 0.10926108062267303, + 0.05770186707377434, + 0.055197130888700485, + 0.08480621874332428, + 0.18861353397369385, + 0.0677645355463028 + ], + "credit_map": [ + 0.02311258763074875, + 0.02506568655371666, + 0.02549304999411106, + 0.02099798433482647, + 0.021448375657200813, + 0.030626744031906128, + 0.017201434820890427, + 0.022128935903310776, + 0.007953699678182602, + 0.011022047139704227, + 0.009006796404719353, + 0.004249675199389458, + 0.010390412993729115, + 0.005910217296332121, + 0.008508258499205112, + 0.006369119044393301, + 0.016805803403258324, + 0.013688351958990097, + 0.014667442999780178, + 0.002543114358559251, + 0.026645202189683914, + 0.00676014693453908, + 0.005703149829059839, + 0.00619880435988307, + 0.006715876515954733, + 0.02353551797568798, + 0.026171213015913963, + 0.005051142536103725, + 0.005515117663890123, + 0.009411679580807686, + 0.003577628405764699, + 0.014180420897901058, + 0.015040682628750801, + 0.012827549129724503, + 0.028261778876185417, + 0.008614441379904747, + 0.009871041402220726, + 0.017232010141015053, + 0.008296815678477287, + 0.00911677349358797, + 0.0065393876284360886, + 0.016852915287017822, + 0.0060018207877874374, + 0.005167132243514061, + 0.005704640876501799, + 0.006179836578667164, + 0.005865379702299833, + 0.00539076142013073, + 0.005937670357525349, + 0.011936228722333908, + 0.0057306415401399136, + 0.003899927483871579, + 0.006482915487140417, + 0.008042984642088413, + 0.007717722561210394, + 0.005449048243463039, + 0.007072561886161566, + 0.02216419391334057, + 0.008234198205173016, + 0.007920818403363228, + 0.010252561420202255, + 0.012931175529956818, + 0.01036850456148386, + 0.005862209014594555, + 0.05156802758574486, + 0.012664815410971642, + 0.009253998287022114, + 0.007684474345296621, + 0.014133647084236145, + 0.07679743319749832, + 0.007676097564399242, + 0.008835063315927982, + 0.01012317556887865, + 0.01637626811861992, + 0.0056978086940944195, + 0.006905687507241964, + 0.008067685179412365, + 0.01062486506998539, + 0.005180579610168934, + 0.004788478370755911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.004634591285139322 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005938741844147444 + ], + "opposite_logprob_tokens": [ + -5.130938529968262 + ], + "logprob_reward": -0.005938741844147444, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005938741844147444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18153560161590576, + 0.05014626309275627, + 0.18365569412708282, + 0.09860303997993469, + 0.11909358948469162, + 0.03479115664958954, + 0.04722094535827637, + 0.0942838191986084, + 0.13271251320838928, + 0.057957444339990616 + ], + "credit_map": [ + 0.02052374929189682, + 0.021071702241897583, + 0.01714223623275757, + 0.029113253578543663, + 0.01903708465397358, + 0.02769782394170761, + 0.023490892723202705, + 0.02345884032547474, + 0.006546497344970703, + 0.009156066924333572, + 0.006047124043107033, + 0.0034125405363738537, + 0.009333615191280842, + 0.0051350705325603485, + 0.006230445578694344, + 0.004284899681806564, + 0.035694073885679245, + 0.01760096289217472, + 0.053047921508550644, + 0.005060160998255014, + 0.030722064897418022, + 0.009644168429076672, + 0.012967291288077831, + 0.01891903206706047, + 0.008088981732726097, + 0.023196451365947723, + 0.030884386971592903, + 0.004643524065613747, + 0.004574309568852186, + 0.008106940425932407, + 0.0023400604259222746, + 0.016768377274274826, + 0.018875062465667725, + 0.012500220909714699, + 0.03355817124247551, + 0.008669957518577576, + 0.01033070869743824, + 0.01549731008708477, + 0.009051777422428131, + 0.010610365308821201, + 0.003885606536641717, + 0.007758140563964844, + 0.0032845877576619387, + 0.004212734755128622, + 0.0034630941227078438, + 0.004103796556591988, + 0.005085351876914501, + 0.0029978386592119932, + 0.004594968631863594, + 0.011820601299405098, + 0.007866282016038895, + 0.0036121499724686146, + 0.004753203596919775, + 0.006654602941125631, + 0.0037870232481509447, + 0.004132108762860298, + 0.007200950291007757, + 0.021831925958395004, + 0.009045504033565521, + 0.009865849278867245, + 0.010701624676585197, + 0.01376840565353632, + 0.012613506056368351, + 0.00925604347139597, + 0.036914534866809845, + 0.008954261429607868, + 0.005901784636080265, + 0.0048673758283257484, + 0.010182052850723267, + 0.05539551004767418, + 0.005034610629081726, + 0.005462386645376682, + 0.006462363991886377, + 0.009793134406208992, + 0.004357711412012577, + 0.007347512058913708, + 0.009549695067107677, + 0.005511745344847441, + 0.009972583502531052, + 0.004962692968547344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.005938741844147444 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11288069933652878 + ], + "opposite_logprob_tokens": [ + -2.2378807067871094 + ], + "logprob_reward": -0.11288069933652878, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11288069933652878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10196503251791, + 0.03715672343969345, + 0.09908351302146912, + 0.15718482434749603, + 0.03855596110224724, + 0.07307404279708862, + 0.07698526233434677, + 0.15088212490081787, + 0.22008618712425232, + 0.04502633213996887 + ], + "credit_map": [ + 0.011914211325347424, + 0.010247286409139633, + 0.018826253712177277, + 0.013693797402083874, + 0.009960752911865711, + 0.011242506094276905, + 0.012582644820213318, + 0.01349758729338646, + 0.004315823782235384, + 0.0061890981160104275, + 0.004500618204474449, + 0.0032766994554549456, + 0.005822287872433662, + 0.0032214955426752567, + 0.0062392172403633595, + 0.003591483226045966, + 0.016221104189753532, + 0.004219759255647659, + 0.04043279215693474, + 0.0037677879445254803, + 0.00755208870396018, + 0.0030687309335917234, + 0.007695374544709921, + 0.016125869005918503, + 0.011739591136574745, + 0.04023675620555878, + 0.047342266887426376, + 0.006725713144987822, + 0.007701302878558636, + 0.015150602906942368, + 0.0031818060670048, + 0.02510679140686989, + 0.004871855489909649, + 0.00761796860024333, + 0.002969961380586028, + 0.003772686468437314, + 0.0051245493814349174, + 0.007054044399410486, + 0.003872733097523451, + 0.003272158792242408, + 0.006670031230896711, + 0.014560182578861713, + 0.006418944802135229, + 0.00687385955825448, + 0.007454623002558947, + 0.010844632983207703, + 0.011049135588109493, + 0.009202629327774048, + 0.005838845390826464, + 0.013546004891395569, + 0.006882868707180023, + 0.009355907328426838, + 0.012141565792262554, + 0.012074598111212254, + 0.011898827739059925, + 0.005246644373983145, + 0.00997432041913271, + 0.02995588816702366, + 0.014717892743647099, + 0.01693161390721798, + 0.016912033781409264, + 0.021767135709524155, + 0.02634689211845398, + 0.014276345260441303, + 0.04808267578482628, + 0.013060897588729858, + 0.022811198607087135, + 0.010066132061183453, + 0.013894101604819298, + 0.07503297924995422, + 0.011282374151051044, + 0.02585582435131073, + 0.005720599554479122, + 0.003653109073638916, + 0.009652303531765938, + 0.004796198569238186, + 0.003532764967530966, + 0.005472476128488779, + 0.005221896339207888, + 0.00697698350995779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.11288069933652878 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, panoramic shot, slowly advancing from a low angle, depicts the heroic stance of the mythical figure Nezha in a distinctively Guocha style. He stands firmly on a pair of rapidly spinning wind-fire wheels, designed as futuristic, suspended pulleys, continuously emitting fiery orange-yellow flame particles and dazzling arcs of light. Nezha himself, sporting his signature black twin buns, bound tightly with red ribbons, possesses a youthful handsomeness, yet his eyes reveal a resolute and hint of rebelliousness. He wears a sleeveless, high-collared red vest, a fusion of modern tailoring and traditional elements. The fabric is intricately woven with auspicious cloud patterns, creating a striking contrast with the black bodysuit underneath. Behind him, the famous Hun Tian Ling (Hun Tian Ling) is no simple fabric; it is a vast ribbon of crimson, translucent silk, its edges ablaze with flowing flames. It swirls and dances majestically and dynamically behind him, like a living being. The entire scene is set against an abstract, dark ink painting of a sea of swirling clouds, creating an epic and mysterious atmosphere. The strong light source comes from the Wind-Fire Wheel and the Hun Tian Ling themselves. The blazing light they emit casts sharp highlights and deep shadows on Nezha's body and clothing, highlighting the lines of his muscles and the texture of his clothes. The whole picture is full of power, rebellious spirit and the unique charm of oriental aesthetics under the strong color contrast of red, gold and black.", + "seed": 33058, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_0/index_4241_seed_33058_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04975638054311275, + "vlm_logprob": -0.04975638054311275, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.04975638054311275, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0368766151368618 + ], + "opposite_logprob_tokens": [ + -3.411876678466797 + ], + "logprob_reward": -0.0368766151368618, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0368766151368618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08419110625982285, + 0.012393967248499393, + 0.018255719915032387, + 0.020125215873122215, + 0.0158443134278059, + 0.017635319381952286, + 0.13469499349594116, + 0.6464729309082031, + 0.04000252112746239, + 0.010383944027125835 + ], + "credit_map": [ + 0.006961964536458254, + 0.005988705903291702, + 0.013393662869930267, + 0.00754882674664259, + 0.02576454170048237, + 0.010526453144848347, + 0.006962112616747618, + 0.007044837810099125, + 0.0010470804991200566, + 0.0017568337498232722, + 0.0015129526145756245, + 0.0014289225218817592, + 0.001288846367970109, + 0.0009570613037794828, + 0.002301769098266959, + 0.002100500511005521, + 0.0017920853570103645, + 0.0018610353581607342, + 0.0028920855838805437, + 0.0027729917783290148, + 0.001823130645789206, + 0.0016613256884738803, + 0.0028628024738281965, + 0.002590260934084654, + 0.002595948288217187, + 0.0021381021942943335, + 0.0021247747354209423, + 0.0030966540798544884, + 0.004139726981520653, + 0.0015921933809295297, + 0.0022568334825336933, + 0.0021809819154441357, + 0.0020999317057430744, + 0.0024894061498343945, + 0.0014883499825373292, + 0.0015896148979663849, + 0.0033741602674126625, + 0.0013194188941270113, + 0.002163342433050275, + 0.0013200882822275162, + 0.0025105997920036316, + 0.0020140372216701508, + 0.0018274291651323438, + 0.002949413610622287, + 0.0031217560172080994, + 0.0011543573345988989, + 0.0026625930331647396, + 0.0013951316941529512, + 0.0335625559091568, + 0.015737079083919525, + 0.01047060638666153, + 0.05281220003962517, + 0.004756891168653965, + 0.006168670020997524, + 0.006897283717989922, + 0.004289685282856226, + 0.10710831731557846, + 0.1639830321073532, + 0.12987524271011353, + 0.13002131879329681, + 0.021574025973677635, + 0.05395492538809776, + 0.02199348621070385, + 0.017962558194994926, + 0.010311750695109367, + 0.004293627105653286, + 0.005587155465036631, + 0.00463923579081893, + 0.003968021832406521, + 0.005849138367921114, + 0.003570671658962965, + 0.0017829224234446883, + 0.0012673917226493359, + 0.001838455442339182, + 0.001327717793174088, + 0.00137935986276716, + 0.001336552668362856, + 0.0014718921156600118, + 0.0009189166012220085, + 0.0008436582866124809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0368766151368618 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820322498679161 + ], + "opposite_logprob_tokens": [ + -4.018203258514404 + ], + "logprob_reward": -0.01820322498679161, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820322498679161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0353245735168457, + 0.02672288939356804, + 0.042886536568403244, + 0.011403382755815983, + 0.009769442491233349, + 0.01359870657324791, + 0.1849893480539322, + 0.5087823867797852, + 0.15650799870491028, + 0.010014825500547886 + ], + "credit_map": [ + 0.00484780129045248, + 0.0033107451163232327, + 0.004464277531951666, + 0.0034524269867688417, + 0.007476511877030134, + 0.00303417956456542, + 0.00538055831566453, + 0.003358071204274893, + 0.002392359310761094, + 0.003396890824660659, + 0.0033367632422596216, + 0.0038299320731312037, + 0.003054650267586112, + 0.0018156504957005382, + 0.003761796047911048, + 0.0051348451524972916, + 0.0035346008371561766, + 0.005956059321761131, + 0.005974080879241228, + 0.0040801349096000195, + 0.006142169237136841, + 0.004182248841971159, + 0.004423929378390312, + 0.008593308739364147, + 0.0014531242195516825, + 0.0016339236171916127, + 0.0017102425917983055, + 0.0012530400417745113, + 0.0018346465658396482, + 0.0010957109043374658, + 0.0011152367806062102, + 0.00130745698697865, + 0.0012723129475489259, + 0.0010378627339378, + 0.001044944510795176, + 0.0009538233280181885, + 0.001656586304306984, + 0.0012194085866212845, + 0.0015922883758321404, + 0.0009922145400196314, + 0.0017483303090557456, + 0.0016973537858575583, + 0.002116571879014373, + 0.001929473364725709, + 0.0018395422957837582, + 0.0012031085789203644, + 0.0017208709614351392, + 0.0013434536522254348, + 0.04812157526612282, + 0.02080438658595085, + 0.013332769274711609, + 0.07276175916194916, + 0.00656817015260458, + 0.00961950421333313, + 0.009062282741069794, + 0.004718886688351631, + 0.0832754373550415, + 0.12451817095279694, + 0.10520590841770172, + 0.10377413779497147, + 0.016819177195429802, + 0.04287943243980408, + 0.017890997231006622, + 0.01441907323896885, + 0.05518705025315285, + 0.01062373910099268, + 0.01292623020708561, + 0.007076257839798927, + 0.008898722007870674, + 0.05215572193264961, + 0.004956996068358421, + 0.004683257546275854, + 0.0014984209556132555, + 0.0011718227760866284, + 0.0016540774377062917, + 0.0018256611656397581, + 0.0009708970319479704, + 0.0012613661820068955, + 0.0009092369582504034, + 0.0007233417709358037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.01820322498679161 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095619544386864 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095619544386864, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095619544386864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0938446894288063, + 0.035608500242233276, + 0.03913920000195503, + 0.052308108657598495, + 0.043771032243967056, + 0.052017927169799805, + 0.181364968419075, + 0.4264800548553467, + 0.04584629833698273, + 0.029619209468364716 + ], + "credit_map": [ + 0.014663002453744411, + 0.00767692644149065, + 0.01173603255301714, + 0.008110509254038334, + 0.023008933290839195, + 0.009337294846773148, + 0.012594320811331272, + 0.006717672571539879, + 0.003441436681896448, + 0.007019962649792433, + 0.005599910859018564, + 0.004729376640170813, + 0.0035433215089142323, + 0.00227543362416327, + 0.004487217869609594, + 0.004511839710175991, + 0.005031112115830183, + 0.005616362206637859, + 0.004881658125668764, + 0.00492903683334589, + 0.005797874182462692, + 0.002587717492133379, + 0.006152319721877575, + 0.00414311746135354, + 0.008332416415214539, + 0.004103993531316519, + 0.00888136588037014, + 0.009244043380022049, + 0.0048073274083435535, + 0.002496726345270872, + 0.0083067761734128, + 0.006135459523648024, + 0.00661289319396019, + 0.005023580975830555, + 0.006074643228203058, + 0.005326863378286362, + 0.0057703847996890545, + 0.004585414193570614, + 0.0049700550734996796, + 0.0054071987979114056, + 0.008389905095100403, + 0.00504844868555665, + 0.006059633567929268, + 0.007171389181166887, + 0.0072251372039318085, + 0.004454102832823992, + 0.007457544095814228, + 0.0062117669731378555, + 0.04605236276984215, + 0.02122250758111477, + 0.011794143356382847, + 0.0700206309556961, + 0.006573168095201254, + 0.010924290865659714, + 0.010018494911491871, + 0.004759365692734718, + 0.08320589363574982, + 0.10730607062578201, + 0.08995502442121506, + 0.07697232067584991, + 0.015220853500068188, + 0.030971579253673553, + 0.012292101047933102, + 0.010556189343333244, + 0.010885667987167835, + 0.00668107345700264, + 0.007376017514616251, + 0.00550285167992115, + 0.004121736623346806, + 0.005465516820549965, + 0.003253738395869732, + 0.002559695392847061, + 0.004421144723892212, + 0.004471689462661743, + 0.0036430892068892717, + 0.004207502119243145, + 0.004580159205943346, + 0.003619002876803279, + 0.0024356734938919544, + 0.0022409483790397644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.004095619544386864 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977255403995514 + ], + "opposite_logprob_tokens": [ + -4.634772777557373 + ], + "logprob_reward": -0.00977255403995514, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977255403995514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1581503003835678, + 0.011880324222147465, + 0.014430036768317223, + 0.012907887808978558, + 0.019524844363331795, + 0.01208910159766674, + 0.05907339230179787, + 0.6450430154800415, + 0.05320437252521515, + 0.013696720823645592 + ], + "credit_map": [ + 0.0348844937980175, + 0.018279653042554855, + 0.020013941451907158, + 0.013436700217425823, + 0.03157726675271988, + 0.013985786586999893, + 0.01214547548443079, + 0.013826987706124783, + 0.0015050543006509542, + 0.0023579131811857224, + 0.0014593200758099556, + 0.0013278304832056165, + 0.0011473086196929216, + 0.000800577865447849, + 0.0016236796509474516, + 0.001658639986999333, + 0.0013581931125372648, + 0.002241385169327259, + 0.0017857099883258343, + 0.0016507789259776473, + 0.0025681862607598305, + 0.000945598294492811, + 0.002142550190910697, + 0.0017376337200403214, + 0.002521769842132926, + 0.0015896353870630264, + 0.0019661495462059975, + 0.001476910663768649, + 0.0013275962555781007, + 0.0010182877304032445, + 0.0014485262800008059, + 0.0015590116381645203, + 0.0031038979068398476, + 0.00239601731300354, + 0.002383157843723893, + 0.001812756061553955, + 0.0031666054856032133, + 0.0033820022363215685, + 0.0016222421545535326, + 0.0016581632662564516, + 0.001768016954883933, + 0.0022651255130767822, + 0.0014779444318264723, + 0.001478311838582158, + 0.0012561050243675709, + 0.001287185586988926, + 0.0015405721496790648, + 0.001015840214677155, + 0.014218267053365707, + 0.007575268391519785, + 0.004606060683727264, + 0.021474746987223625, + 0.0030252994038164616, + 0.0025954986922442913, + 0.0033831761684268713, + 0.002195074688643217, + 0.11500796675682068, + 0.19792421162128448, + 0.10519611835479736, + 0.11767716705799103, + 0.0234328992664814, + 0.05069240182638168, + 0.018830759450793266, + 0.016281472519040108, + 0.017193879932165146, + 0.004374122247099876, + 0.005297697614878416, + 0.004141718149185181, + 0.004017564468085766, + 0.01369347982108593, + 0.0026293827686458826, + 0.0018565257778391242, + 0.0020559991244226694, + 0.002142982091754675, + 0.0018548867665231228, + 0.0018818171229213476, + 0.002009637886658311, + 0.001600171672180295, + 0.0010216974187642336, + 0.001129529089666903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.00977255403995514 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17983388900756836 + ], + "opposite_logprob_tokens": [ + -1.8048338890075684 + ], + "logprob_reward": -0.17983388900756836, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17983388900756836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16797137260437012, + 0.013219598680734634, + 0.01959371194243431, + 0.014645596966147423, + 0.014462798833847046, + 0.014186346903443336, + 0.16896337270736694, + 0.45567741990089417, + 0.11933638900518417, + 0.011943405494093895 + ], + "credit_map": [ + 0.015863966196775436, + 0.012919193133711815, + 0.02372857928276062, + 0.01282251626253128, + 0.044982753694057465, + 0.018434863537549973, + 0.023890770971775055, + 0.01532873697578907, + 0.0013946661492809653, + 0.002487662946805358, + 0.001809638924896717, + 0.0016206934815272689, + 0.0012940027518197894, + 0.0008995764655992389, + 0.0017106676241382957, + 0.0020026916172355413, + 0.0021415711380541325, + 0.0014462884282693267, + 0.002975482027977705, + 0.00275718467310071, + 0.0021126337815076113, + 0.0015380181139335036, + 0.00305786426179111, + 0.0035646718461066484, + 0.0012483905302360654, + 0.0012158275349065661, + 0.0038490411825478077, + 0.0013025051448494196, + 0.0019700611010193825, + 0.0011940182885155082, + 0.001766811590641737, + 0.0020989426411688328, + 0.002432771259918809, + 0.0022308228071779013, + 0.0016070263227447867, + 0.0024323652032762766, + 0.0015789995668455958, + 0.001265074242837727, + 0.0016646283911541104, + 0.0012511127861216664, + 0.002087129047140479, + 0.00155920151155442, + 0.0026390524581074715, + 0.0016652151243761182, + 0.0018896887777373195, + 0.0012276885099709034, + 0.0015436841640621424, + 0.00157468905672431, + 0.04338342696428299, + 0.019063428044319153, + 0.01216899324208498, + 0.06705296784639359, + 0.005871315021067858, + 0.008549178019165993, + 0.008470144122838974, + 0.00440393527969718, + 0.06916584074497223, + 0.13605333864688873, + 0.07337585091590881, + 0.09171342104673386, + 0.01581851951777935, + 0.04123478755354881, + 0.015583471395075321, + 0.012732184492051601, + 0.04256383329629898, + 0.007290221750736237, + 0.009112849831581116, + 0.004553340841084719, + 0.006613065488636494, + 0.041911154985427856, + 0.0036336469929665327, + 0.003658283269032836, + 0.0016784982290118933, + 0.001692080870270729, + 0.0019651572220027447, + 0.0016826778883114457, + 0.001362755661830306, + 0.001684076152741909, + 0.0009683024836704135, + 0.0009098590817302465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.17983388900756836 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, panoramic shot, slowly advancing from a low angle, depicts the heroic stance of the mythical figure Nezha in a distinctively Guocha style. He stands firmly on a pair of rapidly spinning wind-fire wheels, designed as futuristic, suspended pulleys, continuously emitting fiery orange-yellow flame particles and dazzling arcs of light. Nezha himself, sporting his signature black twin buns, bound tightly with red ribbons, possesses a youthful handsomeness, yet his eyes reveal a resolute and hint of rebelliousness. He wears a sleeveless, high-collared red vest, a fusion of modern tailoring and traditional elements. The fabric is intricately woven with auspicious cloud patterns, creating a striking contrast with the black bodysuit underneath. Behind him, the famous Hun Tian Ling (Hun Tian Ling) is no simple fabric; it is a vast ribbon of crimson, translucent silk, its edges ablaze with flowing flames. It swirls and dances majestically and dynamically behind him, like a living being. The entire scene is set against an abstract, dark ink painting of a sea of swirling clouds, creating an epic and mysterious atmosphere. The strong light source comes from the Wind-Fire Wheel and the Hun Tian Ling themselves. The blazing light they emit casts sharp highlights and deep shadows on Nezha's body and clothing, highlighting the lines of his muscles and the texture of his clothes. The whole picture is full of power, rebellious spirit and the unique charm of oriental aesthetics under the strong color contrast of red, gold and black.", + "seed": 33059, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_0/index_4241_seed_33059_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02818070473149419, + "vlm_logprob": -0.02818070473149419, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.02818070473149419, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024189457297325134 + ], + "opposite_logprob_tokens": [ + -3.8991894721984863 + ], + "logprob_reward": -0.024189457297325134, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024189457297325134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16736729443073273, + 0.05356159433722496, + 0.2211732268333435, + 0.08429030328989029, + 0.04741629585623741, + 0.09618601948022842, + 0.07595015317201614, + 0.11877045780420303, + 0.09713304787874222, + 0.03815166652202606 + ], + "credit_map": [ + 0.035035427659749985, + 0.016937578096985817, + 0.013895682990550995, + 0.015676740556955338, + 0.015470857731997967, + 0.022922106087207794, + 0.019206760451197624, + 0.028222136199474335, + 0.004642336163669825, + 0.00979792233556509, + 0.00393653754144907, + 0.0063620698638260365, + 0.01170967798680067, + 0.005597899202257395, + 0.006676769815385342, + 0.004838378634303808, + 0.024324703961610794, + 0.01509370282292366, + 0.034352246671915054, + 0.017962321639060974, + 0.017698803916573524, + 0.056547775864601135, + 0.02594780921936035, + 0.029245847836136818, + 0.016991859301924706, + 0.00941670686006546, + 0.010566938668489456, + 0.005756922531872988, + 0.01153379026800394, + 0.009682734496891499, + 0.007508751470595598, + 0.012832592241466045, + 0.007318669930100441, + 0.0030105519108474255, + 0.01149339322000742, + 0.005360885988920927, + 0.006183256395161152, + 0.003313230350613594, + 0.007025578524917364, + 0.003710726508870721, + 0.01999872364103794, + 0.009140958078205585, + 0.02543465420603752, + 0.007122775074094534, + 0.008367293514311314, + 0.005321856588125229, + 0.012436659075319767, + 0.008363086730241776, + 0.011122044175863266, + 0.007619176991283894, + 0.006976438220590353, + 0.007438020780682564, + 0.006934338249266148, + 0.017967043444514275, + 0.009288454428315163, + 0.008604631759226322, + 0.017214398831129074, + 0.02683064341545105, + 0.009022645652294159, + 0.011692238040268421, + 0.01391309779137373, + 0.010813391767442226, + 0.01326392125338316, + 0.01602010242640972, + 0.014412486925721169, + 0.012522643432021141, + 0.008432302623987198, + 0.01346513256430626, + 0.011745025403797626, + 0.014556030742824078, + 0.014580441638827324, + 0.007418976631015539, + 0.0052820597775280476, + 0.0049867951311171055, + 0.0034064275678247213, + 0.004289758391678333, + 0.00472077913582325, + 0.005563084036111832, + 0.005532784387469292, + 0.004369976930320263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.024189457297325134 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059456892311573 + ], + "opposite_logprob_tokens": [ + -3.895594596862793 + ], + "logprob_reward": -0.02059456892311573, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059456892311573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12378286570310593, + 0.09812112897634506, + 0.29623886942863464, + 0.11458883434534073, + 0.05959554761648178, + 0.04158260300755501, + 0.04162779077887535, + 0.0838884562253952, + 0.058766722679138184, + 0.08180712908506393 + ], + "credit_map": [ + 0.01486410666257143, + 0.012652370147407055, + 0.009661933407187462, + 0.014487138949334621, + 0.014560054056346416, + 0.024252239614725113, + 0.01501229964196682, + 0.01829271763563156, + 0.0069592478685081005, + 0.012070685625076294, + 0.011718470603227615, + 0.013623565435409546, + 0.015594253316521645, + 0.011272515170276165, + 0.015429232269525528, + 0.011453160084784031, + 0.030547326430678368, + 0.021155275404453278, + 0.050820428878068924, + 0.02460174635052681, + 0.024732127785682678, + 0.06999629735946655, + 0.03432990238070488, + 0.04005575552582741, + 0.02063964307308197, + 0.01154917199164629, + 0.013096786104142666, + 0.009693465195596218, + 0.013509185053408146, + 0.01933901011943817, + 0.012242759577929974, + 0.014518807642161846, + 0.01210668496787548, + 0.004518718458712101, + 0.018465271219611168, + 0.0033258332405239344, + 0.005800347309559584, + 0.006157831288874149, + 0.004080329556018114, + 0.005140527617186308, + 0.005064785480499268, + 0.004953321069478989, + 0.007088515441864729, + 0.004188632592558861, + 0.004027087241411209, + 0.004060703329741955, + 0.005743078887462616, + 0.006456479895859957, + 0.006308145355433226, + 0.003341624978929758, + 0.00408942112699151, + 0.0038727675564587116, + 0.003853584174066782, + 0.009595661424100399, + 0.004475470166653395, + 0.006091116461902857, + 0.012153638526797295, + 0.01725820265710354, + 0.006673257332295179, + 0.008109278045594692, + 0.009465738199651241, + 0.009173782542347908, + 0.009989183396100998, + 0.01106537226587534, + 0.00928787887096405, + 0.007801208179444075, + 0.005685129202902317, + 0.007416802458465099, + 0.006472249981015921, + 0.005077947862446308, + 0.01221065130084753, + 0.004814856220036745, + 0.011889555491507053, + 0.007925535552203655, + 0.0076614548452198505, + 0.008928971365094185, + 0.008283732458949089, + 0.01096796989440918, + 0.01427857019007206, + 0.011871338821947575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.02059456892311573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002488137688487768 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.002488137688487768, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002488137688487768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08451410382986069, + 0.06619333475828171, + 0.29059362411499023, + 0.0730937197804451, + 0.057125307619571686, + 0.06956645846366882, + 0.06721305102109909, + 0.07768325507640839, + 0.046467848122119904, + 0.16754929721355438 + ], + "credit_map": [ + 0.011155395768582821, + 0.009312440641224384, + 0.0051777116023004055, + 0.010140097700059414, + 0.0076397922821342945, + 0.014576979912817478, + 0.010827715508639812, + 0.01568397879600525, + 0.005465561058372259, + 0.006942160427570343, + 0.012801505625247955, + 0.006850678473711014, + 0.011560595594346523, + 0.010408309288322926, + 0.006555761676281691, + 0.005608763080090284, + 0.03048192523419857, + 0.02015005797147751, + 0.05031287670135498, + 0.02275758981704712, + 0.025416705757379532, + 0.07156705111265182, + 0.03149024397134781, + 0.03841717541217804, + 0.010090101510286331, + 0.008161263540387154, + 0.009718119166791439, + 0.00669848034158349, + 0.01046737190335989, + 0.009490475989878178, + 0.008933557197451591, + 0.009534354321658611, + 0.007290216628462076, + 0.004397082142531872, + 0.016759086400270462, + 0.0040514362044632435, + 0.004754170775413513, + 0.009925948455929756, + 0.004692531190812588, + 0.005254834890365601, + 0.01281612366437912, + 0.008807851001620293, + 0.013797665946185589, + 0.008770320564508438, + 0.005946498829871416, + 0.004442228004336357, + 0.00764104351401329, + 0.007344732992351055, + 0.010116122663021088, + 0.009719867259263992, + 0.010167721658945084, + 0.006484023295342922, + 0.005818672012537718, + 0.010105335153639317, + 0.007863879203796387, + 0.0069374339655041695, + 0.010564432479441166, + 0.010341730900108814, + 0.008595469407737255, + 0.010677296668291092, + 0.009836548008024693, + 0.006720542907714844, + 0.012506076134741306, + 0.00844116136431694, + 0.006762878969311714, + 0.0046714916825294495, + 0.004401606973260641, + 0.007330678403377533, + 0.004503005184233189, + 0.004567115567624569, + 0.009935295209288597, + 0.004295778460800648, + 0.02476460114121437, + 0.023953087627887726, + 0.023364925757050514, + 0.019447438418865204, + 0.024699266999959946, + 0.011346799321472645, + 0.025972776114940643, + 0.014000398106873035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.002488137688487768 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046312687918543816 + ], + "opposite_logprob_tokens": [ + -5.379631042480469 + ], + "logprob_reward": -0.0046312687918543816, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046312687918543816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1459868848323822, + 0.052920784801244736, + 0.0565778948366642, + 0.11989657580852509, + 0.19332849979400635, + 0.13802526891231537, + 0.05576804280281067, + 0.0679694265127182, + 0.08024981617927551, + 0.0892767384648323 + ], + "credit_map": [ + 0.03537733107805252, + 0.01039696205407381, + 0.00781935639679432, + 0.015957500785589218, + 0.011617185547947884, + 0.019272327423095703, + 0.013143462128937244, + 0.03240276128053665, + 0.004386303015053272, + 0.006171948742121458, + 0.006462574936449528, + 0.00665722880512476, + 0.009341618046164513, + 0.006845258176326752, + 0.006179123651236296, + 0.006876729894429445, + 0.006976209580898285, + 0.006684154737740755, + 0.00524183502420783, + 0.005358940921723843, + 0.011609215289354324, + 0.00944489985704422, + 0.0064963288605213165, + 0.004766312427818775, + 0.027162428945302963, + 0.012228641659021378, + 0.01128335203975439, + 0.007717316970229149, + 0.011998462490737438, + 0.021397031843662262, + 0.011376525275409222, + 0.016732819378376007, + 0.032688502222299576, + 0.01611238904297352, + 0.039233945310115814, + 0.008823518641293049, + 0.015305223874747753, + 0.03757508471608162, + 0.01099055539816618, + 0.03259928151965141, + 0.0236175749450922, + 0.011886032298207283, + 0.027340229600667953, + 0.009028821252286434, + 0.011829298920929432, + 0.024627450853586197, + 0.01836630329489708, + 0.011329560540616512, + 0.008789008483290672, + 0.0038435484748333693, + 0.005148791242390871, + 0.0046848151832818985, + 0.005410767626017332, + 0.013581853359937668, + 0.007038540672510862, + 0.007270719390362501, + 0.009852341376245022, + 0.014537142589688301, + 0.006512415129691362, + 0.009326476603746414, + 0.007644099649041891, + 0.0064214421436190605, + 0.007494788616895676, + 0.006180722266435623, + 0.011339640244841576, + 0.00838928110897541, + 0.00842976476997137, + 0.010838016867637634, + 0.008762875571846962, + 0.005714087747037411, + 0.01772272400557995, + 0.009053424000740051, + 0.011519074440002441, + 0.010941856540739536, + 0.008522771298885345, + 0.013504822738468647, + 0.009116769768297672, + 0.013204369693994522, + 0.014256177470088005, + 0.008210903964936733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.0046312687918543816 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900009095668793 + ], + "opposite_logprob_tokens": [ + -2.4639999866485596 + ], + "logprob_reward": -0.08900009095668793, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900009095668793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3158445656299591, + 0.08802393823862076, + 0.2105865627527237, + 0.07939919829368591, + 0.033157914876937866, + 0.08309299498796463, + 0.06146252527832985, + 0.053405988961458206, + 0.030190983787178993, + 0.044835351407527924 + ], + "credit_map": [ + 0.04735066741704941, + 0.030718928202986717, + 0.0309295691549778, + 0.03529069572687149, + 0.026356371119618416, + 0.06186230480670929, + 0.043417058885097504, + 0.03991895541548729, + 0.007680513430386782, + 0.008615557104349136, + 0.014468447305262089, + 0.011991825886070728, + 0.008285033516585827, + 0.009779422543942928, + 0.013831710442900658, + 0.013371425680816174, + 0.019688909873366356, + 0.013104510493576527, + 0.039398085325956345, + 0.01548114512115717, + 0.019433792680501938, + 0.055343057960271835, + 0.022392787039279938, + 0.025744283571839333, + 0.014489095658063889, + 0.008919847197830677, + 0.01159457303583622, + 0.007694999221712351, + 0.008658846840262413, + 0.011578643694519997, + 0.007234919350594282, + 0.009228276088833809, + 0.003400129033252597, + 0.002529427409172058, + 0.005785115994513035, + 0.0031390460208058357, + 0.004142018500715494, + 0.005058168899267912, + 0.004504587035626173, + 0.004599422682076693, + 0.008057166822254658, + 0.012409945018589497, + 0.009142773225903511, + 0.01075378991663456, + 0.010431593284010887, + 0.005796900484710932, + 0.02056683786213398, + 0.005933994892984629, + 0.009613716043531895, + 0.006221567280590534, + 0.006422224920243025, + 0.00498280068859458, + 0.005167066119611263, + 0.015166152268648148, + 0.006506455596536398, + 0.007382543757557869, + 0.007657090201973915, + 0.01137803215533495, + 0.0036571426317095757, + 0.0052340272814035416, + 0.006542561575770378, + 0.004430965520441532, + 0.007140009663999081, + 0.007366159465163946, + 0.004544702358543873, + 0.0044164215214550495, + 0.0031292710918933153, + 0.0037744154687970877, + 0.0031986902467906475, + 0.0024455487728118896, + 0.0055950721725821495, + 0.003086862852796912, + 0.006641200743615627, + 0.005183754954487085, + 0.004574498627334833, + 0.005209615454077721, + 0.0057359361089766026, + 0.00659880181774497, + 0.007487917318940163, + 0.00340362754650414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.08900009095668793 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, panoramic shot, slowly advancing from a low angle, depicts the heroic stance of the mythical figure Nezha in a distinctively Guocha style. He stands firmly on a pair of rapidly spinning wind-fire wheels, designed as futuristic, suspended pulleys, continuously emitting fiery orange-yellow flame particles and dazzling arcs of light. Nezha himself, sporting his signature black twin buns, bound tightly with red ribbons, possesses a youthful handsomeness, yet his eyes reveal a resolute and hint of rebelliousness. He wears a sleeveless, high-collared red vest, a fusion of modern tailoring and traditional elements. The fabric is intricately woven with auspicious cloud patterns, creating a striking contrast with the black bodysuit underneath. Behind him, the famous Hun Tian Ling (Hun Tian Ling) is no simple fabric; it is a vast ribbon of crimson, translucent silk, its edges ablaze with flowing flames. It swirls and dances majestically and dynamically behind him, like a living being. The entire scene is set against an abstract, dark ink painting of a sea of swirling clouds, creating an epic and mysterious atmosphere. The strong light source comes from the Wind-Fire Wheel and the Hun Tian Ling themselves. The blazing light they emit casts sharp highlights and deep shadows on Nezha's body and clothing, highlighting the lines of his muscles and the texture of his clothes. The whole picture is full of power, rebellious spirit and the unique charm of oriental aesthetics under the strong color contrast of red, gold and black.", + "seed": 33060, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_0/index_4241_seed_33060_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.030288004875183107, + "vlm_logprob": -0.030288004875183107, + "vlm_yes_no_margin": 3.9999999523162844, + "vlm_token_credit": -0.030288004875183107, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019700223580002785 + ], + "opposite_logprob_tokens": [ + -4.019700050354004 + ], + "logprob_reward": -0.019700223580002785, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.019700223580002785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0703483372926712, + 0.35626763105392456, + 0.09374142438173294, + 0.06604021042585373, + 0.07146269083023071, + 0.05385688692331314, + 0.17520304024219513, + 0.049333035945892334, + 0.019084859639406204, + 0.04466190189123154 + ], + "credit_map": [ + 0.007156133186072111, + 0.00495207728818059, + 0.010097162798047066, + 0.009369831532239914, + 0.01066240668296814, + 0.017391471192240715, + 0.006104130297899246, + 0.0046151261776685715, + 0.10564322769641876, + 0.016675205901265144, + 0.03485232964158058, + 0.009250583127140999, + 0.01382362749427557, + 0.1422540545463562, + 0.014994879253208637, + 0.018773727118968964, + 0.01732589304447174, + 0.006667801178991795, + 0.01085332129150629, + 0.011828732676804066, + 0.0058546666987240314, + 0.024394314736127853, + 0.0073202187195420265, + 0.009496474638581276, + 0.007635058369487524, + 0.008270381018519402, + 0.008115874603390694, + 0.009731368161737919, + 0.010526956059038639, + 0.006958923768252134, + 0.007123298943042755, + 0.007678355555981398, + 0.007116040214896202, + 0.010809549130499363, + 0.007841182872653008, + 0.010462827049195766, + 0.01256558671593666, + 0.007016540504992008, + 0.007798489183187485, + 0.007852469570934772, + 0.006151740904897451, + 0.008656317368149757, + 0.004929659888148308, + 0.00459269480779767, + 0.0073925103060901165, + 0.008515335619449615, + 0.009986796416342258, + 0.003631834639236331, + 0.03601309657096863, + 0.018456222489476204, + 0.0070067816413939, + 0.005227142944931984, + 0.07468748092651367, + 0.010164273902773857, + 0.01606334000825882, + 0.007584687322378159, + 0.00486370176076889, + 0.005484719295054674, + 0.0046472689136862755, + 0.006323919631540775, + 0.009780284017324448, + 0.006050190422683954, + 0.006930242292582989, + 0.005252712406218052, + 0.0017664004117250443, + 0.001608052640222013, + 0.0030691996216773987, + 0.001573975314386189, + 0.003832888323813677, + 0.0018480689032003284, + 0.0016506166430190206, + 0.0037356577813625336, + 0.0037932114209979773, + 0.005001402460038662, + 0.003889161627739668, + 0.004180118907243013, + 0.008425050415098667, + 0.005880584008991718, + 0.005853460170328617, + 0.007638911250978708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.019700223580002785 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204584993422031 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204584993422031, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204584993422031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13356952369213104, + 0.32200074195861816, + 0.06765430420637131, + 0.08872703462839127, + 0.09591662138700485, + 0.05658088997006416, + 0.11287242919206619, + 0.05082203447818756, + 0.0206568855792284, + 0.0511995404958725 + ], + "credit_map": [ + 0.01114108320325613, + 0.010849157348275185, + 0.020735736936330795, + 0.01878354325890541, + 0.02227877825498581, + 0.026085801422595978, + 0.01344404648989439, + 0.010251360014081001, + 0.09399823844432831, + 0.014876089058816433, + 0.031017564237117767, + 0.0074736326932907104, + 0.012689916417002678, + 0.13127011060714722, + 0.014027779921889305, + 0.016647404059767723, + 0.015479902736842632, + 0.004637767095118761, + 0.007972891442477703, + 0.007212089840322733, + 0.0038372199051082134, + 0.019813016057014465, + 0.003726831404492259, + 0.004974582698196173, + 0.010172756388783455, + 0.010350664146244526, + 0.008290060795843601, + 0.016535544767975807, + 0.01107272133231163, + 0.011983469128608704, + 0.010839122347533703, + 0.009482696652412415, + 0.010428056120872498, + 0.015065438114106655, + 0.008216270245611668, + 0.010834816843271255, + 0.02022193744778633, + 0.011109493672847748, + 0.010332345962524414, + 0.009708256460726261, + 0.007320047356188297, + 0.008461576886475086, + 0.004682768136262894, + 0.006392870098352432, + 0.01252683438360691, + 0.006699657533317804, + 0.005386357195675373, + 0.0051107825711369514, + 0.026632320135831833, + 0.010828518308699131, + 0.004997794516384602, + 0.005512680392712355, + 0.03645382449030876, + 0.01238578837364912, + 0.010044907219707966, + 0.006016588769853115, + 0.0061804745346307755, + 0.0044482070952653885, + 0.008579796180129051, + 0.00619581900537014, + 0.006665044464170933, + 0.005628350656479597, + 0.005814404226839542, + 0.007309936452656984, + 0.0018145347712561488, + 0.002641198690980673, + 0.0032304376363754272, + 0.001623362535610795, + 0.0031284396536648273, + 0.001914525288157165, + 0.0020547525491565466, + 0.004249634686857462, + 0.0063970633782446384, + 0.008529280312359333, + 0.0034722520504146814, + 0.003658658592030406, + 0.01101449504494667, + 0.006814152002334595, + 0.0076087103225290775, + 0.0037049315869808197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.014204584993422031 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633904159069061 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.008633904159069061, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633904159069061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21623572707176208, + 0.189778670668602, + 0.055999260395765305, + 0.0972704067826271, + 0.06203630566596985, + 0.06199923902750015, + 0.06826778501272202, + 0.021134816110134125, + 0.03983163833618164, + 0.18744611740112305 + ], + "credit_map": [ + 0.03360292688012123, + 0.009896006435155869, + 0.03143981099128723, + 0.010755332186818123, + 0.018771331757307053, + 0.08418165147304535, + 0.012149510905146599, + 0.015439162962138653, + 0.05492972210049629, + 0.008343563415110111, + 0.019108539447188377, + 0.0063324267975986, + 0.007835989817976952, + 0.07491772621870041, + 0.008317350409924984, + 0.009993347339332104, + 0.00661499984562397, + 0.0059152343310415745, + 0.005857530981302261, + 0.012468439526855946, + 0.004225874785333872, + 0.006318965461105108, + 0.006053197663277388, + 0.008545021526515484, + 0.012216855771839619, + 0.012173186056315899, + 0.012239230796694756, + 0.009812602773308754, + 0.019394943490624428, + 0.011994394473731518, + 0.011659031733870506, + 0.007780156563967466, + 0.009433427825570107, + 0.00918445736169815, + 0.005653103347867727, + 0.0061455643735826015, + 0.008088632486760616, + 0.010848733596503735, + 0.007405831478536129, + 0.005276556126773357, + 0.008185564540326595, + 0.008876924403011799, + 0.004136241972446442, + 0.004486257676035166, + 0.016022348776459694, + 0.006957737263292074, + 0.009026151150465012, + 0.0043080151081085205, + 0.011554323136806488, + 0.007933284156024456, + 0.004204096272587776, + 0.006061356049031019, + 0.013318894430994987, + 0.011970343999564648, + 0.006269712932407856, + 0.006955776829272509, + 0.002118645003065467, + 0.00250805402174592, + 0.0024748858995735645, + 0.0029200727585703135, + 0.004122249316424131, + 0.0019752683583647013, + 0.0030061036814004183, + 0.0020095380023121834, + 0.0032253616955131292, + 0.005727788433432579, + 0.004294435493648052, + 0.00284358998760581, + 0.009264166466891766, + 0.00348351476714015, + 0.0033497661352157593, + 0.007643013261258602, + 0.011767555959522724, + 0.029841572046279907, + 0.013826911337673664, + 0.011556846089661121, + 0.05474367365241051, + 0.02374764159321785, + 0.029104894027113914, + 0.012857018038630486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.008633904159069061 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008629058487713337 + ], + "opposite_logprob_tokens": [ + -4.758628845214844 + ], + "logprob_reward": -0.008629058487713337, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008629058487713337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17413565516471863, + 0.35034656524658203, + 0.05898654833436012, + 0.10408266633749008, + 0.07612764090299606, + 0.03539995849132538, + 0.09658496826887131, + 0.04264771193265915, + 0.02855166792869568, + 0.033136650919914246 + ], + "credit_map": [ + 0.020851589739322662, + 0.014942134730517864, + 0.027830008417367935, + 0.026957856491208076, + 0.01746387407183647, + 0.03624160960316658, + 0.016297398135066032, + 0.013551191426813602, + 0.10196506977081299, + 0.01773759350180626, + 0.03363633528351784, + 0.00940584484487772, + 0.01559645775705576, + 0.139382004737854, + 0.01512859296053648, + 0.01749465987086296, + 0.007963605225086212, + 0.00479495944455266, + 0.0052849967032670975, + 0.013147537596523762, + 0.004635579418390989, + 0.008471069857478142, + 0.00677235284820199, + 0.007916446775197983, + 0.01308075524866581, + 0.013458147644996643, + 0.012170912697911263, + 0.012534571811556816, + 0.017222359776496887, + 0.014790606684982777, + 0.011872692033648491, + 0.00895262323319912, + 0.01301612425595522, + 0.007942317984998226, + 0.0106129115447402, + 0.008321121335029602, + 0.008569875732064247, + 0.015311789698898792, + 0.005902563221752644, + 0.0064509352669119835, + 0.004886138252913952, + 0.005457127932459116, + 0.002950417110696435, + 0.0035252345260232687, + 0.008722657337784767, + 0.0030464481096714735, + 0.004173098597675562, + 0.002638838719576597, + 0.021765343844890594, + 0.009184598922729492, + 0.005222962703555822, + 0.003924666438251734, + 0.0349852554500103, + 0.008219759911298752, + 0.009174306876957417, + 0.004108066204935312, + 0.004333716817200184, + 0.005424934905022383, + 0.0039030907209962606, + 0.005294933915138245, + 0.007700950838625431, + 0.005335164722055197, + 0.005373660940676928, + 0.005281258374452591, + 0.002994372509419918, + 0.0019586270209401846, + 0.003213153686374426, + 0.001380489207804203, + 0.00765324616804719, + 0.002530534751713276, + 0.002533163409680128, + 0.006288081407546997, + 0.0038352725096046925, + 0.0038851704448461533, + 0.0037061949260532856, + 0.002622931730002165, + 0.004236770328134298, + 0.006963708437979221, + 0.003973032347857952, + 0.00391356972977519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.008629058487713337 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the mythical figure Nezha standing firmly on wind-fire wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the wind-fire wheels depicted as futuristic, suspended pulleys emitting fiery orange-yellow flame particles and light arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does Nezha have black twin buns bound with red ribbons and wear a sleeveless high-collared red vest with cloud patterns over a black bodysuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a vast crimson translucent silk ribbon (Hun Tian Ling) behind Nezha that swirls and dances with flowing flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot panoramic and slowly advancing from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10027225315570831 + ], + "opposite_logprob_tokens": [ + -2.3502721786499023 + ], + "logprob_reward": -0.10027225315570831, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10027225315570831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.301042377948761, + 0.06779912114143372, + 0.04474721848964691, + 0.07937373220920563, + 0.11656969785690308, + 0.04258213937282562, + 0.06818472594022751, + 0.03969717025756836, + 0.09232979267835617, + 0.14767399430274963 + ], + "credit_map": [ + 0.03937814384698868, + 0.016458280384540558, + 0.048504993319511414, + 0.013850214891135693, + 0.03309253975749016, + 0.11511196941137314, + 0.014645341783761978, + 0.020000897347927094, + 0.008901088498532772, + 0.006811731029301882, + 0.009434737265110016, + 0.008572486229240894, + 0.006246316246688366, + 0.011523211374878883, + 0.006675974931567907, + 0.009633575566112995, + 0.006837079767137766, + 0.007698024623095989, + 0.00657434668391943, + 0.005832741968333721, + 0.004800106398761272, + 0.005058405455201864, + 0.0037317576352506876, + 0.004214760381728411, + 0.0146108353510499, + 0.00932862889021635, + 0.007842672988772392, + 0.0056361183524131775, + 0.009175468236207962, + 0.020748566836118698, + 0.007202396634966135, + 0.004829047713428736, + 0.014252746477723122, + 0.014684687368571758, + 0.012108552269637585, + 0.013755968771874905, + 0.01885388232767582, + 0.020290644839406013, + 0.012715675868093967, + 0.00990754459053278, + 0.005344041623175144, + 0.00743001326918602, + 0.0035903265234082937, + 0.0058248513378202915, + 0.008022844791412354, + 0.0033721798099577427, + 0.005922494921833277, + 0.0030753901228308678, + 0.011697868816554546, + 0.007715276908129454, + 0.004159961361438036, + 0.005217034835368395, + 0.009253785014152527, + 0.0174943245947361, + 0.007373809348791838, + 0.005272669717669487, + 0.005286972038447857, + 0.007042944431304932, + 0.005505972541868687, + 0.004448136780411005, + 0.004911619704216719, + 0.003634912194684148, + 0.003930958453565836, + 0.004935659468173981, + 0.005868466105312109, + 0.005850194022059441, + 0.009951334446668625, + 0.004773029591888189, + 0.025641079992055893, + 0.009716746397316456, + 0.007175134029239416, + 0.0233538169413805, + 0.018084798008203506, + 0.015802152454853058, + 0.019240448251366615, + 0.013290776871144772, + 0.022499985992908478, + 0.017571359872817993, + 0.015716014429926872, + 0.02546847052872181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.10027225315570831 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_1/index_tensor([4676, 4676])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_1/index_tensor([4676, 4676])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bcf09bb238b6cb94bf6ac82d11b0776c2e736cf0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_1/index_tensor([4676, 4676])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up still shot captured from a slightly elevated angle observes a cluster of pristine, deep purple grapes nestled against a flawless, matte white surface. Soft, diffused studio lighting sculpts their plump, spherical forms from the side, casting gentle shadows and creating bright, moist highlights across their taut skin, coated in delicate, powdery blossoms. As the grapes begin to melt, a surreal liquefaction begins; the process begins subtly, with the skin of one grape losing its tension and dissolving as if made of sugar. This transformation triggers a chain reaction as neighboring grapes steadily and deliberately collapse, their solid form deliquescing into a viscous liquid. Rather than cracking, the skins melt, their dark color seeping into a growing pool of dark, rich juice that slowly spreads across the pristine white surface. Its smooth, syrupy surface reflects light, creating a hypnotic, unsettling, and beautifully stark visual effect.", + "seed": 29979, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_1/index_4676_seed_29979_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13437043242156504, + "vlm_logprob": -0.13437043242156504, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.13437043242156504, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009578933008015156 + ], + "opposite_logprob_tokens": [ + -5.634578704833984 + ], + "logprob_reward": -0.009578933008015156, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.009578933008015156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14736667275428772, + 0.21764665842056274, + 0.06367465853691101, + 0.05782819539308548, + 0.07779520750045776, + 0.055455781519412994, + 0.07150524109601974, + 0.15351350605487823, + 0.04896112158894539, + 0.10625304281711578 + ], + "credit_map": [ + 0.01834416203200817, + 0.015832416713237762, + 0.018307426944375038, + 0.014505128376185894, + 0.016492029651999474, + 0.017862673848867416, + 0.02529633603990078, + 0.02072647400200367, + 0.018138021230697632, + 0.01945372484624386, + 0.023673901334404945, + 0.03332791104912758, + 0.0411975271999836, + 0.026819821447134018, + 0.03038148768246174, + 0.02465425617992878, + 0.007711826823651791, + 0.009205215610563755, + 0.008841194212436676, + 0.007611199747771025, + 0.006547896657139063, + 0.008503040298819542, + 0.007916279137134552, + 0.0073379999957978725, + 0.0053804004564881325, + 0.005003861617296934, + 0.00604618014767766, + 0.005292033776640892, + 0.012546740472316742, + 0.010220983996987343, + 0.005975721403956413, + 0.007362272124737501, + 0.012597791850566864, + 0.008896689862012863, + 0.014668800868093967, + 0.011658529751002789, + 0.006084296386688948, + 0.007325074635446072, + 0.009909997694194317, + 0.006654016673564911, + 0.008317478932440281, + 0.007667208556085825, + 0.009270777925848961, + 0.006790648214519024, + 0.006535263732075691, + 0.0044169798493385315, + 0.007573200389742851, + 0.004884219262748957, + 0.011313732713460922, + 0.00590995978564024, + 0.008925704285502434, + 0.007420610170811415, + 0.006351661402732134, + 0.005729380063712597, + 0.01196915004402399, + 0.013885028660297394, + 0.0263384860008955, + 0.009335880167782307, + 0.01408076286315918, + 0.03597583249211311, + 0.00595871452242136, + 0.007618842180818319, + 0.01120708417147398, + 0.04299789294600487, + 0.004646151792258024, + 0.005019669886678457, + 0.008388660848140717, + 0.010012920014560223, + 0.005309863481670618, + 0.004011059179902077, + 0.004748194478452206, + 0.006824593059718609, + 0.02018294297158718, + 0.011070238426327705, + 0.014205517247319221, + 0.019601354375481606, + 0.009610338136553764, + 0.00808643363416195, + 0.008997543714940548, + 0.014498654752969742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.009578933008015156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.632796049118042 + ], + "opposite_logprob_tokens": [ + -0.757796049118042 + ], + "logprob_reward": -0.632796049118042, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.632796049118042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2134491205215454, + 0.11039178818464279, + 0.06590697169303894, + 0.09385189414024353, + 0.06299236416816711, + 0.07458660006523132, + 0.06617791950702667, + 0.12358957529067993, + 0.043105483055114746, + 0.14594833552837372 + ], + "credit_map": [ + 0.01775461621582508, + 0.01762552373111248, + 0.019355367869138718, + 0.017580417916178703, + 0.014901511371135712, + 0.03871640935540199, + 0.052049361169338226, + 0.03546590730547905, + 0.00971191842108965, + 0.007555149495601654, + 0.012964719906449318, + 0.017090309411287308, + 0.016108352690935135, + 0.014845414087176323, + 0.017017731443047523, + 0.015098193660378456, + 0.005664055701345205, + 0.004869540221989155, + 0.009780863299965858, + 0.006065488327294588, + 0.00522213289514184, + 0.015420790761709213, + 0.007422524504363537, + 0.01146158017218113, + 0.011796006932854652, + 0.006755411624908447, + 0.010183330625295639, + 0.009641183540225029, + 0.012834898196160793, + 0.01923547312617302, + 0.009981175884604454, + 0.013424413278698921, + 0.00992271676659584, + 0.005269446410238743, + 0.017813529819250107, + 0.008791462518274784, + 0.004126708954572678, + 0.004858504049479961, + 0.006513284984976053, + 0.005696708802133799, + 0.008143994957208633, + 0.006304569076746702, + 0.011913498863577843, + 0.014288602396845818, + 0.00761432433500886, + 0.005887024104595184, + 0.007941560819745064, + 0.012493024580180645, + 0.010234195739030838, + 0.0058323051780462265, + 0.008423353545367718, + 0.009675594978034496, + 0.007387405261397362, + 0.005109803751111031, + 0.00805157795548439, + 0.01146367471665144, + 0.03670136258006096, + 0.0070251841098070145, + 0.011854234151542187, + 0.03615499287843704, + 0.004494762979447842, + 0.005457632709294558, + 0.009835973381996155, + 0.012065432965755463, + 0.005612113047391176, + 0.004178977105766535, + 0.006746498867869377, + 0.0064346445724368095, + 0.003968331962823868, + 0.004219996277242899, + 0.006071116309612989, + 0.005873802118003368, + 0.014754016883671284, + 0.01671687327325344, + 0.021544814109802246, + 0.019980914890766144, + 0.01885276287794113, + 0.016441799700260162, + 0.01820053905248642, + 0.019456619396805763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.632796049118042 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782706387341022 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.009782706387341022, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782706387341022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17142753303050995, + 0.1697079837322235, + 0.10625159740447998, + 0.1200001910328865, + 0.05192355439066887, + 0.06893615424633026, + 0.07743962854146957, + 0.0862683430314064, + 0.042051732540130615, + 0.10599331557750702 + ], + "credit_map": [ + 0.01657143048942089, + 0.02707197144627571, + 0.024697188287973404, + 0.024701256304979324, + 0.01809065043926239, + 0.020550629124045372, + 0.020108655095100403, + 0.019635746255517006, + 0.011672823689877987, + 0.01108936220407486, + 0.019654391333460808, + 0.023705560714006424, + 0.025588639080524445, + 0.024822818115353584, + 0.032010480761528015, + 0.021163905039429665, + 0.012640858069062233, + 0.00833959225565195, + 0.01480165496468544, + 0.01254374347627163, + 0.009563145227730274, + 0.016419867053627968, + 0.016556784510612488, + 0.015385949984192848, + 0.013103412464261055, + 0.0088358698412776, + 0.016101377084851265, + 0.011405912227928638, + 0.020931674167513847, + 0.020324409008026123, + 0.013336838223040104, + 0.015960700809955597, + 0.006292861886322498, + 0.0039793034084141254, + 0.013809717260301113, + 0.009285531006753445, + 0.0031120122876018286, + 0.0037241512909531593, + 0.00783494208008051, + 0.003885036101564765, + 0.007423002272844315, + 0.007114256266504526, + 0.0145471366122365, + 0.009977038949728012, + 0.0063021560199558735, + 0.005549410358071327, + 0.009341488592326641, + 0.008681661449372768, + 0.011040568351745605, + 0.006545928306877613, + 0.008969338610768318, + 0.023447737097740173, + 0.005216940771788359, + 0.00492393784224987, + 0.009061330929398537, + 0.008233840577304363, + 0.01602044515311718, + 0.005066602025181055, + 0.009540101513266563, + 0.024533269926905632, + 0.005709833465516567, + 0.0034855706617236137, + 0.014060141518712044, + 0.00785237830132246, + 0.004406442400068045, + 0.004063426982611418, + 0.005225961096584797, + 0.011832108721137047, + 0.0035358602181077003, + 0.0027512311935424805, + 0.004758549854159355, + 0.005478151608258486, + 0.01211453601717949, + 0.011533454060554504, + 0.01894620805978775, + 0.01118119154125452, + 0.015616119839251041, + 0.010887522250413895, + 0.014485741034150124, + 0.011228551156818867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.009782706387341022 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036184806376695633 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036184806376695633, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036184806376695633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15577028691768646, + 0.08665210753679276, + 0.0675126388669014, + 0.23573656380176544, + 0.03276761993765831, + 0.04954272508621216, + 0.07977642118930817, + 0.11119727045297623, + 0.0367637500166893, + 0.14428061246871948 + ], + "credit_map": [ + 0.014678836800158024, + 0.02098201774060726, + 0.02428135648369789, + 0.02513069286942482, + 0.02004586160182953, + 0.015295092016458511, + 0.018347645178437233, + 0.017008785158395767, + 0.007282438687980175, + 0.005219659302383661, + 0.00811910629272461, + 0.010739515535533428, + 0.008224702440202236, + 0.016250859946012497, + 0.019249921664595604, + 0.01156590785831213, + 0.0054312944412231445, + 0.0048825680278241634, + 0.01092272438108921, + 0.007179571781307459, + 0.004922195337712765, + 0.01325221173465252, + 0.010631082579493523, + 0.010290988720953465, + 0.022729938849806786, + 0.017135431990027428, + 0.030042126774787903, + 0.02145111933350563, + 0.04233372583985329, + 0.04465627670288086, + 0.02443602867424488, + 0.032951921224594116, + 0.004791672807186842, + 0.003047990845516324, + 0.005359424743801355, + 0.005572592839598656, + 0.0026056799106299877, + 0.0025183947291225195, + 0.005866977386176586, + 0.0030048841144889593, + 0.004778033588081598, + 0.004710610490292311, + 0.006381724961102009, + 0.011036472395062447, + 0.004377071745693684, + 0.003462859196588397, + 0.005471824202686548, + 0.009324129670858383, + 0.007875991985201836, + 0.007427350617945194, + 0.009035313501954079, + 0.015709808096289635, + 0.009323357604444027, + 0.005616235546767712, + 0.01170552708208561, + 0.013082842342555523, + 0.01110164262354374, + 0.007375741843134165, + 0.011571738868951797, + 0.0437801256775856, + 0.004001274239271879, + 0.003546980209648609, + 0.01366098877042532, + 0.016158781945705414, + 0.004018167499452829, + 0.004093122202903032, + 0.005220300517976284, + 0.004215771798044443, + 0.0041601755656301975, + 0.0037558660842478275, + 0.00603778101503849, + 0.005262566264718771, + 0.017027249559760094, + 0.017674194648861885, + 0.024289151653647423, + 0.01792346127331257, + 0.019949566572904587, + 0.014186174608767033, + 0.018322434276342392, + 0.01490836776793003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.0036184806376695633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075992956757545 + ], + "opposite_logprob_tokens": [ + -4.14107608795166 + ], + "logprob_reward": -0.016075992956757545, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075992956757545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14821140468120575, + 0.13632787764072418, + 0.09867233037948608, + 0.05961693823337555, + 0.0640256479382515, + 0.07717806100845337, + 0.09501653164625168, + 0.18401958048343658, + 0.026490468531847, + 0.11044112592935562 + ], + "credit_map": [ + 0.012311703525483608, + 0.014514008536934853, + 0.022091802209615707, + 0.016355838626623154, + 0.0165889710187912, + 0.021002165973186493, + 0.028336184099316597, + 0.017010733485221863, + 0.01136860717087984, + 0.008141709491610527, + 0.018167566508054733, + 0.01991448365151882, + 0.014608540572226048, + 0.01803869754076004, + 0.02946333773434162, + 0.0166249368339777, + 0.007971676997840405, + 0.0050072032026946545, + 0.008887330070137978, + 0.007450580131262541, + 0.006513367872685194, + 0.027384119108319283, + 0.017020151019096375, + 0.01843789964914322, + 0.0065694344229996204, + 0.005611686035990715, + 0.007202776614576578, + 0.007137599401175976, + 0.010058562271296978, + 0.00936143845319748, + 0.006282921880483627, + 0.007392518687993288, + 0.011744935996830463, + 0.005156778264790773, + 0.017538655549287796, + 0.009128398261964321, + 0.004597037564963102, + 0.005242879968136549, + 0.004808146506547928, + 0.0058088116347789764, + 0.005273325834423304, + 0.0047589014284312725, + 0.008738420903682709, + 0.02255556732416153, + 0.007996887899935246, + 0.003844167571514845, + 0.006218897178769112, + 0.01779189147055149, + 0.008609546348452568, + 0.006385746877640486, + 0.010710733011364937, + 0.030295880511403084, + 0.006389797665178776, + 0.004735234659165144, + 0.01541497278958559, + 0.012474620714783669, + 0.014470333233475685, + 0.012228093110024929, + 0.017504099756479263, + 0.07188473641872406, + 0.006802725605666637, + 0.007074886001646519, + 0.022363487631082535, + 0.03169121965765953, + 0.0030526830814778805, + 0.0027991696260869503, + 0.004344376735389233, + 0.0033564886543899775, + 0.0037080063484609127, + 0.002517363056540489, + 0.0037995739839971066, + 0.0029128072783350945, + 0.01320887915790081, + 0.014688712544739246, + 0.017220256850123405, + 0.015852000564336777, + 0.011831942945718765, + 0.010209377855062485, + 0.013196454383432865, + 0.014233493246138096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.016075992956757545 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up still shot captured from a slightly elevated angle observes a cluster of pristine, deep purple grapes nestled against a flawless, matte white surface. Soft, diffused studio lighting sculpts their plump, spherical forms from the side, casting gentle shadows and creating bright, moist highlights across their taut skin, coated in delicate, powdery blossoms. As the grapes begin to melt, a surreal liquefaction begins; the process begins subtly, with the skin of one grape losing its tension and dissolving as if made of sugar. This transformation triggers a chain reaction as neighboring grapes steadily and deliberately collapse, their solid form deliquescing into a viscous liquid. Rather than cracking, the skins melt, their dark color seeping into a growing pool of dark, rich juice that slowly spreads across the pristine white surface. Its smooth, syrupy surface reflects light, creating a hypnotic, unsettling, and beautifully stark visual effect.", + "seed": 29980, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_1/index_4676_seed_29980_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.4356075465679169, + "vlm_logprob": -0.4356075465679169, + "vlm_yes_no_margin": 2.524999976158142, + "vlm_token_credit": -0.4356075465679169, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01191004179418087 + ], + "opposite_logprob_tokens": [ + -5.1369099617004395 + ], + "logprob_reward": -0.01191004179418087, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01191004179418087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2861485183238983, + 0.07687042653560638, + 0.14874812960624695, + 0.09533881396055222, + 0.11154723167419434, + 0.08262067288160324, + 0.04618072137236595, + 0.07467523962259293, + 0.04624786600470543, + 0.031622353941202164 + ], + "credit_map": [ + 0.026137156412005424, + 0.03581120818853378, + 0.03137948364019394, + 0.05604676157236099, + 0.031550224870443344, + 0.03567266836762428, + 0.04507647827267647, + 0.024474531412124634, + 0.012808365747332573, + 0.00890738982707262, + 0.019622983410954475, + 0.00810268335044384, + 0.004793698899447918, + 0.009590847417712212, + 0.007490083575248718, + 0.005554381292313337, + 0.012798694893717766, + 0.007695237640291452, + 0.010824082419276237, + 0.009838399477303028, + 0.009304750710725784, + 0.015031050890684128, + 0.04432038217782974, + 0.038935523480176926, + 0.0089948121458292, + 0.00607488164678216, + 0.024615677073597908, + 0.006524283904582262, + 0.008794468827545643, + 0.005152841564267874, + 0.010355046950280666, + 0.024826806038618088, + 0.007163792848587036, + 0.009089332073926926, + 0.027529368177056313, + 0.011743173003196716, + 0.014532938599586487, + 0.013816443271934986, + 0.016445105895400047, + 0.011227077804505825, + 0.007343619130551815, + 0.013395520858466625, + 0.009304968640208244, + 0.011278380639851093, + 0.006331342272460461, + 0.00900033675134182, + 0.012474204413592815, + 0.013492302969098091, + 0.004357080440968275, + 0.002469788771122694, + 0.0025404144544154406, + 0.014150779694318771, + 0.003938371315598488, + 0.0029739798046648502, + 0.00598190538585186, + 0.009768402203917503, + 0.006501507479697466, + 0.010362265631556511, + 0.007938982918858528, + 0.012451773509383202, + 0.004088950343430042, + 0.005319481249898672, + 0.01694931462407112, + 0.011062964797019958, + 0.005368038546293974, + 0.007863357663154602, + 0.004207838326692581, + 0.010023408569395542, + 0.0035460959188640118, + 0.00399712985381484, + 0.004463304299861193, + 0.006778692360967398, + 0.0042498232796788216, + 0.005764600820839405, + 0.002984663238748908, + 0.005071151070296764, + 0.002357601188123226, + 0.0036287058610469103, + 0.0031734276562929153, + 0.004392382223159075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.01191004179418087 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.0178098678588867 + ], + "opposite_logprob_tokens": [ + -0.14280974864959717 + ], + "logprob_reward": -2.0178098678588867, + "yes_no_margin_reward": -1.8750001192092896, + "token_credit_reward": -2.0178098678588867, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13237343728542328, + 0.07626628130674362, + 0.0775156170129776, + 0.05376580357551575, + 0.043008286505937576, + 0.174510657787323, + 0.06530522555112839, + 0.2779558598995209, + 0.05642956122756004, + 0.04286924749612808 + ], + "credit_map": [ + 0.010336062870919704, + 0.018316322937607765, + 0.013432118110358715, + 0.02807530388236046, + 0.015869563445448875, + 0.016820229589939117, + 0.01802418753504753, + 0.011499647982418537, + 0.017111415043473244, + 0.006579900160431862, + 0.01686943508684635, + 0.006834458094090223, + 0.004435692448168993, + 0.008330529555678368, + 0.009992901235818863, + 0.00611194409430027, + 0.005859582219272852, + 0.0037119020707905293, + 0.006032506003975868, + 0.004317690618336201, + 0.004736010916531086, + 0.00734672648832202, + 0.028295012190937996, + 0.017216181382536888, + 0.005754645448178053, + 0.0039910669438540936, + 0.005334494169801474, + 0.0040570334531366825, + 0.014887386001646519, + 0.004460787400603294, + 0.004271386656910181, + 0.01100900024175644, + 0.004392373841255903, + 0.0035046006087213755, + 0.007785115856677294, + 0.005535165313631296, + 0.004750657361000776, + 0.005247225984930992, + 0.006964833475649357, + 0.004828310571610928, + 0.018343588337302208, + 0.024111809208989143, + 0.021300718188285828, + 0.02077740989625454, + 0.015128502622246742, + 0.01778537966310978, + 0.03378552943468094, + 0.023277711123228073, + 0.006210471037775278, + 0.0025878455489873886, + 0.003288398962467909, + 0.0074297357350587845, + 0.005313715897500515, + 0.004316403064876795, + 0.027506079524755478, + 0.008652562275528908, + 0.021027466282248497, + 0.02565009891986847, + 0.027250636368989944, + 0.059611059725284576, + 0.02281675487756729, + 0.02408403344452381, + 0.032561395317316055, + 0.06495441496372223, + 0.008242101408541203, + 0.005916138645261526, + 0.003884917590767145, + 0.01477514673024416, + 0.006372677627950907, + 0.005526689346879721, + 0.0059605613350868225, + 0.005751322954893112, + 0.0033819111995399, + 0.004133470822125673, + 0.0026389844715595245, + 0.011338871903717518, + 0.0026366959791630507, + 0.002719527343288064, + 0.003100639209151268, + 0.012919142842292786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -2.0178098678588867 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901136577129364 + ], + "opposite_logprob_tokens": [ + -2.5790114402770996 + ], + "logprob_reward": -0.07901136577129364, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901136577129364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18265485763549805, + 0.07402199506759644, + 0.05830689147114754, + 0.09195812046527863, + 0.19024468958377838, + 0.24312463402748108, + 0.026595216244459152, + 0.06914473325014114, + 0.036863092333078384, + 0.027085764333605766 + ], + "credit_map": [ + 0.013627131469547749, + 0.019626667723059654, + 0.02034791000187397, + 0.04511801898479462, + 0.016539303585886955, + 0.018909357488155365, + 0.03260021284222603, + 0.015886258333921432, + 0.011339720338582993, + 0.007264019455760717, + 0.018583791330456734, + 0.007552294060587883, + 0.005221331492066383, + 0.006704119965434074, + 0.011422866955399513, + 0.005933844950050116, + 0.005737281870096922, + 0.005060043651610613, + 0.007131364196538925, + 0.005385705269873142, + 0.009338149800896645, + 0.0046472325921058655, + 0.008430269546806812, + 0.012576844543218613, + 0.008159163407981396, + 0.003781272331252694, + 0.03289560228586197, + 0.0062175109051167965, + 0.013915321789681911, + 0.004687018692493439, + 0.009743322618305683, + 0.012558912858366966, + 0.012113507837057114, + 0.015442597679793835, + 0.045953962951898575, + 0.019256705418229103, + 0.023600643500685692, + 0.02520037442445755, + 0.02949557825922966, + 0.01918131113052368, + 0.02200852707028389, + 0.034988947212696075, + 0.033546701073646545, + 0.034635767340660095, + 0.020426345989108086, + 0.02250712551176548, + 0.04697170481085777, + 0.028039531782269478, + 0.0020977300591766834, + 0.002308470197021961, + 0.0025550087448209524, + 0.008775466121733189, + 0.0019487184472382069, + 0.002111404901370406, + 0.0048055630177259445, + 0.00199285545386374, + 0.00586552731692791, + 0.0075640068389475346, + 0.007982180453836918, + 0.011442926712334156, + 0.005129605531692505, + 0.005271086934953928, + 0.016045289114117622, + 0.009844106622040272, + 0.003409241558983922, + 0.0032985974103212357, + 0.002529214834794402, + 0.012162844650447369, + 0.0021396137308329344, + 0.001702119829133153, + 0.005408666096627712, + 0.006212793756276369, + 0.0039838762022554874, + 0.004397028591483831, + 0.003074180567637086, + 0.004010642878711224, + 0.002383460756391287, + 0.002368918852880597, + 0.0025753274094313383, + 0.004292327910661697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.07901136577129364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602626726031303 + ], + "opposite_logprob_tokens": [ + -3.8956027030944824 + ], + "logprob_reward": -0.020602626726031303, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602626726031303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13190869987010956, + 0.07284164428710938, + 0.07896099984645844, + 0.10679923743009567, + 0.08648635447025299, + 0.16175313293933868, + 0.039396170526742935, + 0.14753437042236328, + 0.06612466275691986, + 0.1081947311758995 + ], + "credit_map": [ + 0.011159015819430351, + 0.01381978765130043, + 0.015283831395208836, + 0.03020416758954525, + 0.012856967747211456, + 0.013577002100646496, + 0.02340617962181568, + 0.011601755395531654, + 0.011397331953048706, + 0.006022538989782333, + 0.02276194840669632, + 0.006878715939819813, + 0.004855589009821415, + 0.008049819618463516, + 0.006757169961929321, + 0.0061185345984995365, + 0.005583661608397961, + 0.004467789549380541, + 0.00932720024138689, + 0.00666563306003809, + 0.011289408430457115, + 0.006633547600358725, + 0.027236776426434517, + 0.0077569931745529175, + 0.018628159537911415, + 0.009313848800957203, + 0.0160745307803154, + 0.0092451898381114, + 0.016118129715323448, + 0.010651177726686, + 0.01285748090595007, + 0.013910714536905289, + 0.0053507923148572445, + 0.007884991355240345, + 0.02098359540104866, + 0.008508063852787018, + 0.011426572687923908, + 0.011196666397154331, + 0.01236905436962843, + 0.008766619488596916, + 0.014872681349515915, + 0.023905953392386436, + 0.022362003102898598, + 0.017191991209983826, + 0.014116999693214893, + 0.013423897325992584, + 0.038645949214696884, + 0.017233652994036674, + 0.002281906083226204, + 0.0023100448306649923, + 0.0035305784549564123, + 0.014503599144518375, + 0.003453482175245881, + 0.0031642497051507235, + 0.00589017616584897, + 0.004262132570147514, + 0.010745727457106113, + 0.013421013951301575, + 0.013898120261728764, + 0.02688920870423317, + 0.017544347792863846, + 0.01867702603340149, + 0.027380628511309624, + 0.018978293985128403, + 0.0053387898951768875, + 0.006230630911886692, + 0.004713938105851412, + 0.020946485921740532, + 0.005494282115250826, + 0.0049928962253034115, + 0.007761229760944843, + 0.010646410286426544, + 0.011276638135313988, + 0.01445951871573925, + 0.008034144528210163, + 0.026716234162449837, + 0.00685620354488492, + 0.006892291363328695, + 0.006903895176947117, + 0.027055801823735237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.020602626726031303 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870383068919182 + ], + "opposite_logprob_tokens": [ + -3.048703908920288 + ], + "logprob_reward": -0.04870383068919182, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870383068919182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08037488162517548, + 0.054055649787187576, + 0.09298962354660034, + 0.09219763427972794, + 0.131369486451149, + 0.1622132658958435, + 0.06104617193341255, + 0.18497784435749054, + 0.05311666429042816, + 0.08765873312950134 + ], + "credit_map": [ + 0.00558708468452096, + 0.007241237908601761, + 0.010055254213511944, + 0.022097334265708923, + 0.007482002954930067, + 0.007025384344160557, + 0.0139617333188653, + 0.006924856919795275, + 0.01018543727695942, + 0.005546177737414837, + 0.008203274570405483, + 0.006214494351297617, + 0.005723604001104832, + 0.006322687957435846, + 0.006976586766541004, + 0.004883389454334974, + 0.010314280167222023, + 0.004714543931186199, + 0.008272475562989712, + 0.0049414061941206455, + 0.007161723915487528, + 0.0077447970397770405, + 0.02897774428129196, + 0.02086266316473484, + 0.008512591011822224, + 0.004125562496483326, + 0.03205028176307678, + 0.006254865787923336, + 0.0062623885460197926, + 0.005056588910520077, + 0.012697014026343822, + 0.017238352447748184, + 0.008397938683629036, + 0.00976212602108717, + 0.03255962207913399, + 0.011725843884050846, + 0.01949499547481537, + 0.015936920419335365, + 0.021152891218662262, + 0.012339163571596146, + 0.015133287757635117, + 0.022846585139632225, + 0.020977243781089783, + 0.022384071722626686, + 0.014131048694252968, + 0.015173045918345451, + 0.030471747741103172, + 0.021096259355545044, + 0.005374227650463581, + 0.003126278519630432, + 0.0026665048208087683, + 0.020844170823693275, + 0.003196575678884983, + 0.0029860124923288822, + 0.009768117219209671, + 0.013084293343126774, + 0.014942570589482784, + 0.017442937940359116, + 0.01922655664384365, + 0.03659417852759361, + 0.018601858988404274, + 0.01819014735519886, + 0.023528192192316055, + 0.03645140677690506, + 0.009370493702590466, + 0.00782091449946165, + 0.0046507129445672035, + 0.005656338296830654, + 0.005218636244535446, + 0.006789695005863905, + 0.006313649006187916, + 0.007296229247003794, + 0.0068582650274038315, + 0.007682301569730043, + 0.006084894295781851, + 0.02422608993947506, + 0.0035806007217615843, + 0.0066717518493533134, + 0.005767724011093378, + 0.02678711526095867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.04870383068919182 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up still shot captured from a slightly elevated angle observes a cluster of pristine, deep purple grapes nestled against a flawless, matte white surface. Soft, diffused studio lighting sculpts their plump, spherical forms from the side, casting gentle shadows and creating bright, moist highlights across their taut skin, coated in delicate, powdery blossoms. As the grapes begin to melt, a surreal liquefaction begins; the process begins subtly, with the skin of one grape losing its tension and dissolving as if made of sugar. This transformation triggers a chain reaction as neighboring grapes steadily and deliberately collapse, their solid form deliquescing into a viscous liquid. Rather than cracking, the skins melt, their dark color seeping into a growing pool of dark, rich juice that slowly spreads across the pristine white surface. Its smooth, syrupy surface reflects light, creating a hypnotic, unsettling, and beautifully stark visual effect.", + "seed": 29981, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_1/index_4676_seed_29981_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2356276720762253, + "vlm_logprob": -0.2356276720762253, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.2356276720762253, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01013360358774662 + ], + "opposite_logprob_tokens": [ + -6.010133743286133 + ], + "logprob_reward": -0.01013360358774662, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.01013360358774662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08701833337545395, + 0.06425873190164566, + 0.07573867589235306, + 0.10094256699085236, + 0.027583526447415352, + 0.1288878470659256, + 0.028532808646559715, + 0.07671856880187988, + 0.3508163392543793, + 0.059502582997083664 + ], + "credit_map": [ + 0.017869744449853897, + 0.012538937851786613, + 0.010705995373427868, + 0.007618551608175039, + 0.00870051421225071, + 0.0060360305942595005, + 0.010845153592526913, + 0.012703410349786282, + 0.007189989555627108, + 0.009546807035803795, + 0.010233283042907715, + 0.005761298350989819, + 0.008450287394225597, + 0.006738841999322176, + 0.007790783885866404, + 0.008547439239919186, + 0.01147979125380516, + 0.006368882954120636, + 0.014719981700181961, + 0.005743113812059164, + 0.0074712238274514675, + 0.010502600111067295, + 0.008347933180630207, + 0.011105154640972614, + 0.01302308589220047, + 0.007505202665925026, + 0.010602787137031555, + 0.006790900602936745, + 0.008512921631336212, + 0.007912809029221535, + 0.009716620668768883, + 0.03687824308872223, + 0.003989519085735083, + 0.004000998567789793, + 0.0022751856595277786, + 0.002023952081799507, + 0.004556742031127214, + 0.0033052286598831415, + 0.003868023632094264, + 0.003563876962289214, + 0.009569223038852215, + 0.0049604088999331, + 0.005402045790106058, + 0.00659519899636507, + 0.01268806867301464, + 0.009724196046590805, + 0.027345940470695496, + 0.052602771669626236, + 0.005152271594852209, + 0.002429975662380457, + 0.002106373431161046, + 0.002219432732090354, + 0.004790269769728184, + 0.0032830408308655024, + 0.0025045310612767935, + 0.00604691356420517, + 0.007537127938121557, + 0.00904011819511652, + 0.004719601012766361, + 0.0048771994188427925, + 0.025566207244992256, + 0.00668638339266181, + 0.006288973614573479, + 0.012002954259514809, + 0.03467300161719322, + 0.02766920067369938, + 0.031082168221473694, + 0.01847452111542225, + 0.11360736191272736, + 0.027733655646443367, + 0.04485131800174713, + 0.05272509530186653, + 0.01777821034193039, + 0.005071454681456089, + 0.003798077115789056, + 0.003757151309400797, + 0.004096924792975187, + 0.004689650610089302, + 0.004611839074641466, + 0.015699276700615883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.01013360358774662 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539616346359253 + ], + "opposite_logprob_tokens": [ + -0.4289616644382477 + ], + "logprob_reward": -1.0539616346359253, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539616346359253, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.03101627342402935, + 0.062225863337516785, + 0.04752219840884209, + 0.058521874248981476, + 0.07775311172008514, + 0.2272094190120697, + 0.010503144934773445, + 0.041762575507164, + 0.3618176281452179, + 0.08166785538196564 + ], + "credit_map": [ + 0.005623762961477041, + 0.0051231528632342815, + 0.004701067227870226, + 0.0037460336461663246, + 0.002697969088330865, + 0.0027183000929653645, + 0.002937162993475795, + 0.003468823619186878, + 0.006723283790051937, + 0.01018908154219389, + 0.010578948073089123, + 0.005974483676254749, + 0.00606803223490715, + 0.006346127483993769, + 0.007070157211273909, + 0.009275748394429684, + 0.008594601415097713, + 0.00546027347445488, + 0.006358110345900059, + 0.003910579718649387, + 0.006465202663093805, + 0.004833000712096691, + 0.007209340576082468, + 0.004691090434789658, + 0.007733587641268969, + 0.005635022651404142, + 0.010307070799171925, + 0.006565814837813377, + 0.005101306363940239, + 0.007707236800342798, + 0.008176115341484547, + 0.0072957174852490425, + 0.013320141471922398, + 0.012861206196248531, + 0.006366678047925234, + 0.007174778264015913, + 0.009469343349337578, + 0.008808051235973835, + 0.010523517616093159, + 0.009229397401213646, + 0.014749466441571712, + 0.010401423089206219, + 0.010542131960391998, + 0.008240917697548866, + 0.019306061789393425, + 0.01876402646303177, + 0.05832255259156227, + 0.08688284456729889, + 0.0011133118532598019, + 0.0007495726458728313, + 0.0008670840179547668, + 0.0006501273019239306, + 0.003427083371207118, + 0.0012390915071591735, + 0.0012888996861875057, + 0.001167975482530892, + 0.004135726951062679, + 0.004429450258612633, + 0.0024925535544753075, + 0.00240884511731565, + 0.01441127434372902, + 0.00357923936098814, + 0.0028018944431096315, + 0.007503595668822527, + 0.0422462522983551, + 0.025436341762542725, + 0.03290809318423271, + 0.017930883914232254, + 0.12376970052719116, + 0.027863744646310806, + 0.04713089391589165, + 0.04453173279762268, + 0.027493374422192574, + 0.00598554452881217, + 0.004158177878707647, + 0.0024530489463359118, + 0.004230083432048559, + 0.006710008718073368, + 0.00398786598816514, + 0.026649752631783485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -1.0539616346359253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700201466679573 + ], + "opposite_logprob_tokens": [ + -2.6950201988220215 + ], + "logprob_reward": -0.0700201466679573, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700201466679573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03847614675760269, + 0.04941890761256218, + 0.03163093328475952, + 0.07079354673624039, + 0.053474705666303635, + 0.25566619634628296, + 0.026217477396130562, + 0.027165904641151428, + 0.40168023109436035, + 0.04547588899731636 + ], + "credit_map": [ + 0.005657003726810217, + 0.004606242291629314, + 0.004550388548523188, + 0.0037388901691883802, + 0.00459267245605588, + 0.005847434513270855, + 0.0049166325479745865, + 0.004566885996609926, + 0.005517229437828064, + 0.0073561654426157475, + 0.005845199804753065, + 0.004166924860328436, + 0.00816455390304327, + 0.005618508905172348, + 0.005804866086691618, + 0.006945461500436068, + 0.005443825386464596, + 0.002428096253424883, + 0.004016354214400053, + 0.0024549744557589293, + 0.002816731110215187, + 0.004263764247298241, + 0.0030297681223601103, + 0.007177420891821384, + 0.008152762427926064, + 0.00670023076236248, + 0.010934283025562763, + 0.008603539317846298, + 0.0062845628708601, + 0.008827464655041695, + 0.009234447963535786, + 0.012056257575750351, + 0.009007428772747517, + 0.007065685000270605, + 0.006185360252857208, + 0.0033707083202898502, + 0.00763910124078393, + 0.006837730761617422, + 0.006396845914423466, + 0.006971842609345913, + 0.015897920355200768, + 0.0101787643507123, + 0.011527935042977333, + 0.009742091409862041, + 0.022733747959136963, + 0.01948942057788372, + 0.06302853673696518, + 0.1030677855014801, + 0.002920945407822728, + 0.0025441187899559736, + 0.0021471173968166113, + 0.0021650006528943777, + 0.008397339843213558, + 0.002348888199776411, + 0.0023684813641011715, + 0.0033255841117352247, + 0.0027184737846255302, + 0.0020353086292743683, + 0.0022085539530962706, + 0.0019101345678791404, + 0.00938772689551115, + 0.0024032744113355875, + 0.00205113529227674, + 0.004451294895261526, + 0.06941766291856766, + 0.026683341711759567, + 0.0339229442179203, + 0.016964023932814598, + 0.14560864865779877, + 0.03146340698003769, + 0.04887568950653076, + 0.028744511306285858, + 0.013567514717578888, + 0.004338426515460014, + 0.0025376630946993828, + 0.003277309238910675, + 0.003047219244763255, + 0.004031520336866379, + 0.003538370132446289, + 0.011137868277728558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.0700201466679573 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207875356078148 + ], + "opposite_logprob_tokens": [ + -4.26420783996582 + ], + "logprob_reward": -0.014207875356078148, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207875356078148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06402235478162766, + 0.044683028012514114, + 0.03734555467963219, + 0.0864020362496376, + 0.04801943525671959, + 0.15674547851085663, + 0.027410713955760002, + 0.052206046879291534, + 0.32658618688583374, + 0.15657921135425568 + ], + "credit_map": [ + 0.008289805613458157, + 0.007663593627512455, + 0.007272124290466309, + 0.007797595579177141, + 0.007294893264770508, + 0.008525822311639786, + 0.008115716278553009, + 0.00906280055642128, + 0.004919741302728653, + 0.0048188273794949055, + 0.00762771675363183, + 0.004785258788615465, + 0.005684292409569025, + 0.004667722154408693, + 0.0058955783024430275, + 0.006283888593316078, + 0.0051319836638867855, + 0.0025141083169728518, + 0.004920140840113163, + 0.002044725464656949, + 0.002528172917664051, + 0.0048528993502259254, + 0.00381391285918653, + 0.011539612896740437, + 0.010531581938266754, + 0.006881697103381157, + 0.009985393844544888, + 0.00825877021998167, + 0.007078357506543398, + 0.00760697154328227, + 0.010294484905898571, + 0.02576478384435177, + 0.006732907611876726, + 0.0049802279099822044, + 0.006452060304582119, + 0.004365059081465006, + 0.006894977297633886, + 0.006925022695213556, + 0.005187380127608776, + 0.006481798831373453, + 0.009332796558737755, + 0.005199538078159094, + 0.00637544272467494, + 0.006516923196613789, + 0.013592246919870377, + 0.010840874165296555, + 0.04071637615561485, + 0.06417126953601837, + 0.004040491301566362, + 0.0032355429138988256, + 0.002249556127935648, + 0.0030897771939635277, + 0.004035394638776779, + 0.0025939170736819506, + 0.003455376485362649, + 0.004710656590759754, + 0.005509061738848686, + 0.005262367427349091, + 0.003287620609626174, + 0.002820416120812297, + 0.01543932780623436, + 0.0046463701874017715, + 0.0040612053126096725, + 0.011179679073393345, + 0.04974022135138512, + 0.024697063490748405, + 0.027971288189291954, + 0.014221316203474998, + 0.11629200726747513, + 0.025998113676905632, + 0.038489796221256256, + 0.0291763786226511, + 0.05438084155321121, + 0.011872119270265102, + 0.0070211696438491344, + 0.005935453809797764, + 0.008361028507351875, + 0.011765365488827229, + 0.009931948967278004, + 0.047311291098594666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.014207875356078148 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029815100133419037 + ], + "opposite_logprob_tokens": [ + -3.529815196990967 + ], + "logprob_reward": -0.029815100133419037, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029815100133419037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07798405736684799, + 0.049556661397218704, + 0.04578413441777229, + 0.028461771085858345, + 0.040063485503196716, + 0.04700418934226036, + 0.025048749521374702, + 0.09281682968139648, + 0.36247140169143677, + 0.23080866038799286 + ], + "credit_map": [ + 0.015043672174215317, + 0.011229117400944233, + 0.00965874083340168, + 0.006297113839536905, + 0.007843813858926296, + 0.006612415891140699, + 0.010977398604154587, + 0.010321796871721745, + 0.006140638142824173, + 0.007458498701453209, + 0.00890676211565733, + 0.0044799186289310455, + 0.006852204445749521, + 0.0033179072197526693, + 0.006272518076002598, + 0.006128219421952963, + 0.003920865245163441, + 0.0024873793590813875, + 0.007498140912503004, + 0.002504076575860381, + 0.003039620816707611, + 0.010671562515199184, + 0.006690046284347773, + 0.00897244643419981, + 0.003957622684538364, + 0.00399082712829113, + 0.0029141271952539682, + 0.0025196243077516556, + 0.0034041127655655146, + 0.00293420790694654, + 0.0036126929335296154, + 0.00512855825945735, + 0.005924458149820566, + 0.005382558796554804, + 0.004104957450181246, + 0.0028922397177666426, + 0.0058029210194945335, + 0.004386302083730698, + 0.0056919255293905735, + 0.005878127180039883, + 0.00523251760751009, + 0.005104537587612867, + 0.0033216814044862986, + 0.002962542697787285, + 0.008333549834787846, + 0.004124367609620094, + 0.0065473527647554874, + 0.011377643793821335, + 0.004248552490025759, + 0.0023513222113251686, + 0.0018176330486312509, + 0.00259856297634542, + 0.0035511336755007505, + 0.0026764192152768373, + 0.0032934113405644894, + 0.004511718172580004, + 0.007756575010716915, + 0.011639098636806011, + 0.00845734216272831, + 0.007681811228394508, + 0.02233067899942398, + 0.006936558987945318, + 0.009360021911561489, + 0.018654754385352135, + 0.04614179581403732, + 0.026014618575572968, + 0.03267141059041023, + 0.01784953847527504, + 0.12436928600072861, + 0.02739870734512806, + 0.04833661764860153, + 0.03968944773077965, + 0.07568392157554626, + 0.017531417310237885, + 0.011164854280650616, + 0.008441862650215626, + 0.014686399139463902, + 0.018689492717385292, + 0.010978935286402702, + 0.07363179326057434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.029815100133419037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up still shot captured from a slightly elevated angle observes a cluster of pristine, deep purple grapes nestled against a flawless, matte white surface. Soft, diffused studio lighting sculpts their plump, spherical forms from the side, casting gentle shadows and creating bright, moist highlights across their taut skin, coated in delicate, powdery blossoms. As the grapes begin to melt, a surreal liquefaction begins; the process begins subtly, with the skin of one grape losing its tension and dissolving as if made of sugar. This transformation triggers a chain reaction as neighboring grapes steadily and deliberately collapse, their solid form deliquescing into a viscous liquid. Rather than cracking, the skins melt, their dark color seeping into a growing pool of dark, rich juice that slowly spreads across the pristine white surface. Its smooth, syrupy surface reflects light, creating a hypnotic, unsettling, and beautifully stark visual effect.", + "seed": 29982, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_1/index_4676_seed_29982_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.31880363803356887, + "vlm_logprob": -0.31880363803356887, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.31880363803356887, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0061168367974460125 + ], + "opposite_logprob_tokens": [ + -6.00611686706543 + ], + "logprob_reward": -0.0061168367974460125, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0061168367974460125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12040980160236359, + 0.04253210499882698, + 0.10039236396551132, + 0.06157607585191727, + 0.19893747568130493, + 0.09802502393722534, + 0.034578319638967514, + 0.1870957911014557, + 0.123404361307621, + 0.03304864466190338 + ], + "credit_map": [ + 0.015975894406437874, + 0.015919888392090797, + 0.012967392802238464, + 0.013029527850449085, + 0.017325730994343758, + 0.018222304061055183, + 0.015717748552560806, + 0.011251316405832767, + 0.006590181961655617, + 0.004965740256011486, + 0.004666885826736689, + 0.00526084192097187, + 0.005432820413261652, + 0.004881598055362701, + 0.004003304988145828, + 0.006730732042342424, + 0.010739095509052277, + 0.010828734375536442, + 0.011879637837409973, + 0.01339514460414648, + 0.013616411946713924, + 0.013843504711985588, + 0.016747239977121353, + 0.009342596866190434, + 0.00601050304248929, + 0.007892866618931293, + 0.003991240635514259, + 0.007981820032000542, + 0.01148788072168827, + 0.009946536272764206, + 0.008913884870707989, + 0.005351340398192406, + 0.02245449833571911, + 0.019974885508418083, + 0.022958384826779366, + 0.01993255317211151, + 0.02511822246015072, + 0.024470556527376175, + 0.03430512547492981, + 0.02972324937582016, + 0.009133805520832539, + 0.009691523388028145, + 0.007281402591615915, + 0.010547962039709091, + 0.016241714358329773, + 0.02807639352977276, + 0.00670802965760231, + 0.010344192385673523, + 0.003482498461380601, + 0.005260997451841831, + 0.0037849722430109978, + 0.006374473217874765, + 0.004106827545911074, + 0.005127735435962677, + 0.003695083549246192, + 0.0027457308024168015, + 0.011856865137815475, + 0.008651360869407654, + 0.006658353377133608, + 0.007977752946317196, + 0.030260972678661346, + 0.0786706954240799, + 0.009730120189487934, + 0.03328966349363327, + 0.00898086465895176, + 0.008936530910432339, + 0.00732792541384697, + 0.008571500889956951, + 0.019326796755194664, + 0.046467214822769165, + 0.01481698639690876, + 0.008976546116173267, + 0.002670645946636796, + 0.0026516697835177183, + 0.004700703080743551, + 0.003870782908052206, + 0.005258835852146149, + 0.005126842297613621, + 0.0045403833501040936, + 0.004228780046105385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.0061168367974460125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020735263824463 + ], + "opposite_logprob_tokens": [ + -0.25207358598709106 + ], + "logprob_reward": -1.5020735263824463, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020735263824463, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2922620177268982, + 0.024460336193442345, + 0.051367130130529404, + 0.0691119059920311, + 0.01277000643312931, + 0.10562031716108322, + 0.05849602818489075, + 0.2197057157754898, + 0.1401071548461914, + 0.02609935775399208 + ], + "credit_map": [ + 0.02606077492237091, + 0.02010834775865078, + 0.02479296550154686, + 0.017582392320036888, + 0.034671708941459656, + 0.06652936339378357, + 0.037807315587997437, + 0.0647091269493103, + 0.0025760005228221416, + 0.0027654606383293867, + 0.0024694378953427076, + 0.0030122643802314997, + 0.004128898493945599, + 0.003871999215334654, + 0.0031490607652813196, + 0.0024872140493243933, + 0.01103818230330944, + 0.005730611737817526, + 0.005531670060008764, + 0.005032427608966827, + 0.0064433878287673, + 0.006895530037581921, + 0.005192493088543415, + 0.005502831190824509, + 0.006562504917383194, + 0.00782750640064478, + 0.0058413054794073105, + 0.011241191066801548, + 0.009342219680547714, + 0.011341117322444916, + 0.009736462496221066, + 0.007219601422548294, + 0.0014501420082524419, + 0.0013361830497160554, + 0.0015324201667681336, + 0.0015327255241572857, + 0.0020625286269932985, + 0.001994355581700802, + 0.001629877951927483, + 0.0012317728251218796, + 0.00984511524438858, + 0.010153796523809433, + 0.009152663871645927, + 0.01284429244697094, + 0.01414466928690672, + 0.020525015890598297, + 0.013717707246541977, + 0.015237061306834221, + 0.006321677938103676, + 0.0054847560822963715, + 0.007478716783225536, + 0.006042161025106907, + 0.006837236229330301, + 0.01400816161185503, + 0.006815851666033268, + 0.005507464986294508, + 0.014858216047286987, + 0.010485059581696987, + 0.007629747036844492, + 0.009304715320467949, + 0.0365169383585453, + 0.09450019150972366, + 0.012150046415627003, + 0.03426080569624901, + 0.009867790155112743, + 0.010119928047060966, + 0.005895575508475304, + 0.006841501221060753, + 0.022081445902585983, + 0.05771442875266075, + 0.012325599789619446, + 0.015260891057550907, + 0.0025974756572395563, + 0.002418726682662964, + 0.004268762655556202, + 0.00279223476536572, + 0.0025034730788320303, + 0.0028022758197039366, + 0.004376932047307491, + 0.004339477512985468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -1.5020735263824463 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067511171102524 + ], + "opposite_logprob_tokens": [ + -3.168067455291748 + ], + "logprob_reward": -0.043067511171102524, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067511171102524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1657898873090744, + 0.08614551275968552, + 0.06632779538631439, + 0.08307931572198868, + 0.03428898751735687, + 0.1586422473192215, + 0.056116070598363876, + 0.11672647297382355, + 0.2106248140335083, + 0.022258935496211052 + ], + "credit_map": [ + 0.01159330178052187, + 0.011303141713142395, + 0.014909511432051659, + 0.011562718078494072, + 0.022964537143707275, + 0.04010147601366043, + 0.020691566169261932, + 0.03266363590955734, + 0.008471278473734856, + 0.007122670765966177, + 0.006011770106852055, + 0.01706923544406891, + 0.0235697403550148, + 0.01035572774708271, + 0.006734753027558327, + 0.006810335908085108, + 0.006779789924621582, + 0.005786971189081669, + 0.008227037265896797, + 0.00914004072546959, + 0.008780092932283878, + 0.008516875095665455, + 0.010378027334809303, + 0.008718960918486118, + 0.007140885107219219, + 0.012231912463903427, + 0.006215095520019531, + 0.010980353690683842, + 0.013780518434941769, + 0.012669648043811321, + 0.01393113937228918, + 0.006129768677055836, + 0.003762182081118226, + 0.0036735718604177237, + 0.00343999988399446, + 0.00290616019628942, + 0.005083953496068716, + 0.004020771011710167, + 0.006663464475423098, + 0.004738887771964073, + 0.009296468459069729, + 0.009044565260410309, + 0.007344019133597612, + 0.011818991042673588, + 0.02433837577700615, + 0.05605095624923706, + 0.010013182647526264, + 0.030735690146684647, + 0.007032122928649187, + 0.005572841968387365, + 0.0054385242983698845, + 0.006286208517849445, + 0.007075730245560408, + 0.015397832728922367, + 0.0047060903161764145, + 0.00460671866312623, + 0.008013567887246609, + 0.006637622602283955, + 0.005292520858347416, + 0.005627520382404327, + 0.015669262036681175, + 0.04617050662636757, + 0.00778662133961916, + 0.021528853103518486, + 0.014239448122680187, + 0.01478701364248991, + 0.009013116359710693, + 0.009162579663097858, + 0.035281505435705185, + 0.08751147985458374, + 0.021248437464237213, + 0.019381234422326088, + 0.0019446535734459758, + 0.0021674756426364183, + 0.0036794620100408792, + 0.002336998237296939, + 0.002873137826099992, + 0.002883617067709565, + 0.0032953389454632998, + 0.003078252775594592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.043067511171102524 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643965046852827 + ], + "opposite_logprob_tokens": [ + -5.37964391708374 + ], + "logprob_reward": -0.004643965046852827, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643965046852827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24974671006202698, + 0.1546896994113922, + 0.06718891859054565, + 0.05100841447710991, + 0.03142392262816429, + 0.09334847331047058, + 0.12001489847898483, + 0.17505864799022675, + 0.020815229043364525, + 0.03670503571629524 + ], + "credit_map": [ + 0.021328752860426903, + 0.017631258815526962, + 0.020318668335676193, + 0.015394407324492931, + 0.031048284843564034, + 0.05773511156439781, + 0.03348426893353462, + 0.052805956453084946, + 0.01387227512896061, + 0.011226393282413483, + 0.007796676363795996, + 0.025060348212718964, + 0.04930780827999115, + 0.0177141260355711, + 0.008460302837193012, + 0.02125176601111889, + 0.00888813566416502, + 0.00539480522274971, + 0.009068384766578674, + 0.007020043209195137, + 0.008130356669425964, + 0.0067397309467196465, + 0.012635491788387299, + 0.00931196752935648, + 0.0041136969812214375, + 0.004800376016646624, + 0.005379921291023493, + 0.0051805670373141766, + 0.009229806251823902, + 0.0070453523658216, + 0.010760070756077766, + 0.004498625174164772, + 0.0028351000510156155, + 0.003168443450704217, + 0.003976290579885244, + 0.0027630196418613195, + 0.005111322738230228, + 0.003925935830920935, + 0.005636067595332861, + 0.0040077404119074345, + 0.007293414324522018, + 0.00537519995123148, + 0.0045132022351026535, + 0.006376686505973339, + 0.014064783230423927, + 0.029666803777217865, + 0.008306324481964111, + 0.017752068117260933, + 0.02066735178232193, + 0.01053609699010849, + 0.008300969377160072, + 0.006062034517526627, + 0.017327455803751945, + 0.04107237234711647, + 0.007745338138192892, + 0.008303278125822544, + 0.011825019493699074, + 0.00833363737910986, + 0.008435124531388283, + 0.009304541163146496, + 0.030193887650966644, + 0.07178250700235367, + 0.011218778789043427, + 0.023965168744325638, + 0.0017366694519296288, + 0.0025010097306221724, + 0.002471323823556304, + 0.003885578364133835, + 0.002210520440712571, + 0.0030366303399205208, + 0.0024555320851504803, + 0.0025179649237543344, + 0.0025739932898432016, + 0.002474809531122446, + 0.004817125853151083, + 0.0030749414581805468, + 0.0031831704545766115, + 0.004098580684512854, + 0.007213844917714596, + 0.009268567897379398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.004643965046852827 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a cluster of deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grapes resting on a flawless, matte white surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the grapes beginning to melt and dissolve into a viscous liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the melting process cause the dark grape color to seep into a growing pool of juice on the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated looking down at the grapes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811635076999664 + ], + "opposite_logprob_tokens": [ + -3.288116455078125 + ], + "logprob_reward": -0.03811635076999664, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811635076999664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16344091296195984, + 0.0850118026137352, + 0.07425197958946228, + 0.08669976145029068, + 0.030662892386317253, + 0.09523436427116394, + 0.07184890657663345, + 0.08427563309669495, + 0.23992802202701569, + 0.06864573806524277 + ], + "credit_map": [ + 0.017021844163537025, + 0.010524456389248371, + 0.01537797786295414, + 0.010597599670290947, + 0.017893744632601738, + 0.03478023037314415, + 0.022892601788043976, + 0.03435245528817177, + 0.007191968150436878, + 0.00681866193190217, + 0.005687538534402847, + 0.010465849190950394, + 0.023872269317507744, + 0.0102535430341959, + 0.007119480054825544, + 0.013602490536868572, + 0.010798578150570393, + 0.0068984259851276875, + 0.0072178104892373085, + 0.007962868548929691, + 0.007944715209305286, + 0.009404574520885944, + 0.013984459452331066, + 0.010040546767413616, + 0.0070674908347427845, + 0.010206446051597595, + 0.008364415727555752, + 0.014270932413637638, + 0.0114629240706563, + 0.01533456053584814, + 0.012201662175357342, + 0.007791324984282255, + 0.0030889210756868124, + 0.0029137672390788794, + 0.0034302559215575457, + 0.0026841668877750635, + 0.005502335261553526, + 0.003622818738222122, + 0.005615611560642719, + 0.003805015701800585, + 0.008759642951190472, + 0.007731142919510603, + 0.007783383596688509, + 0.010743264108896255, + 0.012478229589760303, + 0.019519122317433357, + 0.013280787505209446, + 0.014938784763216972, + 0.00973602943122387, + 0.008969592861831188, + 0.008235537447035313, + 0.01022808812558651, + 0.00863101240247488, + 0.01092529296875, + 0.007880333811044693, + 0.007243019063025713, + 0.00566672021523118, + 0.004282230045646429, + 0.004125978332012892, + 0.00414696242660284, + 0.014743196777999401, + 0.03532405197620392, + 0.0070222229696810246, + 0.008964265696704388, + 0.016415035352110863, + 0.01683199405670166, + 0.008705498650670052, + 0.00887279212474823, + 0.03970487043261528, + 0.10217367112636566, + 0.023361442610621452, + 0.023862730711698532, + 0.005092873703688383, + 0.005716044921427965, + 0.00781932007521391, + 0.006425026338547468, + 0.005469564814120531, + 0.007844444364309311, + 0.011753569357097149, + 0.01852489449083805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.03811635076999664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_10/index_tensor([2467, 2467])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_10/index_tensor([2467, 2467])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8b274c3866e298716baadff91314706cbd8bc4ac --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_10/index_tensor([2467, 2467])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "still, black-and-white close-up shot from a slightly low angle, dominated by high-contrast tones. The focus falls precisely on an elderly hand, its wrinkled skin dry and sagging. The ridges and veins on the back of the hand, etched with age, are clearly outlined in the soft sidelight, the knuckles appearing thick. This hand rests on an old, yellowed photograph, its edges warped and scuffed from repeated handling, its surface minutely scratched. Carefully and with infinite cherishment, the elderly man's rough fingertips trace the worn white border of the photograph, slowly and steadily sliding from one end to the other. The gentle movement is as if he is revisiting the distant memories the photograph holds through touch. The background is an out-of-focus, dark wooden tabletop, its blurred texture providing a quiet contrast to the delicate emotions in the foreground. The entire image is imbued with a tranquil, soulful atmosphere of nostalgia, as if time has stood still in this moment.", + "seed": 10641, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_10/index_2467_seed_10641_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03788773971609771, + "vlm_logprob": -0.03788773971609771, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.03788773971609771, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14667439460754395 + ], + "opposite_logprob_tokens": [ + -2.021674394607544 + ], + "logprob_reward": -0.14667439460754395, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14667439460754395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2434869259595871, + 0.3088816702365875, + 0.13741303980350494, + 0.08969696611166, + 0.05088428035378456, + 0.011133300140500069, + 0.026134179905056953, + 0.012303459458053112, + 0.07210122048854828, + 0.04796496778726578 + ], + "credit_map": [ + 0.019732622429728508, + 0.016677051782608032, + 0.014188650995492935, + 0.0286046601831913, + 0.02629421092569828, + 0.013902042992413044, + 0.08807480335235596, + 0.03601288050413132, + 0.01561625488102436, + 0.02142210491001606, + 0.007571937516331673, + 0.012398475781083107, + 0.02305109053850174, + 0.10721339285373688, + 0.09429901093244553, + 0.027309399098157883, + 0.021044965833425522, + 0.01950896345078945, + 0.017442308366298676, + 0.016916844993829727, + 0.01709248311817646, + 0.021013999357819557, + 0.012669563293457031, + 0.01172391977161169, + 0.00745119946077466, + 0.005333824083209038, + 0.006115760654211044, + 0.003645760240033269, + 0.012297317385673523, + 0.02893705479800701, + 0.0068970369175076485, + 0.01901901327073574, + 0.005001470446586609, + 0.00528648542240262, + 0.005005113314837217, + 0.0029950342141091824, + 0.01736161857843399, + 0.0032968721352517605, + 0.002892155898734927, + 0.00904553011059761, + 0.00138632592279464, + 0.001064022770151496, + 0.003470851806923747, + 0.0008218401344493032, + 0.001045064302161336, + 0.0009676831541582942, + 0.000981029705144465, + 0.0013964818790555, + 0.0024764330592006445, + 0.0025257491506636143, + 0.0035487422719597816, + 0.0030374140478670597, + 0.0031571092549711466, + 0.0034811701625585556, + 0.004146668128669262, + 0.0037608954589813948, + 0.0015503776958212256, + 0.0011851551244035363, + 0.0022796455305069685, + 0.0017768468242138624, + 0.001344966352917254, + 0.0014470221940428019, + 0.0015875723911449313, + 0.0011318728793412447, + 0.007347499020397663, + 0.012655150145292282, + 0.008510163985192776, + 0.009113984182476997, + 0.008279555477201939, + 0.010022121481597424, + 0.010537556372582912, + 0.005635190289467573, + 0.0050945524126291275, + 0.010488546453416348, + 0.005918275099247694, + 0.004362139385193586, + 0.005518442951142788, + 0.008143322542309761, + 0.004748724866658449, + 0.003690961515530944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.14667439460754395 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005261858459562063 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.005261858459562063, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005261858459562063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1114436611533165, + 0.4463069438934326, + 0.033834509551525116, + 0.18750829994678497, + 0.030799023807048798, + 0.020647061988711357, + 0.022227000445127487, + 0.039494726806879044, + 0.04735330119729042, + 0.060385506600141525 + ], + "credit_map": [ + 0.018967030569911003, + 0.011560015380382538, + 0.010411635041236877, + 0.014694432727992535, + 0.015213037841022015, + 0.013230260461568832, + 0.016632476821541786, + 0.010734761133790016, + 0.021238908171653748, + 0.027413450181484222, + 0.009949791245162487, + 0.01643046922981739, + 0.03445635363459587, + 0.15090250968933105, + 0.14331366121768951, + 0.0426018051803112, + 0.003053603693842888, + 0.00270516867749393, + 0.003969288431107998, + 0.003199176862835884, + 0.0035981633700430393, + 0.005656607914716005, + 0.004230467136949301, + 0.007422035094350576, + 0.011950703337788582, + 0.012144459411501884, + 0.009232661686837673, + 0.006448462139815092, + 0.024617085233330727, + 0.06099298223853111, + 0.01629042625427246, + 0.04583151638507843, + 0.0026979693211615086, + 0.003238968551158905, + 0.00403209775686264, + 0.002053994918242097, + 0.00855741836130619, + 0.002688505221158266, + 0.00308006233535707, + 0.00445000734180212, + 0.0022591252345591784, + 0.0019263472640886903, + 0.0024349666200578213, + 0.0014490392059087753, + 0.006451710127294064, + 0.0015914255054667592, + 0.0015391911147162318, + 0.002995256334543228, + 0.0025877216830849648, + 0.0032427008263766766, + 0.00410961639136076, + 0.002933813724666834, + 0.0027391391340643167, + 0.002528608776628971, + 0.0019669278990477324, + 0.002118471311405301, + 0.005858988501131535, + 0.00460786372423172, + 0.0037902933545410633, + 0.004985933657735586, + 0.004497847985476255, + 0.003836950985714793, + 0.007972357794642448, + 0.003944488242268562, + 0.005676492117345333, + 0.00853308942168951, + 0.005246918182820082, + 0.005876804701983929, + 0.00468582846224308, + 0.007259630598127842, + 0.006226842757314444, + 0.0038476965855807066, + 0.008206135593354702, + 0.016420045867562294, + 0.006426319945603609, + 0.004108036402612925, + 0.00669148750603199, + 0.009005754254758358, + 0.004615573212504387, + 0.004912154749035835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.005261858459562063 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093957599252462 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004093957599252462, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093957599252462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04913117364048958, + 0.16071286797523499, + 0.07227670401334763, + 0.17049871385097504, + 0.130793496966362, + 0.04423103109002113, + 0.04809194803237915, + 0.07747271656990051, + 0.14377929270267487, + 0.10301201790571213 + ], + "credit_map": [ + 0.010296870954334736, + 0.006148347165435553, + 0.004825410898774862, + 0.005011039786040783, + 0.005926933139562607, + 0.004004674032330513, + 0.005849021952599287, + 0.0070688785053789616, + 0.00801073107868433, + 0.011122304946184158, + 0.0036826380528509617, + 0.005847416818141937, + 0.012219229713082314, + 0.056772373616695404, + 0.0495394691824913, + 0.013518720865249634, + 0.007852367125451565, + 0.008953632786870003, + 0.006168313790112734, + 0.0053225327283144, + 0.007561069913208485, + 0.01867877133190632, + 0.005513048265129328, + 0.012226981110870838, + 0.0098681366071105, + 0.012300345115363598, + 0.012731922790408134, + 0.008116589859127998, + 0.02244682051241398, + 0.05107273533940315, + 0.016561191529035568, + 0.0374009869992733, + 0.011261464096605778, + 0.01567467860877514, + 0.0104338014498353, + 0.011169453151524067, + 0.03183900564908981, + 0.015207396820187569, + 0.010489600710570812, + 0.024718107655644417, + 0.00397131172940135, + 0.004750285763293505, + 0.006202236283570528, + 0.0039965808391571045, + 0.010976632125675678, + 0.004442133475095034, + 0.0028190561570227146, + 0.007072801236063242, + 0.005932728759944439, + 0.0049353730864822865, + 0.01070182491093874, + 0.00688962172716856, + 0.004054618068039417, + 0.004961851984262466, + 0.003813723335042596, + 0.006802209187299013, + 0.01045551523566246, + 0.008115078322589397, + 0.00855317898094654, + 0.01021356787532568, + 0.009196089580655098, + 0.007780499756336212, + 0.014055699110031128, + 0.009103096090257168, + 0.015586198307573795, + 0.026587320491671562, + 0.016198819503188133, + 0.013623341917991638, + 0.013863920234143734, + 0.03641451522707939, + 0.011276047676801682, + 0.01022913958877325, + 0.01647953875362873, + 0.024390380829572678, + 0.013494696468114853, + 0.007831401191651821, + 0.013063328340649605, + 0.012088160961866379, + 0.008235436864197254, + 0.007429076358675957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.004093957599252462 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793808236718178 + ], + "opposite_logprob_tokens": [ + -3.5297937393188477 + ], + "logprob_reward": -0.029793808236718178, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793808236718178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0750710517168045, + 0.3486049175262451, + 0.025055499747395515, + 0.06303121149539948, + 0.04864956811070442, + 0.045044444501399994, + 0.13047288358211517, + 0.12910491228103638, + 0.05852728709578514, + 0.07643823325634003 + ], + "credit_map": [ + 0.0062214466743171215, + 0.0033211165573447943, + 0.002495632041245699, + 0.004517781548202038, + 0.009082522243261337, + 0.0034667402505874634, + 0.029869576916098595, + 0.016096238046884537, + 0.015583865344524384, + 0.02279292605817318, + 0.007954265922307968, + 0.012700041756033897, + 0.02414277009665966, + 0.11810826510190964, + 0.11373311281204224, + 0.0335896760225296, + 0.003210530849173665, + 0.0032108421437442303, + 0.0045821829698979855, + 0.0030486483592540026, + 0.0030303234234452248, + 0.0027660734485834837, + 0.002155556809157133, + 0.0030513417441397905, + 0.007967036217451096, + 0.005943278316408396, + 0.010336476378142834, + 0.005559172015637159, + 0.007919161580502987, + 0.010585346259176731, + 0.006901695858687162, + 0.007819049060344696, + 0.007410723250359297, + 0.005436798091977835, + 0.007781758438795805, + 0.004814982879906893, + 0.007063476834446192, + 0.004715954884886742, + 0.006042870227247477, + 0.005383005831390619, + 0.00435097748413682, + 0.006024681031703949, + 0.00530684320256114, + 0.004916429053992033, + 0.009114346466958523, + 0.005941924173384905, + 0.00383822713047266, + 0.005551012232899666, + 0.013853226788341999, + 0.015895480290055275, + 0.028764886781573296, + 0.01769774593412876, + 0.01902412250638008, + 0.01324138417840004, + 0.00894250813871622, + 0.013053533621132374, + 0.012227983213961124, + 0.017912354320287704, + 0.015433726832270622, + 0.016259469091892242, + 0.014125523157417774, + 0.011317288503050804, + 0.03049292601644993, + 0.011335632763803005, + 0.00825122743844986, + 0.00873890332877636, + 0.009677181020379066, + 0.006277722306549549, + 0.007133832201361656, + 0.007011279463768005, + 0.007634045090526342, + 0.003803091589361429, + 0.007906940765678883, + 0.01231750100851059, + 0.007416930980980396, + 0.005844049621373415, + 0.010993482545018196, + 0.01575280912220478, + 0.009446574375033379, + 0.00675994623452425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.029793808236718178 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003614679677411914 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.003614679677411914, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003614679677411914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20852480828762054, + 0.3962461054325104, + 0.062394432723522186, + 0.05096346139907837, + 0.05539099872112274, + 0.023849474266171455, + 0.03084169700741768, + 0.03443116322159767, + 0.06632641702890396, + 0.0710313692688942 + ], + "credit_map": [ + 0.025973599404096603, + 0.016605865210294724, + 0.019574487581849098, + 0.01993861049413681, + 0.016890335828065872, + 0.016955403611063957, + 0.058035675436258316, + 0.03455083817243576, + 0.018839262425899506, + 0.026479126885533333, + 0.009549044072628021, + 0.016122039407491684, + 0.028388850390911102, + 0.13108159601688385, + 0.1256079375743866, + 0.04017826169729233, + 0.007844576612114906, + 0.00878150574862957, + 0.008257861249148846, + 0.005401818081736565, + 0.007607195060700178, + 0.00987840723246336, + 0.008397344499826431, + 0.006225726567208767, + 0.005097623448818922, + 0.003945882432162762, + 0.006390921305865049, + 0.0039017084054648876, + 0.007189882453531027, + 0.012866747565567493, + 0.004108601249754429, + 0.007462093140929937, + 0.009774224832654, + 0.007345241028815508, + 0.007642840966582298, + 0.005282931495457888, + 0.0060155196115374565, + 0.006803412456065416, + 0.005186033435165882, + 0.007340794429183006, + 0.0032702013850212097, + 0.0020453259348869324, + 0.004606566391885281, + 0.0024003356229513884, + 0.0030652498826384544, + 0.002533720340579748, + 0.0017974950606003404, + 0.004130578599870205, + 0.0029644726309925318, + 0.0035502936225384474, + 0.005686263088136911, + 0.0035001318901777267, + 0.0066917468793690205, + 0.003186682239174843, + 0.0031196079216897488, + 0.002142498502507806, + 0.0035728709772229195, + 0.006486908067017794, + 0.0049302587285637856, + 0.00468041654676199, + 0.00193011574447155, + 0.0054578641429543495, + 0.0031138586346060038, + 0.004258870612829924, + 0.010441484861075878, + 0.012175822630524635, + 0.009460755623877048, + 0.008438308723270893, + 0.0066920616663992405, + 0.006971150636672974, + 0.004736850969493389, + 0.007409981451928616, + 0.009112928993999958, + 0.012140491977334023, + 0.010466526262462139, + 0.006113527342677116, + 0.010074673220515251, + 0.008738466538488865, + 0.007281963713467121, + 0.007102792151272297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.003614679677411914 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, black-and-white close-up shot from a slightly low angle, dominated by high-contrast tones. The focus falls precisely on an elderly hand, its wrinkled skin dry and sagging. The ridges and veins on the back of the hand, etched with age, are clearly outlined in the soft sidelight, the knuckles appearing thick. This hand rests on an old, yellowed photograph, its edges warped and scuffed from repeated handling, its surface minutely scratched. Carefully and with infinite cherishment, the elderly man's rough fingertips trace the worn white border of the photograph, slowly and steadily sliding from one end to the other. The gentle movement is as if he is revisiting the distant memories the photograph holds through touch. The background is an out-of-focus, dark wooden tabletop, its blurred texture providing a quiet contrast to the delicate emotions in the foreground. The entire image is imbued with a tranquil, soulful atmosphere of nostalgia, as if time has stood still in this moment.", + "seed": 10642, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_10/index_2467_seed_10642_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05043308092281222, + "vlm_logprob": -0.05043308092281222, + "vlm_yes_no_margin": 3.949999976158142, + "vlm_token_credit": -0.05043308092281222, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10476686805486679 + ], + "opposite_logprob_tokens": [ + -2.354766845703125 + ], + "logprob_reward": -0.10476686805486679, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10476686805486679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06397156417369843, + 0.05674474313855171, + 0.05249667167663574, + 0.012975677847862244, + 0.5244876146316528, + 0.23069685697555542, + 0.019092446193099022, + 0.024299684911966324, + 0.007283786311745644, + 0.007950925268232822 + ], + "credit_map": [ + 0.008781536482274532, + 0.007876904681324959, + 0.00613505719229579, + 0.006531992927193642, + 0.008921855129301548, + 0.01144968718290329, + 0.006200982257723808, + 0.008073551580309868, + 0.0029614055529236794, + 0.003219759091734886, + 0.0031192374881356955, + 0.002417248208075762, + 0.01766740344464779, + 0.01655516028404236, + 0.0043998053297400475, + 0.006404721178114414, + 0.00551946833729744, + 0.005300823133438826, + 0.004877866245806217, + 0.0036117732524871826, + 0.014057555235922337, + 0.00846840813755989, + 0.0052925655618309975, + 0.005368211772292852, + 0.001506373519077897, + 0.0016436598962172866, + 0.0014761387137696147, + 0.0021633943542838097, + 0.0016762795858085155, + 0.002290165051817894, + 0.0009039464057423174, + 0.0013157203793525696, + 0.021563125774264336, + 0.012652984820306301, + 0.02858176827430725, + 0.00844611693173647, + 0.23093576729297638, + 0.17434367537498474, + 0.02200663648545742, + 0.025957556441426277, + 0.02083798311650753, + 0.03206125646829605, + 0.023682018741965294, + 0.021578378975391388, + 0.06240764632821083, + 0.028357066214084625, + 0.024377068504691124, + 0.01739543490111828, + 0.0023738269228488207, + 0.0014174793614074588, + 0.00342888874001801, + 0.0024129897356033325, + 0.0025507875252515078, + 0.003553604707121849, + 0.001822118298150599, + 0.0015327493892982602, + 0.002722314326092601, + 0.002426152816042304, + 0.003114639548584819, + 0.0032199136912822723, + 0.003674674779176712, + 0.003190508345142007, + 0.003997781313955784, + 0.001953702187165618, + 0.001261646393686533, + 0.0010660336120054126, + 0.0009560369653627276, + 0.0006900646258145571, + 0.0009929247898980975, + 0.0008077898528426886, + 0.0008211510721594095, + 0.0006881391163915396, + 0.0010783246252685785, + 0.0007519133505411446, + 0.0008892123587429523, + 0.001169850118458271, + 0.0010146374115720391, + 0.0006306252325884998, + 0.0016261158743873239, + 0.000790246413089335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.10476686805486679 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004648118279874325 + ], + "opposite_logprob_tokens": [ + -5.379648208618164 + ], + "logprob_reward": -0.004648118279874325, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004648118279874325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05880161374807358, + 0.1226436197757721, + 0.02349795773625374, + 0.01777651160955429, + 0.5625742077827454, + 0.10751653462648392, + 0.06179189682006836, + 0.017346501350402832, + 0.009233569726347923, + 0.018817510455846786 + ], + "credit_map": [ + 0.005616961512714624, + 0.007429131772369146, + 0.0058508384972810745, + 0.007249508984386921, + 0.00890498235821724, + 0.009606871753931046, + 0.008151895366609097, + 0.005991424899548292, + 0.004555779043585062, + 0.005474971141666174, + 0.004654283635318279, + 0.003706372808665037, + 0.04779670014977455, + 0.041885990649461746, + 0.0062811472453176975, + 0.008288389071822166, + 0.0020183315500617027, + 0.0029338153544813395, + 0.0026699157897382975, + 0.0033094659447669983, + 0.004290105774998665, + 0.0028677701484411955, + 0.002345259068533778, + 0.003063297364860773, + 0.0021761017851531506, + 0.0021503267344087362, + 0.0020329621620476246, + 0.0024859930854290724, + 0.002952261595055461, + 0.0026741023175418377, + 0.0016294830711558461, + 0.0016752835363149643, + 0.024338753893971443, + 0.013674117624759674, + 0.020486177876591682, + 0.010518196038901806, + 0.2445625215768814, + 0.19910584390163422, + 0.02248418889939785, + 0.02740449272096157, + 0.010943044908344746, + 0.014562137424945831, + 0.010469767265021801, + 0.00977356918156147, + 0.028643490746617317, + 0.013596896082162857, + 0.01185646653175354, + 0.00767118064686656, + 0.008955603465437889, + 0.007379339076578617, + 0.009042726829648018, + 0.009270165115594864, + 0.00780089944601059, + 0.006428003776818514, + 0.008256729692220688, + 0.004658436868339777, + 0.001983126625418663, + 0.0017282122280448675, + 0.002334721852093935, + 0.002438511233776808, + 0.0028685342986136675, + 0.0019646361470222473, + 0.0021467015612870455, + 0.0018820584518834949, + 0.0011785362148657441, + 0.0009610482375137508, + 0.0013642472913488746, + 0.0012616387102752924, + 0.001221823855303228, + 0.0012433577794581652, + 0.0010843116324394941, + 0.000918606820050627, + 0.0018193389987573028, + 0.002220605034381151, + 0.0023777433671057224, + 0.002201198600232601, + 0.002871684730052948, + 0.0017835386097431183, + 0.00380202429369092, + 0.0017413794994354248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.004648118279874325 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536294758319855 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536294758319855, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536294758319855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06428590416908264, + 0.15038569271564484, + 0.03846675902605057, + 0.01043921709060669, + 0.6203149557113647, + 0.06384814530611038, + 0.014423104003071785, + 0.016775846481323242, + 0.013712655752897263, + 0.007347732782363892 + ], + "credit_map": [ + 0.012503479607403278, + 0.008758993819355965, + 0.005554733797907829, + 0.006796964444220066, + 0.00795142911374569, + 0.0062850783579051495, + 0.006088972091674805, + 0.010346243157982826, + 0.007028062827885151, + 0.007144494913518429, + 0.006877096835523844, + 0.00646163197234273, + 0.05357137322425842, + 0.04435465857386589, + 0.010384025983512402, + 0.0145643325522542, + 0.002296128310263157, + 0.003333902917802334, + 0.003401238936930895, + 0.004487693775445223, + 0.007554980926215649, + 0.007639004848897457, + 0.0046336473897099495, + 0.005120157264173031, + 0.0009863505838438869, + 0.0009149736142717302, + 0.0011241063475608826, + 0.0015307404100894928, + 0.0019354911055415869, + 0.0014707653317600489, + 0.0012305132113397121, + 0.001246274565346539, + 0.024383842945098877, + 0.014089798554778099, + 0.03931901603937149, + 0.010648616589605808, + 0.27424195408821106, + 0.20003128051757812, + 0.02620486542582512, + 0.031395573168992996, + 0.005922161974012852, + 0.0077969408594071865, + 0.005980285815894604, + 0.004613530822098255, + 0.018057426437735558, + 0.00934716034680605, + 0.007377565838396549, + 0.004753071814775467, + 0.002329970244318247, + 0.001287390012294054, + 0.002123013837262988, + 0.002261783927679062, + 0.0019903958309441805, + 0.0011851268354803324, + 0.0019710403867065907, + 0.001274382695555687, + 0.0025590169243514538, + 0.002158528193831444, + 0.0019772136583924294, + 0.001931195380166173, + 0.002106513362377882, + 0.0021553330589085817, + 0.0018709346186369658, + 0.002017109887674451, + 0.002027088077738881, + 0.0016734610544517636, + 0.0020771510899066925, + 0.0017908205045387149, + 0.0020731871481984854, + 0.001317125977948308, + 0.0014421215746551752, + 0.001311697531491518, + 0.0008698580204509199, + 0.0011278854217380285, + 0.0009465260081924498, + 0.0008793370798230171, + 0.0009069344960153103, + 0.0007097411435097456, + 0.001218132907524705, + 0.0006893167737871408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.012536294758319855 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270228624343872 + ], + "opposite_logprob_tokens": [ + -2.1270227432250977 + ], + "logprob_reward": -0.1270228624343872, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270228624343872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1547863781452179, + 0.2948235869407654, + 0.02660996839404106, + 0.019348010420799255, + 0.33726418018341064, + 0.09113604575395584, + 0.029351867735385895, + 0.014951939694583416, + 0.014248870313167572, + 0.01747908629477024 + ], + "credit_map": [ + 0.02126477286219597, + 0.022305577993392944, + 0.014386488124728203, + 0.01344501506537199, + 0.01935662515461445, + 0.02334778569638729, + 0.01464898232370615, + 0.02603115327656269, + 0.01775520294904709, + 0.019657321274280548, + 0.016601484268903732, + 0.014872639440000057, + 0.09542225301265717, + 0.07912669330835342, + 0.020158812403678894, + 0.031229203566908836, + 0.002306611044332385, + 0.0030516069382429123, + 0.003064684337005019, + 0.0030815135687589645, + 0.00488870358094573, + 0.0037902728654444218, + 0.0030393300112336874, + 0.003387248609215021, + 0.0026252553798258305, + 0.0018261568620800972, + 0.002078223042190075, + 0.002413742942735553, + 0.0024169243406504393, + 0.003604806959629059, + 0.0017780925845727324, + 0.0026048088911920786, + 0.017322970554232597, + 0.008763767778873444, + 0.03130331635475159, + 0.014517842791974545, + 0.11294355988502502, + 0.12661945819854736, + 0.009639824740588665, + 0.01615344174206257, + 0.01978266052901745, + 0.014350554905831814, + 0.009870234876871109, + 0.006099534220993519, + 0.017424093559384346, + 0.010006100870668888, + 0.007401636335998774, + 0.006201247218996286, + 0.0038456395268440247, + 0.0030220288317650557, + 0.004467163700610399, + 0.0043455264531075954, + 0.0033643573988229036, + 0.003537213895469904, + 0.003920463379472494, + 0.0028494761791080236, + 0.0015946817584335804, + 0.0016244297148659825, + 0.002148887375369668, + 0.002291284268721938, + 0.0022478941828012466, + 0.0014371041906997561, + 0.0023872675374150276, + 0.0012203914811834693, + 0.0017391465371474624, + 0.0013625153806060553, + 0.0020408686250448227, + 0.001727381837554276, + 0.0026023369282484055, + 0.001229922636412084, + 0.0023287672083824873, + 0.0012179329060018063, + 0.0015812095953151584, + 0.0017579375999048352, + 0.0018746505957096815, + 0.0024113135877996683, + 0.002220956375822425, + 0.001454832497984171, + 0.004474076442420483, + 0.0017041119281202555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.1270228624343872 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319126108661294 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.00319126108661294, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319126108661294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029532477259635925, + 0.04480879753828049, + 0.01690133661031723, + 0.014170640148222446, + 0.7846209406852722, + 0.07638737559318542, + 0.009987841360270977, + 0.006134895142167807, + 0.010765808634459972, + 0.006689856760203838 + ], + "credit_map": [ + 0.006094804964959621, + 0.004102683160454035, + 0.0031717459205538034, + 0.003491425421088934, + 0.0036410391330718994, + 0.003469916060566902, + 0.0023756097070872784, + 0.0031852577812969685, + 0.002110000466927886, + 0.0028095291927456856, + 0.002462402917444706, + 0.0021572536788880825, + 0.015785589814186096, + 0.01324501633644104, + 0.0028225795831531286, + 0.003416431602090597, + 0.0018757438519969583, + 0.0020143475849181414, + 0.001799434539861977, + 0.002043116604909301, + 0.0028866406064480543, + 0.001988820731639862, + 0.001757063902914524, + 0.0025361725129187107, + 0.0017836971674114466, + 0.0011664850171655416, + 0.001573146553710103, + 0.0018937347922474146, + 0.0024635624140501022, + 0.002120730932801962, + 0.0014349623816087842, + 0.001734323101118207, + 0.02932499721646309, + 0.01776408776640892, + 0.07351455092430115, + 0.017389420419931412, + 0.3443686366081238, + 0.22557960450649261, + 0.03512736037373543, + 0.041552331298589706, + 0.00828984659165144, + 0.011075896210968494, + 0.00729813426733017, + 0.006194370333105326, + 0.019442018121480942, + 0.009668358601629734, + 0.008542395196855068, + 0.005876361858099699, + 0.0014151649083942175, + 0.0011103057768195868, + 0.0015845083398744464, + 0.0018974399426952004, + 0.001179870800115168, + 0.000816020299680531, + 0.0010801060125231743, + 0.0009044270263984799, + 0.0008186320774257183, + 0.0006870419601909816, + 0.0007429055403918028, + 0.0007449585827998817, + 0.000976037175860256, + 0.0007674756343476474, + 0.0008219434530474246, + 0.0005759020568802953, + 0.0011931704357266426, + 0.0009539463790133595, + 0.0015151615953072906, + 0.001394979190081358, + 0.0015844331355765462, + 0.0010868518147617579, + 0.0020191934891045094, + 0.0010180743411183357, + 0.0009652225417084992, + 0.0010236110538244247, + 0.000808064709417522, + 0.0006267951685003936, + 0.0008619021391496062, + 0.0008868606528267264, + 0.00095803861040622, + 0.0005593625246547163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.00319126108661294 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, black-and-white close-up shot from a slightly low angle, dominated by high-contrast tones. The focus falls precisely on an elderly hand, its wrinkled skin dry and sagging. The ridges and veins on the back of the hand, etched with age, are clearly outlined in the soft sidelight, the knuckles appearing thick. This hand rests on an old, yellowed photograph, its edges warped and scuffed from repeated handling, its surface minutely scratched. Carefully and with infinite cherishment, the elderly man's rough fingertips trace the worn white border of the photograph, slowly and steadily sliding from one end to the other. The gentle movement is as if he is revisiting the distant memories the photograph holds through touch. The background is an out-of-focus, dark wooden tabletop, its blurred texture providing a quiet contrast to the delicate emotions in the foreground. The entire image is imbued with a tranquil, soulful atmosphere of nostalgia, as if time has stood still in this moment.", + "seed": 10643, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_10/index_2467_seed_10643_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.060046311002224684, + "vlm_logprob": -0.060046311002224684, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.060046311002224684, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.185354545712471 + ], + "opposite_logprob_tokens": [ + -1.8103545904159546 + ], + "logprob_reward": -0.185354545712471, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.185354545712471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13855372369289398, + 0.05247202888131142, + 0.015478365123271942, + 0.014664987102150917, + 0.09123895317316055, + 0.13097742199897766, + 0.16843225061893463, + 0.07132512331008911, + 0.2420678287744522, + 0.07478924840688705 + ], + "credit_map": [ + 0.016207056120038033, + 0.028196638450026512, + 0.01975063793361187, + 0.01907135359942913, + 0.020258920267224312, + 0.01408738736063242, + 0.009295978583395481, + 0.011685756035149097, + 0.00630250433459878, + 0.01119215041399002, + 0.00608195411041379, + 0.005574052222073078, + 0.010508614592254162, + 0.006580832879990339, + 0.003525847801938653, + 0.002706073457375169, + 0.002277584280818701, + 0.0037833331152796745, + 0.0020946403965353966, + 0.0020440020598471165, + 0.0015046123880892992, + 0.0015129217645153403, + 0.001129266107454896, + 0.0011320046614855528, + 0.001723259687423706, + 0.0033949860371649265, + 0.002241478068754077, + 0.0014186716871336102, + 0.0020025677513331175, + 0.0014345800736919045, + 0.0012142375344410539, + 0.0012352059129625559, + 0.011459297500550747, + 0.01728069968521595, + 0.010354449972510338, + 0.010489046573638916, + 0.008306379429996014, + 0.008268455043435097, + 0.01612047478556633, + 0.008960156701505184, + 0.018811291083693504, + 0.03936241567134857, + 0.010675271041691303, + 0.01303008385002613, + 0.026425186544656754, + 0.011286740191280842, + 0.006987959612160921, + 0.004398486111313105, + 0.029504911974072456, + 0.058307405561208725, + 0.015330861322581768, + 0.008802044205367565, + 0.02823738008737564, + 0.013550191186368465, + 0.007272400893270969, + 0.007427061442285776, + 0.009552828967571259, + 0.017138749361038208, + 0.007709426339715719, + 0.006572735495865345, + 0.013080107048153877, + 0.006795821711421013, + 0.004892739001661539, + 0.005582709331065416, + 0.045549456030130386, + 0.0920385792851448, + 0.0076823765411973, + 0.00822249986231327, + 0.05606383457779884, + 0.016905948519706726, + 0.006211938802152872, + 0.009393204003572464, + 0.011119852773845196, + 0.012650723569095135, + 0.008055860176682472, + 0.007208783645182848, + 0.012987310066819191, + 0.011071321554481983, + 0.006574664264917374, + 0.00512072816491127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.185354545712471 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797580540180206 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.009797580540180206, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797580540180206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17542526125907898, + 0.048292454332113266, + 0.04616382718086243, + 0.02802908979356289, + 0.20949117839336395, + 0.042705513536930084, + 0.08725431561470032, + 0.022849399596452713, + 0.2894101142883301, + 0.050378911197185516 + ], + "credit_map": [ + 0.02750224620103836, + 0.03556935116648674, + 0.022843481972813606, + 0.02158607542514801, + 0.021697884425520897, + 0.018526019528508186, + 0.013829515315592289, + 0.013870674185454845, + 0.007224516477435827, + 0.012648345902562141, + 0.005335318855941296, + 0.005990626756101847, + 0.00556528102606535, + 0.005212975200265646, + 0.0036473500076681376, + 0.0026680361479520798, + 0.007345176301896572, + 0.009829451330006123, + 0.006869941018521786, + 0.004508059471845627, + 0.006779734045267105, + 0.003938119858503342, + 0.0035210030619055033, + 0.003372341161593795, + 0.003170133102685213, + 0.006073133088648319, + 0.0029478855431079865, + 0.0020761776249855757, + 0.006078875157982111, + 0.004490272607654333, + 0.0016733647789806128, + 0.0015192487044259906, + 0.027019724249839783, + 0.03758776932954788, + 0.023167796432971954, + 0.025483636185526848, + 0.017885901033878326, + 0.0200512558221817, + 0.03809897601604462, + 0.020196126773953438, + 0.005578317679464817, + 0.0138687239959836, + 0.003391062840819359, + 0.002664220752194524, + 0.010051830671727657, + 0.0032150843180716038, + 0.0019547182600945234, + 0.0019815578125417233, + 0.013030115514993668, + 0.02968756854534149, + 0.005215499550104141, + 0.004600673448294401, + 0.0202869214117527, + 0.006078789476305246, + 0.004232155624777079, + 0.004122593440115452, + 0.0029278863221406937, + 0.005034503526985645, + 0.0033304505050182343, + 0.002298590261489153, + 0.0029901661910116673, + 0.002030631061643362, + 0.0021252206061035395, + 0.002111950656399131, + 0.058442529290914536, + 0.1132488027215004, + 0.00968110840767622, + 0.0070769027806818485, + 0.06528770923614502, + 0.020589571446180344, + 0.007276955060660839, + 0.007806519046425819, + 0.007842703722417355, + 0.010196519084274769, + 0.004752991255372763, + 0.0044620102271437645, + 0.006548338569700718, + 0.007841350510716438, + 0.004354090895503759, + 0.0043809120543301105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.009797580540180206 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328656055033207 + ], + "opposite_logprob_tokens": [ + -3.7732865810394287 + ], + "logprob_reward": -0.02328656055033207, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328656055033207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21888768672943115, + 0.0649852529168129, + 0.08021577447652817, + 0.05373189598321915, + 0.25707724690437317, + 0.04753643646836281, + 0.06264090538024902, + 0.0572734959423542, + 0.07032440602779388, + 0.08732698857784271 + ], + "credit_map": [ + 0.03264835104346275, + 0.03395422175526619, + 0.03542985022068024, + 0.030264131724834442, + 0.0229590255767107, + 0.017820779234170914, + 0.018994364887475967, + 0.026816941797733307, + 0.00996550265699625, + 0.011875299736857414, + 0.011493559926748276, + 0.009005974046885967, + 0.0056501151993870735, + 0.005570732522755861, + 0.006658941973000765, + 0.004765118006616831, + 0.012612237595021725, + 0.015788914635777473, + 0.012898743152618408, + 0.007505510933697224, + 0.011174315586686134, + 0.0063504925929009914, + 0.008017896674573421, + 0.005867650732398033, + 0.006450332701206207, + 0.008751878514885902, + 0.007840212434530258, + 0.00966732483357191, + 0.004172688350081444, + 0.003018945688381791, + 0.006717817857861519, + 0.007112687919288874, + 0.032191772013902664, + 0.04498174414038658, + 0.0296774972230196, + 0.02934892661869526, + 0.022489488124847412, + 0.026802456006407738, + 0.04606649652123451, + 0.025518840178847313, + 0.006224656943231821, + 0.01007880549877882, + 0.007224798668175936, + 0.004915344528853893, + 0.005989233031868935, + 0.00445598503574729, + 0.004613049328327179, + 0.0040345573797822, + 0.010055560618638992, + 0.017607426270842552, + 0.0062746312469244, + 0.004287439864128828, + 0.012279518879950047, + 0.004685681778937578, + 0.00391732482239604, + 0.003533310955390334, + 0.009405764751136303, + 0.008685034699738026, + 0.008680776692926884, + 0.0073544178158044815, + 0.005680815316736698, + 0.005052889231592417, + 0.004961192142218351, + 0.007452595978975296, + 0.008878333494067192, + 0.011472607031464577, + 0.01190039049834013, + 0.012286422774195671, + 0.005721190012991428, + 0.007317402400076389, + 0.004841179121285677, + 0.007906878367066383, + 0.01339437160640955, + 0.016712171956896782, + 0.005931803025305271, + 0.004506197292357683, + 0.0210547037422657, + 0.012687595561146736, + 0.008905479684472084, + 0.004134662449359894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.02328656055033207 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897313684225082 + ], + "opposite_logprob_tokens": [ + -2.5789730548858643 + ], + "logprob_reward": -0.07897313684225082, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897313684225082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03564062714576721, + 0.038383252918720245, + 0.015379661694169044, + 0.014830829575657845, + 0.13485468924045563, + 0.025908352807164192, + 0.02973959594964981, + 0.054432567209005356, + 0.6242133975028992, + 0.02661694958806038 + ], + "credit_map": [ + 0.004378225654363632, + 0.006075651850551367, + 0.004484676290303469, + 0.005473586264997721, + 0.0033992540556937456, + 0.0036366593558341265, + 0.003764370921999216, + 0.004428205080330372, + 0.003780882339924574, + 0.006712895818054676, + 0.005680830217897892, + 0.006168352905660868, + 0.005109559278935194, + 0.0052583771757781506, + 0.0029134347569197416, + 0.002758919959887862, + 0.002082184189930558, + 0.0032764431089162827, + 0.0021866902243345976, + 0.0019275194499641657, + 0.001823651953600347, + 0.001375764375552535, + 0.001539182965643704, + 0.0011682268232107162, + 0.0016254732618108392, + 0.002725237049162388, + 0.0019723298028111458, + 0.001408182317391038, + 0.0022395469713956118, + 0.0017703338526189327, + 0.0014416939811781049, + 0.0016480335034430027, + 0.01699751615524292, + 0.023175004869699478, + 0.017782699316740036, + 0.01332949660718441, + 0.012616139836609364, + 0.015319498255848885, + 0.02212190441787243, + 0.01351243443787098, + 0.003130663186311722, + 0.006755034904927015, + 0.0029196091927587986, + 0.0030097882263362408, + 0.003955740015953779, + 0.002204537857323885, + 0.0026778618339449167, + 0.0012551182880997658, + 0.004016609862446785, + 0.006672580260783434, + 0.0037363520823419094, + 0.002951397793367505, + 0.005065238103270531, + 0.0023901727981865406, + 0.0021392530761659145, + 0.002767994999885559, + 0.00850762240588665, + 0.007235032506287098, + 0.006817386951297522, + 0.010376510210335255, + 0.00497918576002121, + 0.0032922907266765833, + 0.005714047234505415, + 0.007510493975132704, + 0.119841568171978, + 0.24480460584163666, + 0.014682823792099953, + 0.014366138726472855, + 0.1519654542207718, + 0.043820030987262726, + 0.015020671300590038, + 0.019712112843990326, + 0.0038554382044821978, + 0.004086378030478954, + 0.0032816564198583364, + 0.003985940478742123, + 0.0024729224387556314, + 0.002839075168594718, + 0.0024654387962073088, + 0.003630101215094328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.07897313684225082 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819731365889311 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819731365889311, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819731365889311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07591187953948975, + 0.03467358276247978, + 0.017526526004076004, + 0.018202438950538635, + 0.18590642511844635, + 0.045691922307014465, + 0.07454536855220795, + 0.05307469144463539, + 0.4394237995147705, + 0.055043384432792664 + ], + "credit_map": [ + 0.00908893533051014, + 0.015614661388099194, + 0.012266328558325768, + 0.007688894867897034, + 0.008038055151700974, + 0.00772456731647253, + 0.007609057240188122, + 0.007881383411586285, + 0.005114697851240635, + 0.006532290019094944, + 0.005084846168756485, + 0.00455419160425663, + 0.003453379962593317, + 0.003578027244657278, + 0.003640999086201191, + 0.002715152921155095, + 0.0020656839478760958, + 0.003627705853432417, + 0.0026669693179428577, + 0.0018918877467513084, + 0.003058947855606675, + 0.0018916947301477194, + 0.001050880178809166, + 0.0012727583525702357, + 0.002099091187119484, + 0.004444495774805546, + 0.0024180642794817686, + 0.0014156081015244126, + 0.002852286444976926, + 0.0022911345586180687, + 0.001298087532632053, + 0.0013836727011948824, + 0.023119457066059113, + 0.03240353241562843, + 0.023694386705756187, + 0.02054750546813011, + 0.016036510467529297, + 0.02016877941787243, + 0.03156514838337898, + 0.018371107056736946, + 0.005729556083679199, + 0.01140380185097456, + 0.004778921138495207, + 0.004753194283694029, + 0.008162742480635643, + 0.005067598074674606, + 0.0034064631909132004, + 0.0023896496277302504, + 0.012392478063702583, + 0.02687741070985794, + 0.004978657700121403, + 0.003601707285270095, + 0.015839116647839546, + 0.006020865868777037, + 0.0024888694752007723, + 0.002346268855035305, + 0.007131653372198343, + 0.008741665631532669, + 0.007008029147982597, + 0.007592132780700922, + 0.0068993475288152695, + 0.004846316762268543, + 0.004678709432482719, + 0.00617683632299304, + 0.0849192887544632, + 0.17353932559490204, + 0.012488476000726223, + 0.009349383413791656, + 0.10863227397203445, + 0.03030174784362316, + 0.00902360025793314, + 0.01116975024342537, + 0.00840627308934927, + 0.007421912159770727, + 0.008994423784315586, + 0.006350270472466946, + 0.006578353233635426, + 0.008332882076501846, + 0.004183514509350061, + 0.004775759298354387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.002819731365889311 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, black-and-white close-up shot from a slightly low angle, dominated by high-contrast tones. The focus falls precisely on an elderly hand, its wrinkled skin dry and sagging. The ridges and veins on the back of the hand, etched with age, are clearly outlined in the soft sidelight, the knuckles appearing thick. This hand rests on an old, yellowed photograph, its edges warped and scuffed from repeated handling, its surface minutely scratched. Carefully and with infinite cherishment, the elderly man's rough fingertips trace the worn white border of the photograph, slowly and steadily sliding from one end to the other. The gentle movement is as if he is revisiting the distant memories the photograph holds through touch. The background is an out-of-focus, dark wooden tabletop, its blurred texture providing a quiet contrast to the delicate emotions in the foreground. The entire image is imbued with a tranquil, soulful atmosphere of nostalgia, as if time has stood still in this moment.", + "seed": 10644, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_10/index_2467_seed_10644_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06805367516353726, + "vlm_logprob": -0.06805367516353726, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.06805367516353726, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1889408975839615 + ], + "opposite_logprob_tokens": [ + -1.8139408826828003 + ], + "logprob_reward": -0.1889408975839615, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1889408975839615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2107921689748764, + 0.018415285274386406, + 0.04707586020231247, + 0.09200090914964676, + 0.01693408377468586, + 0.4224984347820282, + 0.03606903925538063, + 0.08908257633447647, + 0.02320130728185177, + 0.0439302921295166 + ], + "credit_map": [ + 0.02599763683974743, + 0.025324532762169838, + 0.027271633967757225, + 0.030255036428570747, + 0.025855643674731255, + 0.02125520259141922, + 0.029955141246318817, + 0.024877337738871574, + 0.0018820881377905607, + 0.002048908732831478, + 0.0023872649762779474, + 0.0030198702588677406, + 0.001749309478327632, + 0.0025422037579119205, + 0.001959774410352111, + 0.0028258641250431538, + 0.00423314468935132, + 0.005647817626595497, + 0.006371303461492062, + 0.006997617892920971, + 0.006668094079941511, + 0.006675885058939457, + 0.0057535553351044655, + 0.004728443454951048, + 0.012491460889577866, + 0.010958063416182995, + 0.012920159846544266, + 0.013292467221617699, + 0.00839991308748722, + 0.011343171820044518, + 0.016460562124848366, + 0.006135109346359968, + 0.0013771889498457313, + 0.0013992490712553263, + 0.0034139566123485565, + 0.0014722166815772653, + 0.0017191612860187888, + 0.0036917130928486586, + 0.0026854067109525204, + 0.001175189740024507, + 0.024437634274363518, + 0.012389246374368668, + 0.1530883014202118, + 0.01867780089378357, + 0.012490606866776943, + 0.1797230839729309, + 0.008299164474010468, + 0.013392604887485504, + 0.002879219362512231, + 0.003747850889340043, + 0.005515876691788435, + 0.004949195310473442, + 0.00549275754019618, + 0.005655215587466955, + 0.00508410157635808, + 0.0027448239270597696, + 0.007267882116138935, + 0.007525255437940359, + 0.010874273255467415, + 0.01908273808658123, + 0.011732916347682476, + 0.008913509547710419, + 0.014560991898179054, + 0.009125006385147572, + 0.0019954415038228035, + 0.003052501007914543, + 0.003019791329279542, + 0.0037370966747403145, + 0.0028838571161031723, + 0.0027302801609039307, + 0.004052175674587488, + 0.0017301643965765834, + 0.004317426588386297, + 0.0064169419929385185, + 0.005155675113201141, + 0.009480089880526066, + 0.005263919476419687, + 0.0043883537873625755, + 0.005832439288496971, + 0.0030754448380321264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.1889408975839615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012561723589897156 + ], + "opposite_logprob_tokens": [ + -4.387561798095703 + ], + "logprob_reward": -0.012561723589897156, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012561723589897156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07557327300310135, + 0.023373940959572792, + 0.029058266431093216, + 0.04199879243969917, + 0.038906585425138474, + 0.6540273427963257, + 0.051784444600343704, + 0.0203134436160326, + 0.02032379060983658, + 0.04464013874530792 + ], + "credit_map": [ + 0.009276816621422768, + 0.009958063252270222, + 0.010129927657544613, + 0.010171473026275635, + 0.008841740898787975, + 0.006604997906833887, + 0.011413327418267727, + 0.009176922962069511, + 0.0022808623034507036, + 0.0026478159707039595, + 0.003033752553164959, + 0.0036228506360203028, + 0.0021146193612366915, + 0.0025543002411723137, + 0.0033800718374550343, + 0.003739667357876897, + 0.0031495431903749704, + 0.003495030803605914, + 0.004112419206649065, + 0.0037833331152796745, + 0.003660803195089102, + 0.0032912467140704393, + 0.003594788955524564, + 0.0039711007848382, + 0.004862514324486256, + 0.003763014916330576, + 0.0070454771630465984, + 0.007068506442010403, + 0.003928497899323702, + 0.005047755315899849, + 0.0071359011344611645, + 0.0031471282709389925, + 0.0026793600991368294, + 0.0027886698953807354, + 0.010479278862476349, + 0.0028265996370464563, + 0.0024089093785732985, + 0.010988254100084305, + 0.005013887770473957, + 0.0017216262640431523, + 0.037419144064188004, + 0.018726790323853493, + 0.23753486573696136, + 0.02757527492940426, + 0.019813938066363335, + 0.2790548801422119, + 0.013012482784688473, + 0.020889965817332268, + 0.003740538377314806, + 0.004839407280087471, + 0.008622958324849606, + 0.0062416922301054, + 0.007462318986654282, + 0.011842278763651848, + 0.005419949535280466, + 0.003615298541262746, + 0.0021181555930525064, + 0.0022285415325313807, + 0.0031471571419388056, + 0.0037689553573727608, + 0.002069673966616392, + 0.0027204561047255993, + 0.0024030953645706177, + 0.001857408438809216, + 0.0019362624734640121, + 0.0025460473261773586, + 0.002985857892781496, + 0.0050996290519833565, + 0.0018310032319277525, + 0.0016172229079529643, + 0.002299650339409709, + 0.00200811680406332, + 0.005237778183072805, + 0.003921416588127613, + 0.008769584819674492, + 0.005671232007443905, + 0.00431028800085187, + 0.004660612903535366, + 0.0076093487441539764, + 0.004459876101464033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.012561723589897156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058943174779415 + ], + "opposite_logprob_tokens": [ + -3.8955893516540527 + ], + "logprob_reward": -0.02058943174779415, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058943174779415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050062716007232666, + 0.019130535423755646, + 0.010441907681524754, + 0.05280173197388649, + 0.26949945092201233, + 0.24197961390018463, + 0.20537936687469482, + 0.030483389273285866, + 0.0562974214553833, + 0.06392385810613632 + ], + "credit_map": [ + 0.005334763787686825, + 0.005611879751086235, + 0.0076162731274962425, + 0.007708940654993057, + 0.005269673187285662, + 0.004218870773911476, + 0.008449270389974117, + 0.005853050388395786, + 0.0021451886277645826, + 0.0026549361646175385, + 0.0023588354233652353, + 0.0033690123818814754, + 0.0017505405703559518, + 0.0025075548328459263, + 0.002067022258415818, + 0.0022774464450776577, + 0.0010842358460649848, + 0.0008861691458150744, + 0.0017496797954663634, + 0.0015860141720622778, + 0.0011590735521167517, + 0.0014013488544151187, + 0.0015792731428518891, + 0.00099611422047019, + 0.006456343922764063, + 0.006057373713701963, + 0.008301609195768833, + 0.007910488173365593, + 0.0048134238459169865, + 0.005835549905896187, + 0.008547969162464142, + 0.004878976382315159, + 0.017149485647678375, + 0.010087470524013042, + 0.09026392549276352, + 0.014114201068878174, + 0.010385874658823013, + 0.10306378453969955, + 0.015326708555221558, + 0.009108021855354309, + 0.013803423382341862, + 0.00683471467345953, + 0.08557815849781036, + 0.01051370706409216, + 0.009023457765579224, + 0.1002841666340828, + 0.007456633727997541, + 0.008485349826514721, + 0.016984477639198303, + 0.020829342305660248, + 0.030476167798042297, + 0.030247271060943604, + 0.03585563227534294, + 0.03323538973927498, + 0.02266327664256096, + 0.015087817795574665, + 0.002395524410530925, + 0.0025775558315217495, + 0.004092428833246231, + 0.00564868887886405, + 0.003528372384607792, + 0.0033629550598561764, + 0.005753489676862955, + 0.0031243781559169292, + 0.0032932821195572615, + 0.009581662714481354, + 0.006937507539987564, + 0.006534289103001356, + 0.00804985873401165, + 0.005278815049678087, + 0.008360318839550018, + 0.008261692710220814, + 0.0057996646501123905, + 0.009136179462075233, + 0.008163602091372013, + 0.009395990520715714, + 0.007083696313202381, + 0.0074170553125441074, + 0.006911847740411758, + 0.010015822015702724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.02058943174779415 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292276531457901 + ], + "opposite_logprob_tokens": [ + -2.2379226684570312 + ], + "logprob_reward": -0.11292276531457901, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292276531457901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.011278001591563225, + 0.008414660580456257, + 0.012448737397789955, + 0.013889928348362446, + 0.05888834223151207, + 0.852272093296051, + 0.008483578450977802, + 0.016569364815950394, + 0.012308210134506226, + 0.005447077099233866 + ], + "credit_map": [ + 0.0010993772884830832, + 0.0015956611605361104, + 0.0017501972615718842, + 0.0014834253815934062, + 0.0011480427347123623, + 0.0012063764734193683, + 0.0013176323845982552, + 0.0016772900708019733, + 0.0008578011183999479, + 0.000917288358323276, + 0.0011136182583868504, + 0.0011297835735604167, + 0.000815133040305227, + 0.0013642871053889394, + 0.0011819859500974417, + 0.0010347634088248014, + 0.0010121410014107823, + 0.001159673323854804, + 0.0017787703545764089, + 0.0017579804407432675, + 0.0017466142307966948, + 0.0019801678135991096, + 0.0016044218791648746, + 0.0014089688193053007, + 0.0018843115540221334, + 0.0014504499267786741, + 0.0021899843122810125, + 0.002046627691015601, + 0.0013867935631424189, + 0.0014301332412287593, + 0.002243467839434743, + 0.0012581611517816782, + 0.003812758019194007, + 0.002296583727002144, + 0.01982693001627922, + 0.0029146303422749043, + 0.0020631796214729548, + 0.022444529458880424, + 0.00345971854403615, + 0.0020700127352029085, + 0.04921954497694969, + 0.024786539375782013, + 0.3097781240940094, + 0.037589263170957565, + 0.024795137345790863, + 0.3629947304725647, + 0.01607588119804859, + 0.027032924816012383, + 0.0006677378551103175, + 0.0007414928404614329, + 0.0014136956306174397, + 0.0015638185432180762, + 0.0011389236897230148, + 0.001252993126399815, + 0.0008027405710890889, + 0.0009021764271892607, + 0.0016257623210549355, + 0.0019832178950309753, + 0.002233949489891529, + 0.002609569812193513, + 0.0016382980393245816, + 0.002185993129387498, + 0.002918879734352231, + 0.0013736957916989923, + 0.001327458187006414, + 0.001477850484661758, + 0.0023684995248913765, + 0.0014727122616022825, + 0.0010901519563049078, + 0.001602331642061472, + 0.0013425409561023116, + 0.0016266644233837724, + 0.000725801452063024, + 0.000414115289459005, + 0.001309522776864469, + 0.0009126347140409052, + 0.0004996994975954294, + 0.00044363088090904057, + 0.0006588191026821733, + 0.0004828529490623623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.11292276531457901 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in black-and-white with high-contrast tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an elderly hand with wrinkled, dry, and sagging skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand resting on an old, yellowed photograph with warped edges and surface scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand slowly and steadily trace the white border of the photograph from one end to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly low angle with a close-up framing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052535575814545155 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.0052535575814545155, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052535575814545155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058478131890296936, + 0.02859262190759182, + 0.026441490277647972, + 0.0585864819586277, + 0.16071759164333344, + 0.41201895475387573, + 0.1598348170518875, + 0.054228298366069794, + 0.02244224213063717, + 0.01865937002003193 + ], + "credit_map": [ + 0.007662961725145578, + 0.007765383925288916, + 0.0075793215073645115, + 0.006768268067389727, + 0.00627501355484128, + 0.005122687667608261, + 0.00806909054517746, + 0.009235406294465065, + 0.004291254561394453, + 0.002581472508609295, + 0.004562117159366608, + 0.00358305755071342, + 0.002002069028094411, + 0.002954470692202449, + 0.003213342744857073, + 0.00540483882650733, + 0.0023269180674105883, + 0.0030795461498200893, + 0.004129049368202686, + 0.004002682864665985, + 0.003006892278790474, + 0.004056024365127087, + 0.0030941106379032135, + 0.0027462663128972054, + 0.005896872840821743, + 0.0059858993627130985, + 0.008280027657747269, + 0.009326498955488205, + 0.0056278579868376255, + 0.009332680143415928, + 0.010714245960116386, + 0.00342240440659225, + 0.010363796725869179, + 0.006296807900071144, + 0.054369114339351654, + 0.008286290802061558, + 0.0057595656253397465, + 0.06080351024866104, + 0.009372706525027752, + 0.005465799476951361, + 0.02345540188252926, + 0.012197447940707207, + 0.14972713589668274, + 0.017347609624266624, + 0.012394423596560955, + 0.17568060755729675, + 0.00819579791277647, + 0.013020548969507217, + 0.014556271024048328, + 0.01640358753502369, + 0.023999737575650215, + 0.022996382787823677, + 0.022848350927233696, + 0.028704609721899033, + 0.017403636127710342, + 0.012922246009111404, + 0.00448314705863595, + 0.005874216090887785, + 0.008168053813278675, + 0.007431515492498875, + 0.0060669430531561375, + 0.0064891609363257885, + 0.011038700118660927, + 0.0046765608713030815, + 0.0019531125435605645, + 0.004623574670404196, + 0.002192756161093712, + 0.002665227511897683, + 0.002558027161285281, + 0.0021412381902337074, + 0.0033346775453537703, + 0.002973626833409071, + 0.0022269152104854584, + 0.0019464063225314021, + 0.0032839032355695963, + 0.003008214058354497, + 0.0017763017676770687, + 0.001807135995477438, + 0.0027240393683314323, + 0.001886454876512289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.0052535575814545155 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_11/index_tensor([7420, 7420])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_11/index_tensor([7420, 7420])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4a0264cb4129dc06cf02cd5838cbcd0956529598 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_11/index_tensor([7420, 7420])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro shot, captured from a slight overhead angle, maintains a still frame, focusing squarely on the intricate texture of a pristine white cotton pad, set against a deep, non-reflective, dark background. A bright, precise, clinical overhead light illuminates the scene, bringing each soft, interwoven fiber of the pad into sharp focus, casting subtle shadows and dramatically enhancing its three-dimensional quality. A drop of crystal-clear, slightly viscous disinfectant begins to trickle cautiously onto the pad's surface, capturing the microscopic process of the liquid soaking into the fibers in astonishing detail: the clear liquid's front advances steadily and deliberately, instantly saturating the dry, fluffy white fibers wherever it passes, turning them slightly darker and translucent. As the liquid continues to spread silently, the entire picture seamlessly transitions to a final static picture through a smooth and flowing dissolve shot - at this point, the entire cotton piece has been completely and evenly wetted, the fibers are tightly attached together, and they shimmer with a wet luster under the light, creating an extremely clean, focused and tranquil atmosphere with a scientific beauty.", + "seed": 33079, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_11/index_7420_seed_33079_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05852806307375431, + "vlm_logprob": -0.05852806307375431, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.05852806307375431, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07666224241256714 + ], + "opposite_logprob_tokens": [ + -2.701662302017212 + ], + "logprob_reward": -0.07666224241256714, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07666224241256714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14012302458286285, + 0.12199516594409943, + 0.24909809231758118, + 0.047014910727739334, + 0.08520819991827011, + 0.07764443755149841, + 0.050063591450452805, + 0.09892390668392181, + 0.09182355552911758, + 0.038105204701423645 + ], + "credit_map": [ + 0.011305680498480797, + 0.021547600626945496, + 0.018826911225914955, + 0.019877487793564796, + 0.02072260156273842, + 0.011657698079943657, + 0.02150324173271656, + 0.014681781642138958, + 0.011941657401621342, + 0.02292616292834282, + 0.011640035547316074, + 0.015287268906831741, + 0.01959470845758915, + 0.012163660489022732, + 0.017835957929491997, + 0.010605700314044952, + 0.02143673226237297, + 0.05987498164176941, + 0.01731340028345585, + 0.023036973550915718, + 0.05769675597548485, + 0.018456995487213135, + 0.023106519132852554, + 0.0281757153570652, + 0.004464468453079462, + 0.006467884872108698, + 0.004848368465900421, + 0.00689838919788599, + 0.006890971213579178, + 0.005332139320671558, + 0.007003098260611296, + 0.005109584890305996, + 0.004296466708183289, + 0.013560709543526173, + 0.005928360391408205, + 0.01058373972773552, + 0.012723141349852085, + 0.009105042554438114, + 0.01995420642197132, + 0.009056526236236095, + 0.008503170683979988, + 0.01030401699244976, + 0.006971023045480251, + 0.010371851734817028, + 0.009612932801246643, + 0.015719138085842133, + 0.008076198399066925, + 0.008086098358035088, + 0.005748600233346224, + 0.007555002346634865, + 0.003741151886060834, + 0.0055022151209414005, + 0.007487919181585312, + 0.009815851226449013, + 0.004706420004367828, + 0.005506429821252823, + 0.009776465594768524, + 0.014778036624193192, + 0.009695064276456833, + 0.013309420086443424, + 0.012028297409415245, + 0.01806025765836239, + 0.008392004296183586, + 0.012884345836937428, + 0.008330614306032658, + 0.01303957961499691, + 0.009423721581697464, + 0.009385312907397747, + 0.016234587877988815, + 0.011972207576036453, + 0.010868934914469719, + 0.012568587437272072, + 0.005350728053599596, + 0.005614123307168484, + 0.0038582254201173782, + 0.0042011491023004055, + 0.005247538909316063, + 0.005655061919242144, + 0.0037133730947971344, + 0.004465001169592142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.07666224241256714 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901114225387573 + ], + "opposite_logprob_tokens": [ + -2.5790112018585205 + ], + "logprob_reward": -0.07901114225387573, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901114225387573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2109927535057068, + 0.031034961342811584, + 0.19314275681972504, + 0.13814736902713776, + 0.1118299588561058, + 0.19490325450897217, + 0.03753701224923134, + 0.031683772802352905, + 0.03777129203081131, + 0.012956860475242138 + ], + "credit_map": [ + 0.01699819415807724, + 0.02952706068754196, + 0.02626391500234604, + 0.030628163367509842, + 0.030070550739765167, + 0.01677711494266987, + 0.03778887540102005, + 0.022938886657357216, + 0.00251235393807292, + 0.005837898701429367, + 0.003547082422301173, + 0.003707410302013159, + 0.0047165462747216225, + 0.00285098678432405, + 0.0048484536819159985, + 0.0030142278410494328, + 0.008929642848670483, + 0.05583104118704796, + 0.010514256544411182, + 0.009637110866606236, + 0.06293883174657822, + 0.01288105733692646, + 0.0142156220972538, + 0.018195170909166336, + 0.017044883221387863, + 0.014035732485353947, + 0.01071971282362938, + 0.01584445685148239, + 0.01636817865073681, + 0.010131622664630413, + 0.02534041367471218, + 0.028662361204624176, + 0.006794427055865526, + 0.01511791255325079, + 0.010480045340955257, + 0.023900285363197327, + 0.013834795914590359, + 0.01171803567558527, + 0.015117905102670193, + 0.014866547659039497, + 0.021533697843551636, + 0.020292703062295914, + 0.016893766820430756, + 0.02337563969194889, + 0.022532006725668907, + 0.03632555156946182, + 0.02546703815460205, + 0.028482865542173386, + 0.0032855705358088017, + 0.005097340792417526, + 0.0031815259717404842, + 0.0048531475476920605, + 0.005105653777718544, + 0.005017952527850866, + 0.006835819222033024, + 0.004160003270953894, + 0.001831629779189825, + 0.004068170208483934, + 0.0033074927050620317, + 0.005023985635489225, + 0.004920692183077335, + 0.0044968584552407265, + 0.004029680974781513, + 0.004005263093858957, + 0.0033675790764391422, + 0.0050913384184241295, + 0.0025826713535934687, + 0.0038633421063423157, + 0.006748291198164225, + 0.004650412127375603, + 0.004282549023628235, + 0.0071851094253361225, + 0.00142784439958632, + 0.002040209248661995, + 0.0018444627057760954, + 0.0015963909681886435, + 0.00134176854044199, + 0.0019260329427197576, + 0.001166331465356052, + 0.0016138211358338594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.07901114225387573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075875610113144 + ], + "opposite_logprob_tokens": [ + -4.14107608795166 + ], + "logprob_reward": -0.016075875610113144, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075875610113144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12024682760238647, + 0.02411714568734169, + 0.1527559459209442, + 0.19043125212192535, + 0.22117556631565094, + 0.04213489964604378, + 0.08234348893165588, + 0.09074441343545914, + 0.05208534747362137, + 0.023965131491422653 + ], + "credit_map": [ + 0.010269002057611942, + 0.022879241034388542, + 0.013861363753676414, + 0.0143998172134161, + 0.01852116361260414, + 0.011001759208738804, + 0.014322485774755478, + 0.014991994015872478, + 0.002064255764707923, + 0.004664205946028233, + 0.002467565005645156, + 0.0028639952652156353, + 0.003952088765799999, + 0.0021187588572502136, + 0.0033552730455994606, + 0.0026310011744499207, + 0.005394525360316038, + 0.05281814932823181, + 0.006225018296390772, + 0.008900073356926441, + 0.05685608088970184, + 0.005750060081481934, + 0.008607516996562481, + 0.008204522542655468, + 0.01517496444284916, + 0.03076300583779812, + 0.015573459677398205, + 0.021835779771208763, + 0.031218213960528374, + 0.026307718828320503, + 0.020728113129734993, + 0.028829991817474365, + 0.015557521022856236, + 0.029383119195699692, + 0.014632473699748516, + 0.05391345173120499, + 0.030218632891774178, + 0.023308103904128075, + 0.03679230436682701, + 0.017369963228702545, + 0.004212039522826672, + 0.006881706416606903, + 0.002878570929169655, + 0.005144977942109108, + 0.006067761220037937, + 0.007876128889620304, + 0.00391368055716157, + 0.0051600332371890545, + 0.007453920319676399, + 0.010700958780944347, + 0.006542397662997246, + 0.010605075396597385, + 0.011569940485060215, + 0.010185866616666317, + 0.015508780255913734, + 0.009776553139090538, + 0.006612610071897507, + 0.016582660377025604, + 0.007628168910741806, + 0.0089056845754385, + 0.01300783734768629, + 0.016384385526180267, + 0.007654184475541115, + 0.013968885876238346, + 0.0059263259172439575, + 0.009720714762806892, + 0.0052488455548882484, + 0.004318539518862963, + 0.010740193538367748, + 0.007437515538185835, + 0.0034394836984574795, + 0.0052537331357598305, + 0.003392900340259075, + 0.004048035014420748, + 0.0031090225093066692, + 0.0025842536706477404, + 0.0034338824916630983, + 0.0031950725242495537, + 0.002034246688708663, + 0.0021677196491509676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.016075875610113144 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031096637248993 + ], + "opposite_logprob_tokens": [ + -2.350311040878296 + ], + "logprob_reward": -0.10031096637248993, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031096637248993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05870451033115387, + 0.03209493309259415, + 0.25469279289245605, + 0.2919650375843048, + 0.12211766839027405, + 0.08324676752090454, + 0.047308698296546936, + 0.0381876565515995, + 0.053700026124715805, + 0.01798188127577305 + ], + "credit_map": [ + 0.0064519126899540424, + 0.008129661902785301, + 0.006581690162420273, + 0.008486952632665634, + 0.008245664648711681, + 0.005162643734365702, + 0.009733179584145546, + 0.005912803113460541, + 0.0029334065038710833, + 0.004913276992738247, + 0.0034336016979068518, + 0.003840309800580144, + 0.0048220157623291016, + 0.003807811066508293, + 0.004549023229628801, + 0.003795489203184843, + 0.011701776646077633, + 0.07343657314777374, + 0.01680276356637478, + 0.024065939709544182, + 0.07651276886463165, + 0.011084252968430519, + 0.029233312234282494, + 0.011855397373437881, + 0.01992383971810341, + 0.04789202660322189, + 0.02343226969242096, + 0.03427579253911972, + 0.04911729320883751, + 0.045680686831474304, + 0.0363631509244442, + 0.035279951989650726, + 0.011228659190237522, + 0.016938073560595512, + 0.013580569997429848, + 0.014770431444048882, + 0.016464978456497192, + 0.018151111900806427, + 0.016835667192935944, + 0.01414817851036787, + 0.0075160665437579155, + 0.013846266083419323, + 0.005234424490481615, + 0.012173112481832504, + 0.01201479323208332, + 0.01373845525085926, + 0.009814986027777195, + 0.008908660151064396, + 0.004085934720933437, + 0.0068600960075855255, + 0.0031267686281353235, + 0.004131427500396967, + 0.007288453169167042, + 0.01321579609066248, + 0.003885398618876934, + 0.004714823793619871, + 0.0034514986909925938, + 0.00510023720562458, + 0.003614185145124793, + 0.005262819584459066, + 0.00533121544867754, + 0.0072569577023386955, + 0.004646125249564648, + 0.0035246184561401606, + 0.004813676699995995, + 0.007637344766408205, + 0.005813431926071644, + 0.0064735496416687965, + 0.006860787980258465, + 0.007673646789044142, + 0.00606928626075387, + 0.008358300663530827, + 0.001984394621104002, + 0.003167944261804223, + 0.0017007270362228155, + 0.0015917292330414057, + 0.0026387583930045366, + 0.002215953543782234, + 0.0019564079120755196, + 0.0027259662747383118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.10031096637248993 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058008871972561 + ], + "opposite_logprob_tokens": [ + -3.8955800533294678 + ], + "logprob_reward": -0.02058008871972561, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058008871972561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07882897555828094, + 0.01956758461892605, + 0.1429147869348526, + 0.13538429141044617, + 0.13742288947105408, + 0.05415099859237671, + 0.070103220641613, + 0.0929868146777153, + 0.18406976759433746, + 0.08457064628601074 + ], + "credit_map": [ + 0.005525283981114626, + 0.014843552373349667, + 0.008853515610098839, + 0.009742665104568005, + 0.016498947516083717, + 0.006532101426273584, + 0.008521913550794125, + 0.008310993202030659, + 0.0016701089916750789, + 0.0035797657910734415, + 0.0017577032558619976, + 0.002183786826208234, + 0.003917609341442585, + 0.0021385636646300554, + 0.0026925457641482353, + 0.0016274995869025588, + 0.004772198386490345, + 0.045035235583782196, + 0.0073698824271559715, + 0.010573668405413628, + 0.04885387048125267, + 0.006262694951146841, + 0.012037461623549461, + 0.008009771816432476, + 0.01105677243322134, + 0.02237241342663765, + 0.011836009100079536, + 0.014017293229699135, + 0.022295434027910233, + 0.019792653620243073, + 0.013262646272778511, + 0.020751075819134712, + 0.009993712417781353, + 0.020013172179460526, + 0.01000306848436594, + 0.028048263862729073, + 0.018428632989525795, + 0.016582366079092026, + 0.02346082031726837, + 0.010892840102314949, + 0.005431297235190868, + 0.0061408537440001965, + 0.004432171583175659, + 0.006301505025476217, + 0.006538409274071455, + 0.012561767362058163, + 0.006082694977521896, + 0.006662293802946806, + 0.005692790262401104, + 0.009639561176300049, + 0.005402631126344204, + 0.010327501222491264, + 0.009140035137534142, + 0.00842441525310278, + 0.013497944921255112, + 0.007978332228958607, + 0.0070872060023248196, + 0.01803596317768097, + 0.0075775813311338425, + 0.00983839575201273, + 0.013917849399149418, + 0.017622658982872963, + 0.007528682705014944, + 0.01137846801429987, + 0.021626019850373268, + 0.031247878447175026, + 0.01891380175948143, + 0.014305341988801956, + 0.035062603652477264, + 0.0277414433658123, + 0.013859299011528492, + 0.021313374862074852, + 0.010726410895586014, + 0.015304522588849068, + 0.010463329032063484, + 0.009158912114799023, + 0.014599177055060863, + 0.010663286782801151, + 0.006682717707008123, + 0.00697228591889143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.02058008871972561 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro shot, captured from a slight overhead angle, maintains a still frame, focusing squarely on the intricate texture of a pristine white cotton pad, set against a deep, non-reflective, dark background. A bright, precise, clinical overhead light illuminates the scene, bringing each soft, interwoven fiber of the pad into sharp focus, casting subtle shadows and dramatically enhancing its three-dimensional quality. A drop of crystal-clear, slightly viscous disinfectant begins to trickle cautiously onto the pad's surface, capturing the microscopic process of the liquid soaking into the fibers in astonishing detail: the clear liquid's front advances steadily and deliberately, instantly saturating the dry, fluffy white fibers wherever it passes, turning them slightly darker and translucent. As the liquid continues to spread silently, the entire picture seamlessly transitions to a final static picture through a smooth and flowing dissolve shot - at this point, the entire cotton piece has been completely and evenly wetted, the fibers are tightly attached together, and they shimmer with a wet luster under the light, creating an extremely clean, focused and tranquil atmosphere with a scientific beauty.", + "seed": 33080, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_11/index_7420_seed_33080_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1098796185106039, + "vlm_logprob": -0.1098796185106039, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.1098796185106039, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017033355310559273 + ], + "opposite_logprob_tokens": [ + -4.267033576965332 + ], + "logprob_reward": -0.017033355310559273, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017033355310559273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24277319014072418, + 0.2029530107975006, + 0.09805301576852798, + 0.06529849022626877, + 0.08133739233016968, + 0.04491201043128967, + 0.05775929242372513, + 0.11151111871004105, + 0.03324802219867706, + 0.06215445324778557 + ], + "credit_map": [ + 0.013480325229465961, + 0.07817748188972473, + 0.013269585557281971, + 0.0191095732152462, + 0.08324527740478516, + 0.009529181756079197, + 0.01805676706135273, + 0.007904999889433384, + 0.015046107582747936, + 0.051789309829473495, + 0.011693654581904411, + 0.030247611925005913, + 0.056215833872556686, + 0.01202551368623972, + 0.01750960573554039, + 0.00842536985874176, + 0.007765580900013447, + 0.021335303783416748, + 0.007038915995508432, + 0.015027492307126522, + 0.011217913590371609, + 0.005521085113286972, + 0.023593733087182045, + 0.006552995648235083, + 0.004753408953547478, + 0.011623894795775414, + 0.005181578919291496, + 0.014824291691184044, + 0.008846208453178406, + 0.005530127789825201, + 0.01010175235569477, + 0.004437230993062258, + 0.006687732879072428, + 0.011306211352348328, + 0.006292687263339758, + 0.024997763335704803, + 0.009188593365252018, + 0.004909151699393988, + 0.010779216885566711, + 0.0071760318242013454, + 0.0034646436106413603, + 0.0078043402172625065, + 0.0033585703931748867, + 0.010156536474823952, + 0.007275673560798168, + 0.0038729559164494276, + 0.005548052955418825, + 0.003431235905736685, + 0.003003123216331005, + 0.01462030503898859, + 0.003003241028636694, + 0.007371553685516119, + 0.018361078575253487, + 0.002510820282623172, + 0.005355418659746647, + 0.0035337484441697598, + 0.004266350530087948, + 0.04178279638290405, + 0.004777996335178614, + 0.008190101012587547, + 0.03831769898533821, + 0.0044968146830797195, + 0.006473096553236246, + 0.0032062726095318794, + 0.0016449486138299108, + 0.008987956680357456, + 0.002976048970595002, + 0.0033653886057436466, + 0.00809402298182249, + 0.0021339785307645798, + 0.004228324629366398, + 0.001817353069782257, + 0.004440648946911097, + 0.008123252540826797, + 0.00805890653282404, + 0.01563020795583725, + 0.008266019634902477, + 0.005225005559623241, + 0.006357983220368624, + 0.0060524288564920425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.017033355310559273 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22565752267837524 + ], + "opposite_logprob_tokens": [ + -1.6006574630737305 + ], + "logprob_reward": -0.22565752267837524, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22565752267837524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19635851681232452, + 0.10344124585390091, + 0.015975510701537132, + 0.02367187850177288, + 0.044230468571186066, + 0.03447140380740166, + 0.0144021762534976, + 0.15984727442264557, + 0.2597897946834564, + 0.1478116661310196 + ], + "credit_map": [ + 0.011927268467843533, + 0.05297960340976715, + 0.013119969516992569, + 0.019846390932798386, + 0.05894006788730621, + 0.013332696631550789, + 0.016698207706212997, + 0.009514323435723782, + 0.005682984367012978, + 0.028266696259379387, + 0.005316855851560831, + 0.011394355446100235, + 0.03135346621274948, + 0.00869936216622591, + 0.00820860080420971, + 0.004518931731581688, + 0.0018016707617789507, + 0.002349813934415579, + 0.001546630752272904, + 0.002666021464392543, + 0.002224873285740614, + 0.0011947316816076636, + 0.002884204499423504, + 0.0013075658353045583, + 0.0016287319594994187, + 0.0043128477409482, + 0.0035062513779848814, + 0.004243793897330761, + 0.003311407519504428, + 0.0015450733480975032, + 0.0034820723813027143, + 0.0016417006263509393, + 0.004046095069497824, + 0.0073003014549613, + 0.003935876302421093, + 0.011847935616970062, + 0.006560980342328548, + 0.0021399781107902527, + 0.005509180948138237, + 0.0028901253826916218, + 0.0032094535417854786, + 0.005969644989818335, + 0.003016088157892227, + 0.006586906034499407, + 0.008041071705520153, + 0.0019863429479300976, + 0.003293567104265094, + 0.0023683293256908655, + 0.0011179071152582765, + 0.0024129562079906464, + 0.0015049976063892245, + 0.002748829545453191, + 0.00276589416898787, + 0.000855252961628139, + 0.0019649204332381487, + 0.0010314204264432192, + 0.0044045341201126575, + 0.0657363310456276, + 0.004369170870631933, + 0.007736001163721085, + 0.06435543298721313, + 0.004354765173047781, + 0.005396408960223198, + 0.003494642674922943, + 0.008180702105164528, + 0.11050320416688919, + 0.0058236271142959595, + 0.009785451926290989, + 0.1034228578209877, + 0.007322145625948906, + 0.008784706704318523, + 0.005967141594737768, + 0.010723954997956753, + 0.02284148707985878, + 0.02068304643034935, + 0.026474038138985634, + 0.027287805452942848, + 0.010600137524306774, + 0.015713538974523544, + 0.013487662188708782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.22565752267837524 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206072121858597 + ], + "opposite_logprob_tokens": [ + -2.812060832977295 + ], + "logprob_reward": -0.06206072121858597, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206072121858597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24100230634212494, + 0.2747329771518707, + 0.0645836740732193, + 0.09245593845844269, + 0.04787794500589371, + 0.028378689661622047, + 0.04453372210264206, + 0.1567228138446808, + 0.02644164115190506, + 0.023270340636372566 + ], + "credit_map": [ + 0.020214125514030457, + 0.06031816080212593, + 0.01675221137702465, + 0.025892112404108047, + 0.06793934851884842, + 0.013688888400793076, + 0.023888031020760536, + 0.012309412471950054, + 0.022756658494472504, + 0.05783762410283089, + 0.02113019861280918, + 0.05608862638473511, + 0.048206817358732224, + 0.02084282785654068, + 0.03308403491973877, + 0.01478615216910839, + 0.004755305126309395, + 0.01250974740833044, + 0.0051002646796405315, + 0.006992132868617773, + 0.006449239794164896, + 0.002500843955203891, + 0.021411865949630737, + 0.004864272195845842, + 0.007160681299865246, + 0.01794780232012272, + 0.006809237878769636, + 0.020954927429556847, + 0.012735595926642418, + 0.006340124644339085, + 0.0145854027941823, + 0.005922158248722553, + 0.004403633065521717, + 0.00772243645042181, + 0.003781807143241167, + 0.013783692382276058, + 0.005643689073622227, + 0.0025152273010462523, + 0.006462339777499437, + 0.003565115388482809, + 0.0026120834518224, + 0.004899357445538044, + 0.002231657737866044, + 0.006019510794430971, + 0.005277569405734539, + 0.0020423063542693853, + 0.002885200548917055, + 0.0024110027588903904, + 0.002873081248253584, + 0.00982737448066473, + 0.0038382953498512506, + 0.006402840372174978, + 0.009190049022436142, + 0.003386409254744649, + 0.004908601753413677, + 0.004107065033167601, + 0.00588228739798069, + 0.055129267275333405, + 0.008037767373025417, + 0.01630542241036892, + 0.04803299531340599, + 0.006753738038241863, + 0.011842023581266403, + 0.004739303141832352, + 0.0020459636580199003, + 0.003952390979975462, + 0.0037341045681387186, + 0.005041264463216066, + 0.0041580586694180965, + 0.002066431799903512, + 0.0035830156411975622, + 0.0018604099750518799, + 0.0015752332983538508, + 0.004936369135975838, + 0.002478868467733264, + 0.003249381436035037, + 0.004776790272444487, + 0.00167421018704772, + 0.00280408700928092, + 0.0017753982683643699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.06206072121858597 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158183574676514 + ], + "opposite_logprob_tokens": [ + -1.7015818357467651 + ], + "logprob_reward": -0.20158183574676514, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158183574676514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25122228264808655, + 0.23558564484119415, + 0.07283686101436615, + 0.07704278826713562, + 0.04147220402956009, + 0.057358238846063614, + 0.036163490265607834, + 0.07356561720371246, + 0.11631826311349869, + 0.038434695452451706 + ], + "credit_map": [ + 0.016440967097878456, + 0.07447808235883713, + 0.01439455896615982, + 0.02051277458667755, + 0.08306646347045898, + 0.011683194898068905, + 0.02080601640045643, + 0.009840190410614014, + 0.01823917217552662, + 0.05094531178474426, + 0.016655094921588898, + 0.04295353963971138, + 0.05020752549171448, + 0.018016578629612923, + 0.02759418822824955, + 0.010974222794175148, + 0.005205275025218725, + 0.012591457925736904, + 0.006023991387337446, + 0.01193750649690628, + 0.011256061494350433, + 0.004683779086917639, + 0.01702965795993805, + 0.004109130706638098, + 0.005146485287696123, + 0.014850561507046223, + 0.007156169507652521, + 0.015526353381574154, + 0.010204315185546875, + 0.005965850315988064, + 0.012172834947705269, + 0.006020205561071634, + 0.0031828354112803936, + 0.00815457385033369, + 0.003726859111338854, + 0.010394884273409843, + 0.007164019625633955, + 0.002076581818982959, + 0.004673939198255539, + 0.0020985077135264874, + 0.00451269093900919, + 0.009564967826008797, + 0.004500240087509155, + 0.01092142891138792, + 0.008918109349906445, + 0.005887152161449194, + 0.009044396691024303, + 0.004009246826171875, + 0.0026520711835473776, + 0.00598524697124958, + 0.002803088864311576, + 0.004541253205388784, + 0.008737797848880291, + 0.003632364561781287, + 0.005407663062214851, + 0.0024039994459599257, + 0.002403320511803031, + 0.019294941797852516, + 0.0044411951676011086, + 0.010829933919012547, + 0.02155539207160473, + 0.0038525329437106848, + 0.007969561032950878, + 0.003218727884814143, + 0.003910282626748085, + 0.04888148605823517, + 0.002543800976127386, + 0.005830971524119377, + 0.04582546651363373, + 0.0032240513246506453, + 0.003486976493149996, + 0.00261521153151989, + 0.0026454601902514696, + 0.006899902131408453, + 0.005233255680650473, + 0.0042495038360357285, + 0.008628793992102146, + 0.0031636718194931746, + 0.004485568962991238, + 0.0031285365112125874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.20158183574676514 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430646575987339 + ], + "opposite_logprob_tokens": [ + -3.168064594268799 + ], + "logprob_reward": -0.0430646575987339, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430646575987339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1671290248632431, + 0.2100442349910736, + 0.10345079749822617, + 0.059369657188653946, + 0.0454215481877327, + 0.04458247497677803, + 0.04282683506608009, + 0.15221405029296875, + 0.1321931779384613, + 0.042768191546201706 + ], + "credit_map": [ + 0.011818964034318924, + 0.043774478137493134, + 0.00911775417625904, + 0.020682932808995247, + 0.05287226289510727, + 0.008373825810849667, + 0.01453046128153801, + 0.005958350375294685, + 0.01237405464053154, + 0.058745160698890686, + 0.01106791291385889, + 0.028344206511974335, + 0.0638888031244278, + 0.010788475163280964, + 0.017283717170357704, + 0.007551912218332291, + 0.008234163746237755, + 0.016190964728593826, + 0.006883479189127684, + 0.018705278635025024, + 0.012980121187865734, + 0.006303155329078436, + 0.027324078604578972, + 0.006829564925283194, + 0.004266634117811918, + 0.01021566241979599, + 0.0064537907019257545, + 0.011922424659132957, + 0.00823405385017395, + 0.0056484052911400795, + 0.008182667195796967, + 0.004446025472134352, + 0.003856819588690996, + 0.006908058654516935, + 0.003589927451685071, + 0.011902537196874619, + 0.006146530155092478, + 0.0029254406690597534, + 0.00636734813451767, + 0.0037248863372951746, + 0.0038827969692647457, + 0.007652829401195049, + 0.003506652545183897, + 0.008826478384435177, + 0.0070791286416351795, + 0.004339242819696665, + 0.0056677632965147495, + 0.0036275899037718773, + 0.0027202111668884754, + 0.009498735889792442, + 0.0036257884930819273, + 0.0065794833935797215, + 0.008100289851427078, + 0.0034803245216608047, + 0.005138007458299398, + 0.0036839977838099003, + 0.004676594398915768, + 0.06176023557782173, + 0.005034811329096556, + 0.007922658696770668, + 0.057645391672849655, + 0.004724545404314995, + 0.007078838534653187, + 0.003370999824255705, + 0.00588900875300169, + 0.049116019159555435, + 0.0061634560115635395, + 0.009796408005058765, + 0.04435855895280838, + 0.0045883809216320515, + 0.008249938488006592, + 0.004031406249850988, + 0.0022170227020978928, + 0.009090269915759563, + 0.0031158318743109703, + 0.0059195468202233315, + 0.010898293927311897, + 0.003175188321620226, + 0.005568656139075756, + 0.0027833839412778616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0430646575987339 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro shot, captured from a slight overhead angle, maintains a still frame, focusing squarely on the intricate texture of a pristine white cotton pad, set against a deep, non-reflective, dark background. A bright, precise, clinical overhead light illuminates the scene, bringing each soft, interwoven fiber of the pad into sharp focus, casting subtle shadows and dramatically enhancing its three-dimensional quality. A drop of crystal-clear, slightly viscous disinfectant begins to trickle cautiously onto the pad's surface, capturing the microscopic process of the liquid soaking into the fibers in astonishing detail: the clear liquid's front advances steadily and deliberately, instantly saturating the dry, fluffy white fibers wherever it passes, turning them slightly darker and translucent. As the liquid continues to spread silently, the entire picture seamlessly transitions to a final static picture through a smooth and flowing dissolve shot - at this point, the entire cotton piece has been completely and evenly wetted, the fibers are tightly attached together, and they shimmer with a wet luster under the light, creating an extremely clean, focused and tranquil atmosphere with a scientific beauty.", + "seed": 33081, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_11/index_7420_seed_33081_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2628777753561735, + "vlm_logprob": -0.2628777753561735, + "vlm_yes_no_margin": 1.775, + "vlm_token_credit": -0.2628777753561735, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03000969998538494 + ], + "opposite_logprob_tokens": [ + -3.6550097465515137 + ], + "logprob_reward": -0.03000969998538494, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03000969998538494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20999440550804138, + 0.13106003403663635, + 0.07725232094526291, + 0.0621996708214283, + 0.07278485596179962, + 0.023645145818591118, + 0.04032990708947182, + 0.07636544853448868, + 0.06497329473495483, + 0.24139493703842163 + ], + "credit_map": [ + 0.02772827260196209, + 0.029083337634801865, + 0.02548755146563053, + 0.04226209968328476, + 0.02642267383635044, + 0.02082931436598301, + 0.019590148702263832, + 0.01859101466834545, + 0.007636915892362595, + 0.02042919583618641, + 0.014512925408780575, + 0.014764907769858837, + 0.03556256741285324, + 0.016287660226225853, + 0.013965736143290997, + 0.007900124415755272, + 0.004992347676306963, + 0.010988854803144932, + 0.011039644479751587, + 0.011395825073122978, + 0.014698201790452003, + 0.009700466878712177, + 0.0090707466006279, + 0.00536623690277338, + 0.0041509210132062435, + 0.009457871317863464, + 0.010519566014409065, + 0.008208227343857288, + 0.01188197173178196, + 0.005516070872545242, + 0.008486246690154076, + 0.003978792577981949, + 0.006297578569501638, + 0.009810179471969604, + 0.012934690341353416, + 0.007625317666679621, + 0.01382139977067709, + 0.008113888092339039, + 0.008461595512926579, + 0.005720209795981646, + 0.001935621490702033, + 0.0035886068362742662, + 0.002885302295908332, + 0.003452667035162449, + 0.004728042054921389, + 0.002550723496824503, + 0.0026773579884320498, + 0.0018268248531967402, + 0.0049405936151742935, + 0.005942939314991236, + 0.005004859063774347, + 0.0049976771697402, + 0.0058308434672653675, + 0.0046368190087378025, + 0.0048165940679609776, + 0.004159580450505018, + 0.004933624062687159, + 0.010343439877033234, + 0.012353197671473026, + 0.009044215083122253, + 0.01302618719637394, + 0.01046151202172041, + 0.009920010343194008, + 0.006283256690949202, + 0.0045341807417571545, + 0.010259210132062435, + 0.010637928731739521, + 0.009655173867940903, + 0.012145658023655415, + 0.00655737379565835, + 0.006907090079039335, + 0.00427668122574687, + 0.01273786835372448, + 0.027548300102353096, + 0.028962310403585434, + 0.08635396510362625, + 0.027644623070955276, + 0.013077985495328903, + 0.03236697241663933, + 0.012702926993370056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.03000969998538494 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16047309339046478 + ], + "opposite_logprob_tokens": [ + -1.910473108291626 + ], + "logprob_reward": -0.16047309339046478, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16047309339046478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2434513419866562, + 0.03232256695628166, + 0.08838814496994019, + 0.013768943957984447, + 0.01604207418859005, + 0.0206573698669672, + 0.08499042689800262, + 0.04495997726917267, + 0.3677333891391754, + 0.08768577873706818 + ], + "credit_map": [ + 0.026272224262356758, + 0.04091522470116615, + 0.030908439308404922, + 0.02423991449177265, + 0.0448702834546566, + 0.035397130995988846, + 0.01699146442115307, + 0.023856664076447487, + 0.0025565633550286293, + 0.004612368065863848, + 0.002820153720676899, + 0.005431815050542355, + 0.00816852506250143, + 0.003285395447164774, + 0.00347900390625, + 0.0019687418825924397, + 0.005689116194844246, + 0.013258352875709534, + 0.014558984898030758, + 0.011742121540009975, + 0.015844877809286118, + 0.011061832308769226, + 0.009732011705636978, + 0.006500853691250086, + 0.0017851563170552254, + 0.002165020676329732, + 0.0013831302057951689, + 0.0013032887363806367, + 0.002541252411901951, + 0.0012806756421923637, + 0.001831290777772665, + 0.001479129190556705, + 0.0012760680401697755, + 0.0021967170760035515, + 0.0020395321771502495, + 0.002110895235091448, + 0.003087483812123537, + 0.0021078968420624733, + 0.0022045744117349386, + 0.0010189078748226166, + 0.0015359068056568503, + 0.003291813423857093, + 0.003719212720170617, + 0.0028356893453747034, + 0.0035263265017420053, + 0.002018337370827794, + 0.0022238038945943117, + 0.00150628003757447, + 0.005909832660108805, + 0.011366366408765316, + 0.019891494885087013, + 0.010824380442500114, + 0.012400120496749878, + 0.007952666841447353, + 0.009308489970862865, + 0.00733707332983613, + 0.0041298530995845795, + 0.007337669841945171, + 0.004914942663162947, + 0.005316075403243303, + 0.009067130275070667, + 0.005725986324250698, + 0.004915439523756504, + 0.00355288153514266, + 0.024374734610319138, + 0.053979434072971344, + 0.07305508106946945, + 0.04637230932712555, + 0.06475383788347244, + 0.03471041098237038, + 0.038756925612688065, + 0.03173067420721054, + 0.0060926261357963085, + 0.010904944501817226, + 0.01728776842355728, + 0.012366767041385174, + 0.014063415117561817, + 0.009195405058562756, + 0.010591196827590466, + 0.007183655630797148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.16047309339046478 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328692436218262 + ], + "opposite_logprob_tokens": [ + -0.7578692436218262 + ], + "logprob_reward": -0.6328692436218262, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328692436218262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07420602440834045, + 0.09889287501573563, + 0.027653872966766357, + 0.029817704111337662, + 0.031151868402957916, + 0.02869427390396595, + 0.037960175424814224, + 0.05987638607621193, + 0.09627595543861389, + 0.5154708623886108 + ], + "credit_map": [ + 0.00610905047506094, + 0.015917664393782616, + 0.00880418810993433, + 0.007791379000991583, + 0.016132181510329247, + 0.006956927943974733, + 0.006360966246575117, + 0.006133668124675751, + 0.0063361977227032185, + 0.015640350058674812, + 0.00930661428719759, + 0.00982619822025299, + 0.027233434841036797, + 0.011824231594800949, + 0.013062074780464172, + 0.005663774441927671, + 0.003243301995098591, + 0.0036413120105862617, + 0.004043008666485548, + 0.003853504080325365, + 0.004254800733178854, + 0.0026897857896983624, + 0.0027533869724720716, + 0.0031747729517519474, + 0.002242566319182515, + 0.00420739920809865, + 0.006165428087115288, + 0.003509911708533764, + 0.006003777030855417, + 0.0025919133331626654, + 0.0028213984332978725, + 0.002275308361276984, + 0.0021914748940616846, + 0.005208592861890793, + 0.002935485914349556, + 0.004880670923739672, + 0.006810348015278578, + 0.0037457600701600313, + 0.0037764091975986958, + 0.0016031268751248717, + 0.0026954265777021646, + 0.0041743675246834755, + 0.0031841867603361607, + 0.004075676202774048, + 0.005841882433742285, + 0.0030445335432887077, + 0.0034028736408799887, + 0.00227532722055912, + 0.002891650190576911, + 0.005183499306440353, + 0.006515675690025091, + 0.005830592475831509, + 0.006571378093212843, + 0.003509468398988247, + 0.0046338364481925964, + 0.00282407458871603, + 0.004452706780284643, + 0.009648186154663563, + 0.008255363442003727, + 0.006344437133520842, + 0.010665890760719776, + 0.0057096234522759914, + 0.010779262520372868, + 0.004020916763693094, + 0.007856194861233234, + 0.01389826089143753, + 0.015621015802025795, + 0.017613038420677185, + 0.012830027379095554, + 0.007800589315593243, + 0.011706031858921051, + 0.008950797840952873, + 0.029265504330396652, + 0.05939489230513573, + 0.06541097909212112, + 0.1680225133895874, + 0.06333238631486893, + 0.03160754218697548, + 0.06888443231582642, + 0.029552588239312172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.6328692436218262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42897936701774597 + ], + "opposite_logprob_tokens": [ + -1.0539793968200684 + ], + "logprob_reward": -0.42897936701774597, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42897936701774597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11672572791576385, + 0.057294487953186035, + 0.04024559259414673, + 0.031200751662254333, + 0.033022455871105194, + 0.06579595059156418, + 0.06295373290777206, + 0.056558866053819656, + 0.3702581822872162, + 0.16594424843788147 + ], + "credit_map": [ + 0.012449616566300392, + 0.02093883790075779, + 0.011759519577026367, + 0.01313500665128231, + 0.017833929508924484, + 0.016808554530143738, + 0.015483652241528034, + 0.00831660721451044, + 0.003533397102728486, + 0.008815230801701546, + 0.006075815297663212, + 0.007323805242776871, + 0.01525108888745308, + 0.006539121735841036, + 0.006020371336489916, + 0.003735658247023821, + 0.0031879241578280926, + 0.006566070485860109, + 0.006071062292903662, + 0.005283300764858723, + 0.007626265287399292, + 0.004276040475815535, + 0.003802817314863205, + 0.0034321104176342487, + 0.0022519861813634634, + 0.004877880681306124, + 0.0031629714649170637, + 0.00445740157738328, + 0.00691301841288805, + 0.0035526948049664497, + 0.004427049774676561, + 0.0015577469021081924, + 0.003023770870640874, + 0.004565006121993065, + 0.005447348579764366, + 0.004052706528455019, + 0.006798501592129469, + 0.0027823629789054394, + 0.0038479617796838284, + 0.002504796488210559, + 0.006614564452320337, + 0.008397315628826618, + 0.01311526633799076, + 0.00812019407749176, + 0.010567287914454937, + 0.006994635798037052, + 0.006408160086721182, + 0.0055785272270441055, + 0.004601642023772001, + 0.00865279883146286, + 0.013108484447002411, + 0.007376131135970354, + 0.011149141006171703, + 0.006845090072602034, + 0.0062928954139351845, + 0.004927552305161953, + 0.0031373987440019846, + 0.00895834993571043, + 0.005173929501324892, + 0.005285452585667372, + 0.013315238058567047, + 0.007274125702679157, + 0.009833093732595444, + 0.0035812784917652607, + 0.02462868019938469, + 0.05216871201992035, + 0.06981504708528519, + 0.055258478969335556, + 0.063327856361866, + 0.035682328045368195, + 0.04046010598540306, + 0.0289169792085886, + 0.009133826941251755, + 0.017869580537080765, + 0.016958944499492645, + 0.06365573406219482, + 0.017256302759051323, + 0.009364431723952293, + 0.02283615618944168, + 0.008869265206158161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.42897936701774597 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205747276544571 + ], + "opposite_logprob_tokens": [ + -2.8120574951171875 + ], + "logprob_reward": -0.06205747276544571, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205747276544571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13895238935947418, + 0.19889003038406372, + 0.07293590158224106, + 0.044558085501194, + 0.05302786827087402, + 0.045664042234420776, + 0.046782802790403366, + 0.06241421774029732, + 0.1426563858985901, + 0.19411823153495789 + ], + "credit_map": [ + 0.01759139820933342, + 0.02111002244055271, + 0.016471780836582184, + 0.021802252158522606, + 0.018626391887664795, + 0.01897975243628025, + 0.011197772808372974, + 0.013173015788197517, + 0.013950024731457233, + 0.029546206817030907, + 0.01813933625817299, + 0.024501625448465347, + 0.04568398743867874, + 0.026037000119686127, + 0.02802692912518978, + 0.013004925101995468, + 0.005600806325674057, + 0.011412416584789753, + 0.008842562325298786, + 0.010118217207491398, + 0.015025656670331955, + 0.007897577248513699, + 0.008074875921010971, + 0.005963790230453014, + 0.003520622383803129, + 0.006723780184984207, + 0.005981503985822201, + 0.0057687764056026936, + 0.011045165359973907, + 0.003736654296517372, + 0.005350908730179071, + 0.0024306776467710733, + 0.005695154424756765, + 0.009244206361472607, + 0.004678973462432623, + 0.007004844024777412, + 0.010382191278040409, + 0.006015424150973558, + 0.006436309311538935, + 0.0035707629285752773, + 0.004676178563386202, + 0.006217089481651783, + 0.007408598437905312, + 0.006221082992851734, + 0.00812524650245905, + 0.0046831523068249226, + 0.004739402327686548, + 0.003593290923163295, + 0.0029890332370996475, + 0.008082297630608082, + 0.008570187725126743, + 0.005427574273198843, + 0.008849791251122952, + 0.0044308979995548725, + 0.004642954096198082, + 0.003790064947679639, + 0.006125228479504585, + 0.010712490417063236, + 0.0075832842849195, + 0.006191813852638006, + 0.01343495398759842, + 0.007004030514508486, + 0.006533082574605942, + 0.004829337354749441, + 0.009978635236620903, + 0.02074851281940937, + 0.027788832783699036, + 0.018596410751342773, + 0.024100588634610176, + 0.013730784878134727, + 0.01556185632944107, + 0.01215075608342886, + 0.01205546222627163, + 0.026112554594874382, + 0.028150074183940887, + 0.04045071825385094, + 0.03200581297278404, + 0.018325654789805412, + 0.024028703570365906, + 0.012989250011742115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.06205747276544571 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro shot, captured from a slight overhead angle, maintains a still frame, focusing squarely on the intricate texture of a pristine white cotton pad, set against a deep, non-reflective, dark background. A bright, precise, clinical overhead light illuminates the scene, bringing each soft, interwoven fiber of the pad into sharp focus, casting subtle shadows and dramatically enhancing its three-dimensional quality. A drop of crystal-clear, slightly viscous disinfectant begins to trickle cautiously onto the pad's surface, capturing the microscopic process of the liquid soaking into the fibers in astonishing detail: the clear liquid's front advances steadily and deliberately, instantly saturating the dry, fluffy white fibers wherever it passes, turning them slightly darker and translucent. As the liquid continues to spread silently, the entire picture seamlessly transitions to a final static picture through a smooth and flowing dissolve shot - at this point, the entire cotton piece has been completely and evenly wetted, the fibers are tightly attached together, and they shimmer with a wet luster under the light, creating an extremely clean, focused and tranquil atmosphere with a scientific beauty.", + "seed": 33082, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_11/index_7420_seed_33082_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11395914554595947, + "vlm_logprob": -0.11395914554595947, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.11395914554595947, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.042198989540338516 + ], + "opposite_logprob_tokens": [ + -3.292198896408081 + ], + "logprob_reward": -0.042198989540338516, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.042198989540338516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2073553204536438, + 0.11358613520860672, + 0.1559799313545227, + 0.14796526730060577, + 0.06423100084066391, + 0.06135886535048485, + 0.0946262925863266, + 0.055102311074733734, + 0.060796383768320084, + 0.03899848088622093 + ], + "credit_map": [ + 0.014589742757380009, + 0.03188629448413849, + 0.02295553870499134, + 0.033402714878320694, + 0.04487226903438568, + 0.016300220042467117, + 0.026895489543676376, + 0.016453027725219727, + 0.008331495337188244, + 0.014969742856919765, + 0.011906975880265236, + 0.021022172644734383, + 0.023379141464829445, + 0.008164302445948124, + 0.017725281417369843, + 0.008087015710771084, + 0.013975114561617374, + 0.018128283321857452, + 0.01775898039340973, + 0.024945709854364395, + 0.020502546802163124, + 0.02600579708814621, + 0.024532215669751167, + 0.010131271556019783, + 0.014767344109714031, + 0.01760345697402954, + 0.016713125631213188, + 0.018087605014443398, + 0.022724293172359467, + 0.02261584997177124, + 0.018676504492759705, + 0.016777079552412033, + 0.004590671043843031, + 0.009531553834676743, + 0.008206572383642197, + 0.010650519281625748, + 0.009919489733874798, + 0.003716695588082075, + 0.011764531023800373, + 0.005850963294506073, + 0.0035966935101896524, + 0.009851246140897274, + 0.003569125197827816, + 0.011292475275695324, + 0.012920431792736053, + 0.003340173978358507, + 0.013102677650749683, + 0.0036860364489257336, + 0.008116756565868855, + 0.014297201298177242, + 0.011595371179282665, + 0.015306093730032444, + 0.013300263322889805, + 0.0067048510536551476, + 0.014906482771039009, + 0.010399265214800835, + 0.0049836584366858006, + 0.00837623793631792, + 0.007465645205229521, + 0.0075989775359630585, + 0.008416402153670788, + 0.00511838449165225, + 0.007025107741355896, + 0.006117894314229488, + 0.0068945190869271755, + 0.009322667494416237, + 0.008316703140735626, + 0.008979749865829945, + 0.00658507039770484, + 0.005283143371343613, + 0.00935717485845089, + 0.00605735182762146, + 0.0035242720041424036, + 0.006864383816719055, + 0.0033893140498548746, + 0.006074798759073019, + 0.007877921685576439, + 0.0035094842314720154, + 0.004438068717718124, + 0.0033202343620359898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.042198989540338516 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902271300554276 + ], + "opposite_logprob_tokens": [ + -2.5790226459503174 + ], + "logprob_reward": -0.07902271300554276, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902271300554276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15422067046165466, + 0.19103386998176575, + 0.22410164773464203, + 0.06486190110445023, + 0.10296759009361267, + 0.06577850133180618, + 0.0876329094171524, + 0.05338388308882713, + 0.03015640191733837, + 0.025862667709589005 + ], + "credit_map": [ + 0.016422642394900322, + 0.022354403510689735, + 0.018959391862154007, + 0.028512977063655853, + 0.025377236306667328, + 0.011569141410291195, + 0.021969517692923546, + 0.009055348113179207, + 0.013858720660209656, + 0.03105969727039337, + 0.01791565492749214, + 0.029804019257426262, + 0.043241918087005615, + 0.01412754412740469, + 0.028560973703861237, + 0.01246535312384367, + 0.019541259855031967, + 0.02634984254837036, + 0.026290223002433777, + 0.03605133295059204, + 0.02860582061111927, + 0.038780421018600464, + 0.031908150762319565, + 0.016574595123529434, + 0.010472225956618786, + 0.008034964092075825, + 0.010715730488300323, + 0.00823771208524704, + 0.008394826203584671, + 0.005198019556701183, + 0.009122390300035477, + 0.004686032421886921, + 0.009669613093137741, + 0.01726844720542431, + 0.011371089145541191, + 0.01624048687517643, + 0.01799304597079754, + 0.007509967777878046, + 0.016011040657758713, + 0.006903898436576128, + 0.005501513835042715, + 0.008760168217122555, + 0.006637756712734699, + 0.009918516501784325, + 0.011981124989688396, + 0.006467273458838463, + 0.010201042518019676, + 0.0063111064955592155, + 0.007403105963021517, + 0.013118692673742771, + 0.007324790116399527, + 0.015739988535642624, + 0.019237138330936432, + 0.007157464977353811, + 0.012222218327224255, + 0.005429507233202457, + 0.0041959634982049465, + 0.0076669128611683846, + 0.005393055733293295, + 0.009172261692583561, + 0.007785695139318705, + 0.005736044142395258, + 0.007861403748393059, + 0.005572544410824776, + 0.0035724726039916277, + 0.004439729731529951, + 0.003996885381639004, + 0.0047991047613322735, + 0.004738873336464167, + 0.001832342124544084, + 0.004713112488389015, + 0.00206388090737164, + 0.002535151084885001, + 0.003926422912627459, + 0.003115741303190589, + 0.004099941346794367, + 0.0030771316960453987, + 0.0017303446074947715, + 0.005448976531624794, + 0.0019289578776806593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.07902271300554276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991027235984802 + ], + "opposite_logprob_tokens": [ + -1.8049103021621704 + ], + "logprob_reward": -0.17991027235984802, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991027235984802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1918250322341919, + 0.2030499279499054, + 0.10079623758792877, + 0.050422076135873795, + 0.10884526371955872, + 0.1133481040596962, + 0.11331571638584137, + 0.042630232870578766, + 0.04213739186525345, + 0.033629994839429855 + ], + "credit_map": [ + 0.01260723453015089, + 0.029501108452677727, + 0.02307315729558468, + 0.02548603154718876, + 0.04729579761624336, + 0.013595451600849628, + 0.024950683116912842, + 0.015315563417971134, + 0.009636863134801388, + 0.03235888108611107, + 0.022448161616921425, + 0.031170442700386047, + 0.04318308085203171, + 0.012674374505877495, + 0.035855937749147415, + 0.015722185373306274, + 0.006944629363715649, + 0.015458895824849606, + 0.014097643084824085, + 0.01667363941669464, + 0.014387107454240322, + 0.009711491875350475, + 0.015574580989778042, + 0.007948240265250206, + 0.004630325362086296, + 0.008218729868531227, + 0.005763740744441748, + 0.008353161625564098, + 0.00797494500875473, + 0.004088946618139744, + 0.007171496283262968, + 0.004220730159431696, + 0.012654798105359077, + 0.012797870673239231, + 0.017565876245498657, + 0.013875313103199005, + 0.014616000466048717, + 0.011211647652089596, + 0.016901953145861626, + 0.009221803396940231, + 0.00985762756317854, + 0.016554860398173332, + 0.014048650860786438, + 0.017154233530163765, + 0.01893928088247776, + 0.00943514984101057, + 0.01737103797495365, + 0.009987265802919865, + 0.0141771100461483, + 0.01541810017079115, + 0.015630826354026794, + 0.01658342033624649, + 0.015363862738013268, + 0.010180668905377388, + 0.016350077465176582, + 0.009611645713448524, + 0.004058902617543936, + 0.0065989503636956215, + 0.005253964569419622, + 0.0057913195341825485, + 0.006535446736961603, + 0.0040785158053040504, + 0.005644295830279589, + 0.004668832290917635, + 0.004932889249175787, + 0.005727788899093866, + 0.005320193711668253, + 0.005520279053598642, + 0.005554306786507368, + 0.0045822616666555405, + 0.0059110005386173725, + 0.004588668700307608, + 0.0021636742167174816, + 0.005722089670598507, + 0.0034207492135465145, + 0.007673338055610657, + 0.0045156884007155895, + 0.0021009398624300957, + 0.005680499132722616, + 0.002353016985580325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.17991027235984802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559978067874908 + ], + "opposite_logprob_tokens": [ + -1.600599765777588 + ], + "logprob_reward": -0.22559978067874908, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559978067874908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20543068647384644, + 0.140753373503685, + 0.15195302665233612, + 0.0999903529882431, + 0.05066974461078644, + 0.12679457664489746, + 0.11570864170789719, + 0.04917793720960617, + 0.03028731979429722, + 0.029234375804662704 + ], + "credit_map": [ + 0.011347220279276371, + 0.031853705644607544, + 0.024623237550258636, + 0.030328169465065002, + 0.048155345022678375, + 0.016187869012355804, + 0.026410678401589394, + 0.01652446575462818, + 0.00802999921143055, + 0.023957358673214912, + 0.0163209680467844, + 0.01937863975763321, + 0.034864023327827454, + 0.011578104458749294, + 0.018917778506875038, + 0.007706489413976669, + 0.009857233613729477, + 0.02179187722504139, + 0.01873897947371006, + 0.026488367468118668, + 0.02257147990167141, + 0.019446389749646187, + 0.022187786176800728, + 0.010870907455682755, + 0.014923561364412308, + 0.012701369822025299, + 0.016462435945868492, + 0.012332849204540253, + 0.01253573503345251, + 0.006445244420319796, + 0.017636455595493317, + 0.006952703930437565, + 0.004751367028802633, + 0.007574012037366629, + 0.0071311453357338905, + 0.006199763156473637, + 0.007331393659114838, + 0.0039767129346728325, + 0.008692990988492966, + 0.00501236179843545, + 0.010906406678259373, + 0.01801862195134163, + 0.015672963112592697, + 0.01814878173172474, + 0.02532607689499855, + 0.009585157036781311, + 0.01895054057240486, + 0.010186036117374897, + 0.00854217354208231, + 0.018794121220707893, + 0.011109894141554832, + 0.014876279048621655, + 0.03283480182290077, + 0.0077456762082874775, + 0.014506944455206394, + 0.007298746146261692, + 0.0035424057859927416, + 0.007887838408350945, + 0.004830322228372097, + 0.005145004019141197, + 0.011126162484288216, + 0.005467680282890797, + 0.006054136902093887, + 0.00512438639998436, + 0.0020254412665963173, + 0.004909054841846228, + 0.0026551070623099804, + 0.005007161758840084, + 0.005920174531638622, + 0.0017370942514389753, + 0.006095347460359335, + 0.0019379362929612398, + 0.0023719947785139084, + 0.00463863043114543, + 0.0026557245291769505, + 0.005702975671738386, + 0.005974760744720697, + 0.001447824644856155, + 0.004613883327692747, + 0.001828580629080534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.22559978067874908 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on the texture of a pristine white cotton pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background deep, non-reflective, and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a drop of crystal-clear, slightly viscous disinfectant trickle onto the pad and soak into the fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video transition from the soaking process to a final static picture via a smooth and flowing dissolve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"In the final static picture, are the cotton fibers completely wetted, tightly attached, and shimmering with a wet luster?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043063972145318985 + ], + "opposite_logprob_tokens": [ + -3.1680638790130615 + ], + "logprob_reward": -0.043063972145318985, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043063972145318985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2078079730272293, + 0.17504623532295227, + 0.10023431479930878, + 0.09018605947494507, + 0.09745833277702332, + 0.06341148167848587, + 0.10103782266378403, + 0.05977002531290054, + 0.0502314493060112, + 0.05481638386845589 + ], + "credit_map": [ + 0.011248909868299961, + 0.031311873346567154, + 0.022587250918149948, + 0.0336131788790226, + 0.03894130885601044, + 0.018569713458418846, + 0.033809974789619446, + 0.017725754529237747, + 0.011251406744122505, + 0.023069551214575768, + 0.02453368715941906, + 0.024504724889993668, + 0.03310682624578476, + 0.011763514019548893, + 0.029087692499160767, + 0.01772882603108883, + 0.009299504570662975, + 0.014211971312761307, + 0.013804936781525612, + 0.013507034629583359, + 0.01484372466802597, + 0.013184993527829647, + 0.014115692116320133, + 0.007266451604664326, + 0.008568350225687027, + 0.014233467169106007, + 0.009212153032422066, + 0.014159697107970715, + 0.01719396747648716, + 0.006291422992944717, + 0.012858467176556587, + 0.007668516132980585, + 0.010370400734245777, + 0.014625982381403446, + 0.013291163370013237, + 0.01481324527412653, + 0.015477529726922512, + 0.005576136521995068, + 0.016874108463525772, + 0.006429754197597504, + 0.004948154091835022, + 0.009417325258255005, + 0.00713286641985178, + 0.010297256521880627, + 0.010611945763230324, + 0.003749306546524167, + 0.013043414801359177, + 0.004211205989122391, + 0.0055590420961380005, + 0.015819404274225235, + 0.007728542201220989, + 0.027741676196455956, + 0.013139674440026283, + 0.004327052738517523, + 0.022458048537373543, + 0.00426437146961689, + 0.007145068608224392, + 0.00973847508430481, + 0.01056758314371109, + 0.008357061073184013, + 0.006221809424459934, + 0.004551077261567116, + 0.007943461649119854, + 0.005245479289442301, + 0.003856791416183114, + 0.007591390982270241, + 0.004473164211958647, + 0.007465135771781206, + 0.0100356824696064, + 0.003923787269741297, + 0.008381623774766922, + 0.004503870848566294, + 0.0057373554445803165, + 0.009592798538506031, + 0.006545406766235828, + 0.008754692040383816, + 0.008419261313974857, + 0.0034812800586223602, + 0.007935353554785252, + 0.004350232891738415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.043063972145318985 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_12/index_tensor([9321, 9321])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_12/index_tensor([9321, 9321])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8ee562d49cbe0a186f402de05b0c3aa78e5b6b10 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_12/index_tensor([9321, 9321])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, ultra-high-definition macro close-up captures a Cape Pygmy Chameleon (Bradypodion pumilum) clinging motionless to a grotesque biomechanical structure—a nightmarish amalgamation of cold, ribbed metal pipes and pulsating, translucent organic tubing slick with a thin, gooey sheen. In a biomechanical horror style similar to H.R. Giger, the creature's body is a study in absolute stillness, its elastic tail tightly wrapped around a black metal tube, its peculiar feet gripping the alien surface. As it fades into itself, a slow, mesmerizing transformation ripples through its skin; its natural dermal pattern methodically dissolves, replaced by the textures and colors of its inhabitants. The skin morphs, mimicking the polished, bronze sheen of metal and the textured, fleshy tones of the organic tubes. The changes flow like a liquid algorithm until the chameleon becomes a seamless, living extension of the eerie architecture. As its body remains completely frozen, its conical, turret-like eyes move with unsettling independence: one slowly rotates, gazing directly into the camera with a cold, intelligent gleam, while the other swivels to scan the shadowy depths of the environment. The scene is illuminated by a single, harsh, cold-toned light from above, casting deep, dramatic shadows and creating sharp, specular highlights on metallic surfaces, defining an eerie, otherworldly atmosphere. The color palette is a desaturated mix of obsidian black, cold gray, and sickly ochre, creating a mood of claustrophobia and deep unease, as if the biological and the mechanical have merged into a nightmarish new form.", + "seed": 1985, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_12/index_9321_seed_1985_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06437506787478924, + "vlm_logprob": -0.06437506787478924, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.06437506787478924, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037450917065143585 + ], + "opposite_logprob_tokens": [ + -3.5374510288238525 + ], + "logprob_reward": -0.037450917065143585, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.037450917065143585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39208245277404785, + 0.07551664113998413, + 0.1154172420501709, + 0.09366324543952942, + 0.04069197550415993, + 0.0597728006541729, + 0.037993982434272766, + 0.05403149500489235, + 0.051334500312805176, + 0.07949569076299667 + ], + "credit_map": [ + 0.05277656391263008, + 0.05043676123023033, + 0.08170760422945023, + 0.0671953409910202, + 0.03864673897624016, + 0.046651165932416916, + 0.03409971296787262, + 0.02056855335831642, + 0.0044806706719100475, + 0.008170479908585548, + 0.011557268910109997, + 0.013536259531974792, + 0.012045014649629593, + 0.005523951258510351, + 0.012258746661245823, + 0.00794424582272768, + 0.008345494978129864, + 0.009095486253499985, + 0.02875841222703457, + 0.03213842958211899, + 0.009681385941803455, + 0.007300357334315777, + 0.012706751935184002, + 0.007390941027551889, + 0.006539579480886459, + 0.013299720361828804, + 0.015242772176861763, + 0.017867570742964745, + 0.009443351067602634, + 0.006895761005580425, + 0.01648877188563347, + 0.007885715924203396, + 0.002901955973356962, + 0.006379162427037954, + 0.004839921835809946, + 0.006674434524029493, + 0.006460065022110939, + 0.004667933564633131, + 0.0059243966825306416, + 0.0028441071044653654, + 0.006755604408681393, + 0.006576717831194401, + 0.006317385472357273, + 0.008243688382208347, + 0.008901375345885754, + 0.007640236523002386, + 0.00721007352694869, + 0.008127720095217228, + 0.004335757344961166, + 0.005123547278344631, + 0.003825004445388913, + 0.006637864746153355, + 0.00435325363650918, + 0.005649905651807785, + 0.004524506628513336, + 0.003544145030900836, + 0.005332779139280319, + 0.00801969226449728, + 0.006034350488334894, + 0.006237816531211138, + 0.006323650944977999, + 0.004494841676205397, + 0.007131192833185196, + 0.010457166470587254, + 0.005303715821355581, + 0.007879301905632019, + 0.005267600063234568, + 0.005845417734235525, + 0.006732860114425421, + 0.0048630088567733765, + 0.008191605098545551, + 0.0072509897872805595, + 0.009006496518850327, + 0.00939421821385622, + 0.010185115039348602, + 0.00852399691939354, + 0.01076489221304655, + 0.006441369187086821, + 0.01380321104079485, + 0.011376394890248775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.037450917065143585 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02333943359553814 + ], + "opposite_logprob_tokens": [ + -3.7733395099639893 + ], + "logprob_reward": -0.02333943359553814, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02333943359553814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28925779461860657, + 0.12065554410219193, + 0.1225234866142273, + 0.13880112767219543, + 0.02893466129899025, + 0.042677294462919235, + 0.026163635775446892, + 0.03731613978743553, + 0.0354454331099987, + 0.15822486579418182 + ], + "credit_map": [ + 0.036008480936288834, + 0.026338886469602585, + 0.0568402074277401, + 0.04302055016160011, + 0.029711231589317322, + 0.040283285081386566, + 0.035083457827568054, + 0.021971680223941803, + 0.008396069519221783, + 0.018131807446479797, + 0.018800294026732445, + 0.023657256737351418, + 0.013658581301569939, + 0.006754680071026087, + 0.02097134292125702, + 0.010285514406859875, + 0.011525520123541355, + 0.00931960716843605, + 0.02675243280827999, + 0.028116745874285698, + 0.011627774685621262, + 0.0100429467856884, + 0.01906539499759674, + 0.006073069758713245, + 0.012453106231987476, + 0.01614110730588436, + 0.02462209016084671, + 0.02662128396332264, + 0.012344324961304665, + 0.010211150161921978, + 0.024117279797792435, + 0.012290798127651215, + 0.0023305939976125956, + 0.003911288920789957, + 0.003034672699868679, + 0.005061520263552666, + 0.004165485501289368, + 0.004225784447044134, + 0.003862581681460142, + 0.0023427323903888464, + 0.003399978158995509, + 0.004386599641293287, + 0.003878579940646887, + 0.005898232571780682, + 0.008462836965918541, + 0.0068850452080369, + 0.005588827654719353, + 0.004177192226052284, + 0.0027520416770130396, + 0.0035162456333637238, + 0.002789219841361046, + 0.0035928376019001007, + 0.0033316167537122965, + 0.004939132835716009, + 0.0033226460218429565, + 0.0019198948284611106, + 0.0033782431855797768, + 0.004142289981245995, + 0.004270787816494703, + 0.004484361968934536, + 0.006090708076953888, + 0.0050499942153692245, + 0.005252668168395758, + 0.004647086374461651, + 0.0030792327597737312, + 0.004370353650301695, + 0.00471251318231225, + 0.004771022591739893, + 0.004063236061483622, + 0.004028186202049255, + 0.00750553235411644, + 0.002915353048592806, + 0.014915958978235722, + 0.011700070463120937, + 0.0160535778850317, + 0.017684848979115486, + 0.023067623376846313, + 0.01526616234332323, + 0.025910058990120888, + 0.03362656757235527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.02333943359553814 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708143889904022 + ], + "opposite_logprob_tokens": [ + -2.1270813941955566 + ], + "logprob_reward": -0.12708143889904022, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12708143889904022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23334001004695892, + 0.1394374966621399, + 0.14523117244243622, + 0.0591447576880455, + 0.09890724718570709, + 0.07876099646091461, + 0.04231668636202812, + 0.06235438585281372, + 0.08822321146726608, + 0.05228402838110924 + ], + "credit_map": [ + 0.02473507821559906, + 0.026685815304517746, + 0.031415678560733795, + 0.041578009724617004, + 0.040573347359895706, + 0.017360946163535118, + 0.03523372858762741, + 0.01575740985572338, + 0.012495766393840313, + 0.021294452250003815, + 0.018799293786287308, + 0.01892862468957901, + 0.023187773302197456, + 0.008010520599782467, + 0.024925749748945236, + 0.011795317754149437, + 0.014557546004652977, + 0.010998058132827282, + 0.028240766376256943, + 0.02768818289041519, + 0.012711155228316784, + 0.011452670209109783, + 0.030341261997818947, + 0.009241534397006035, + 0.004721473436802626, + 0.006932269781827927, + 0.00891253910958767, + 0.01163422130048275, + 0.005463243927806616, + 0.005873690824955702, + 0.009436541236937046, + 0.0061707766726613045, + 0.005926721729338169, + 0.010537923313677311, + 0.01139350701123476, + 0.023593200370669365, + 0.011301746591925621, + 0.011478433385491371, + 0.014311610721051693, + 0.010364104993641376, + 0.005555891431868076, + 0.007771771866828203, + 0.009508124552667141, + 0.01229797676205635, + 0.016576940193772316, + 0.008264069445431232, + 0.010318980552256107, + 0.008467244915664196, + 0.006208465900272131, + 0.005748899653553963, + 0.004970704670995474, + 0.0073557160794734955, + 0.004521028138697147, + 0.004574515391141176, + 0.0055212657898664474, + 0.0034160898067057133, + 0.005547099746763706, + 0.008033369667828083, + 0.008067060261964798, + 0.007055885624140501, + 0.007661847863346338, + 0.005511293653398752, + 0.00861585047096014, + 0.011861978098750114, + 0.007731219753623009, + 0.011367790400981903, + 0.0112520856782794, + 0.010642034932971, + 0.009915181435644627, + 0.009589784778654575, + 0.014098775573074818, + 0.013626339845359325, + 0.005012895911931992, + 0.005988329648971558, + 0.00920272246003151, + 0.005500814877450466, + 0.007836265489459038, + 0.0046377358958125114, + 0.008568109944462776, + 0.005537154618650675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.12708143889904022 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370778262615204 + ], + "opposite_logprob_tokens": [ + -3.408707857131958 + ], + "logprob_reward": -0.03370778262615204, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370778262615204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4316733479499817, + 0.0850478857755661, + 0.05889878794550896, + 0.0918722003698349, + 0.04361026734113693, + 0.0668698102235794, + 0.052324119955301285, + 0.017560521140694618, + 0.11997473984956741, + 0.03216828778386116 + ], + "credit_map": [ + 0.04356308653950691, + 0.03981278836727142, + 0.09888453036546707, + 0.08778224885463715, + 0.03807120397686958, + 0.05280878394842148, + 0.04158638417720795, + 0.029164304956793785, + 0.01149752363562584, + 0.010172532871365547, + 0.015013635158538818, + 0.012322600930929184, + 0.01241145096719265, + 0.004775106441229582, + 0.011188055388629436, + 0.0076669794507324696, + 0.0064134555868804455, + 0.005725836381316185, + 0.009781040251255035, + 0.011247387155890465, + 0.005448686890304089, + 0.008590912446379662, + 0.007737245410680771, + 0.003954221494495869, + 0.006223595235496759, + 0.01356733962893486, + 0.012607943266630173, + 0.017650537192821503, + 0.008884740993380547, + 0.007169046904891729, + 0.017412742599844933, + 0.008356250822544098, + 0.0028513947036117315, + 0.005614355206489563, + 0.004812541883438826, + 0.009320372715592384, + 0.007207036484032869, + 0.0033860853873193264, + 0.006098669953644276, + 0.004319813568145037, + 0.005623164586722851, + 0.010166590102016926, + 0.007660428062081337, + 0.01149427518248558, + 0.011032259091734886, + 0.007321162149310112, + 0.009152032434940338, + 0.004419897682964802, + 0.007273753639310598, + 0.005808026529848576, + 0.0064486730843782425, + 0.007280760444700718, + 0.005806597415357828, + 0.00792553648352623, + 0.007172152865678072, + 0.00460862135514617, + 0.0018063216703012586, + 0.002610986353829503, + 0.0017234765691682696, + 0.0036934753879904747, + 0.00225734687410295, + 0.0014186190674081445, + 0.002292749471962452, + 0.0017575464444234967, + 0.012203511781990528, + 0.01720842719078064, + 0.013641441240906715, + 0.015846138820052147, + 0.012750937603414059, + 0.014972800388932228, + 0.020599788054823875, + 0.012751691974699497, + 0.002742738462984562, + 0.0035038148052990437, + 0.0026199694257229567, + 0.003536609699949622, + 0.004977700300514698, + 0.003004136960953474, + 0.0066648609936237335, + 0.005118455737829208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.03370778262615204 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1002957671880722 + ], + "opposite_logprob_tokens": [ + -2.3502957820892334 + ], + "logprob_reward": -0.1002957671880722, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002957671880722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1701343059539795, + 0.12085635215044022, + 0.2570551931858063, + 0.13715890049934387, + 0.04975273460149765, + 0.05137080326676369, + 0.046052321791648865, + 0.04367329180240631, + 0.041830144822597504, + 0.08211594074964523 + ], + "credit_map": [ + 0.01325837429612875, + 0.012508065439760685, + 0.0336139053106308, + 0.03481461480259895, + 0.022539256140589714, + 0.023661307990550995, + 0.020115798339247704, + 0.009622977115213871, + 0.01423578429967165, + 0.011997414752840996, + 0.021371914073824883, + 0.027068527415394783, + 0.013763736933469772, + 0.0070727807469666, + 0.019634608179330826, + 0.005711585283279419, + 0.03539767861366272, + 0.02718210592865944, + 0.0375494584441185, + 0.04468967020511627, + 0.025537002831697464, + 0.02706753835082054, + 0.03836360201239586, + 0.02126813493669033, + 0.011133997701108456, + 0.011652585119009018, + 0.0257937740534544, + 0.03163216635584831, + 0.01202508807182312, + 0.010891943238675594, + 0.02558731473982334, + 0.008442041464149952, + 0.005815549753606319, + 0.004951147828251123, + 0.0059102741070091724, + 0.008525104261934757, + 0.006998896598815918, + 0.006488963030278683, + 0.007041309494525194, + 0.004021492786705494, + 0.004003280773758888, + 0.006941392086446285, + 0.007039657793939114, + 0.008498698472976685, + 0.007638575974851847, + 0.0075602224096655846, + 0.005742732435464859, + 0.0039462437853217125, + 0.005576813127845526, + 0.007491078227758408, + 0.0047436486929655075, + 0.008294648490846157, + 0.005952464882284403, + 0.004837808199226856, + 0.005541673395782709, + 0.003614186542108655, + 0.0033162443432956934, + 0.003545390209183097, + 0.00465007359161973, + 0.004734253045171499, + 0.0058662258088588715, + 0.006242125295102596, + 0.006599274929612875, + 0.008719702251255512, + 0.004102729260921478, + 0.004981104284524918, + 0.004477565176784992, + 0.005469716154038906, + 0.0049077775329351425, + 0.00398135744035244, + 0.00693264976143837, + 0.006977242883294821, + 0.00690303323790431, + 0.009157431311905384, + 0.013840082101523876, + 0.010245799086987972, + 0.010420983657240868, + 0.009609607979655266, + 0.01322571188211441, + 0.008713285438716412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.1002957671880722 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, ultra-high-definition macro close-up captures a Cape Pygmy Chameleon (Bradypodion pumilum) clinging motionless to a grotesque biomechanical structure—a nightmarish amalgamation of cold, ribbed metal pipes and pulsating, translucent organic tubing slick with a thin, gooey sheen. In a biomechanical horror style similar to H.R. Giger, the creature's body is a study in absolute stillness, its elastic tail tightly wrapped around a black metal tube, its peculiar feet gripping the alien surface. As it fades into itself, a slow, mesmerizing transformation ripples through its skin; its natural dermal pattern methodically dissolves, replaced by the textures and colors of its inhabitants. The skin morphs, mimicking the polished, bronze sheen of metal and the textured, fleshy tones of the organic tubes. The changes flow like a liquid algorithm until the chameleon becomes a seamless, living extension of the eerie architecture. As its body remains completely frozen, its conical, turret-like eyes move with unsettling independence: one slowly rotates, gazing directly into the camera with a cold, intelligent gleam, while the other swivels to scan the shadowy depths of the environment. The scene is illuminated by a single, harsh, cold-toned light from above, casting deep, dramatic shadows and creating sharp, specular highlights on metallic surfaces, defining an eerie, otherworldly atmosphere. The color palette is a desaturated mix of obsidian black, cold gray, and sickly ochre, creating a mood of claustrophobia and deep unease, as if the biological and the mechanical have merged into a nightmarish new form.", + "seed": 1986, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_12/index_9321_seed_1986_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08940926603972912, + "vlm_logprob": -0.08940926603972912, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.08940926603972912, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06667880713939667 + ], + "opposite_logprob_tokens": [ + -2.816678762435913 + ], + "logprob_reward": -0.06667880713939667, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06667880713939667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23783868551254272, + 0.062341686338186264, + 0.08019524067640305, + 0.07071088254451752, + 0.08577054738998413, + 0.07198455929756165, + 0.0649871900677681, + 0.06146721541881561, + 0.18251848220825195, + 0.08218555897474289 + ], + "credit_map": [ + 0.037519942969083786, + 0.026689676567912102, + 0.04486088454723358, + 0.0234900563955307, + 0.015198868699371815, + 0.026756221428513527, + 0.02812834270298481, + 0.03519466891884804, + 0.0095641715452075, + 0.007365565747022629, + 0.01082689967006445, + 0.006496200803667307, + 0.005083457101136446, + 0.009124498814344406, + 0.006046372931450605, + 0.007834517396986485, + 0.01098573673516512, + 0.005997064057737589, + 0.014010421931743622, + 0.005607623606920242, + 0.005803081206977367, + 0.012701306492090225, + 0.007531097158789635, + 0.01755889505147934, + 0.007749043405056, + 0.005200339946895838, + 0.010332201607525349, + 0.008689009584486485, + 0.004440172575414181, + 0.008599772118031979, + 0.008575539104640484, + 0.01712479628622532, + 0.014180167578160763, + 0.008588723838329315, + 0.01608135923743248, + 0.02188372239470482, + 0.005075740162283182, + 0.007335153874009848, + 0.005557672586292028, + 0.007067997474223375, + 0.006841691210865974, + 0.011399731040000916, + 0.009588680230081081, + 0.012898962944746017, + 0.006248862016946077, + 0.0076036350801587105, + 0.008664335124194622, + 0.008738655596971512, + 0.0160231851041317, + 0.006287186872214079, + 0.010471679270267487, + 0.009001955389976501, + 0.004940259736031294, + 0.006000879220664501, + 0.005552997346967459, + 0.006709039211273193, + 0.009071937762200832, + 0.007619732990860939, + 0.011491701938211918, + 0.0075829001143574715, + 0.004071539733558893, + 0.007086677476763725, + 0.00896544847637415, + 0.005577269941568375, + 0.028973516076803207, + 0.01582162268459797, + 0.04334895312786102, + 0.019852109253406525, + 0.01770981214940548, + 0.016432875767350197, + 0.02482493221759796, + 0.015554645098745823, + 0.013494091108441353, + 0.007766085676848888, + 0.015702666714787483, + 0.00883542001247406, + 0.007942226715385914, + 0.008571825921535492, + 0.012588808313012123, + 0.007284428924322128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.06667880713939667 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011107398197054863 + ], + "opposite_logprob_tokens": [ + -4.511107444763184 + ], + "logprob_reward": -0.011107398197054863, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011107398197054863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13499027490615845, + 0.1063641831278801, + 0.12419125437736511, + 0.061795756220817566, + 0.18183578550815582, + 0.06950658559799194, + 0.08467963337898254, + 0.042800769209861755, + 0.12479886412620544, + 0.06903690099716187 + ], + "credit_map": [ + 0.022036179900169373, + 0.015428908169269562, + 0.0230503361672163, + 0.01290077157318592, + 0.008772996254265308, + 0.0198226235806942, + 0.012835785746574402, + 0.02014267072081566, + 0.020909445360302925, + 0.015365364030003548, + 0.01938902586698532, + 0.011174492537975311, + 0.006528998725116253, + 0.010496326722204685, + 0.014452984556555748, + 0.008047547191381454, + 0.020252151414752007, + 0.007118333596736193, + 0.019589299336075783, + 0.0076235258020460606, + 0.007377335336059332, + 0.01507134735584259, + 0.010814535431563854, + 0.03634471818804741, + 0.0057015009224414825, + 0.007560695521533489, + 0.012432307936251163, + 0.011980235576629639, + 0.0033078882843255997, + 0.005894965957850218, + 0.00645440723747015, + 0.008463756181299686, + 0.019794177263975143, + 0.013761780224740505, + 0.04617878049612045, + 0.05753626674413681, + 0.008881007321178913, + 0.011038662865757942, + 0.01117821503430605, + 0.013466892763972282, + 0.012457406148314476, + 0.00723103154450655, + 0.01381372008472681, + 0.008637775667011738, + 0.00525695038959384, + 0.007659063674509525, + 0.006356172263622284, + 0.008094464428722858, + 0.01550478395074606, + 0.010202523320913315, + 0.01631889119744301, + 0.009934625588357449, + 0.007120307069271803, + 0.007779521867632866, + 0.008386009372770786, + 0.009432971477508545, + 0.006327580660581589, + 0.004586636554449797, + 0.00672599533572793, + 0.004733779933303595, + 0.0029525284189730883, + 0.0056373621337115765, + 0.006114773917943239, + 0.005722112953662872, + 0.024418095126748085, + 0.01022595725953579, + 0.0320674404501915, + 0.013205485418438911, + 0.008736359886825085, + 0.012320767156779766, + 0.013051696121692657, + 0.010773064568638802, + 0.014511764980852604, + 0.004695590119808912, + 0.014978732913732529, + 0.006620253901928663, + 0.004705070052295923, + 0.006813371554017067, + 0.00830843299627304, + 0.008403684012591839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.011107398197054863 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906334638595581 + ], + "opposite_logprob_tokens": [ + -2.4640634059906006 + ], + "logprob_reward": -0.08906334638595581, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906334638595581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12993444502353668, + 0.05137009918689728, + 0.04477153345942497, + 0.090103879570961, + 0.20234869420528412, + 0.039274610579013824, + 0.09901707619428635, + 0.09869435429573059, + 0.1940261423587799, + 0.05045915022492409 + ], + "credit_map": [ + 0.014101282693445683, + 0.013724304735660553, + 0.021829836070537567, + 0.012638858519494534, + 0.009848983027040958, + 0.016328325495123863, + 0.01781284064054489, + 0.023650022223591805, + 0.007255423814058304, + 0.004869230091571808, + 0.007049701176583767, + 0.004391747061163187, + 0.002970128320157528, + 0.005616400856524706, + 0.005519663915038109, + 0.01369780395179987, + 0.004269703291356564, + 0.0042412374168634415, + 0.007071252446621656, + 0.006665580440312624, + 0.00258782971650362, + 0.005772809032350779, + 0.004329453222453594, + 0.009833668358623981, + 0.016857722774147987, + 0.007872512564063072, + 0.021280838176608086, + 0.016159506514668465, + 0.005037764552980661, + 0.008118185214698315, + 0.006518928334116936, + 0.00825842097401619, + 0.021315190941095352, + 0.014382297173142433, + 0.052265338599681854, + 0.06083006411790848, + 0.010155191645026207, + 0.016219770535826683, + 0.01153398398309946, + 0.01564684882760048, + 0.006830296479165554, + 0.0042296419851481915, + 0.005967979319393635, + 0.006617373321205378, + 0.0026605413295328617, + 0.004431142471730709, + 0.003499489976093173, + 0.005038147792220116, + 0.022855153307318687, + 0.008027841337025166, + 0.021930888295173645, + 0.012870932929217815, + 0.005848538130521774, + 0.009155402891337872, + 0.009643587283790112, + 0.008684732019901276, + 0.018145056441426277, + 0.011000948026776314, + 0.016305705532431602, + 0.015509233810007572, + 0.006830764934420586, + 0.011823629960417747, + 0.011696839705109596, + 0.007382172625511885, + 0.03700433298945427, + 0.015663355588912964, + 0.04338638484477997, + 0.029718849807977676, + 0.010429251939058304, + 0.02093781717121601, + 0.020440200343728065, + 0.016445957124233246, + 0.007456903345882893, + 0.004060675855726004, + 0.009804228320717812, + 0.005079932510852814, + 0.004979007411748171, + 0.00708897365257144, + 0.00756973447278142, + 0.004419694654643536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.08906334638595581 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032714158296585 + ], + "opposite_logprob_tokens": [ + -2.350327253341675 + ], + "logprob_reward": -0.10032714158296585, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032714158296585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11033759266138077, + 0.045110203325748444, + 0.0766979306936264, + 0.10657669603824615, + 0.11210918426513672, + 0.08225858211517334, + 0.121820367872715, + 0.08534295111894608, + 0.1450788676738739, + 0.11466765403747559 + ], + "credit_map": [ + 0.010239090770483017, + 0.008277663961052895, + 0.019517997279763222, + 0.00885638315230608, + 0.005161725450307131, + 0.014803558588027954, + 0.010390877723693848, + 0.03309030085802078, + 0.007270597852766514, + 0.00626868661493063, + 0.007353764493018389, + 0.005191728938370943, + 0.00298001104965806, + 0.007368477061390877, + 0.004515646491199732, + 0.004161290358752012, + 0.007627931423485279, + 0.006134746130555868, + 0.01135324314236641, + 0.006428782362490892, + 0.004685790278017521, + 0.010199780575931072, + 0.010358002968132496, + 0.019909659400582314, + 0.018232932314276695, + 0.008229374885559082, + 0.025125892832875252, + 0.015522751025855541, + 0.005691573489457369, + 0.011274746619164944, + 0.007622207049280405, + 0.014877207577228546, + 0.011718807741999626, + 0.00805229227989912, + 0.028919536620378494, + 0.030050361528992653, + 0.008018613792955875, + 0.010522689670324326, + 0.006045902613550425, + 0.008780973963439465, + 0.01695307530462742, + 0.008680016733705997, + 0.015505029819905758, + 0.009768853895366192, + 0.006613961886614561, + 0.008546006865799427, + 0.007717264350503683, + 0.008474376052618027, + 0.023501580581068993, + 0.011672439984977245, + 0.021554481238126755, + 0.017996400594711304, + 0.008103306405246258, + 0.013858145102858543, + 0.013380312360823154, + 0.011753701604902744, + 0.015401833690702915, + 0.009350704960525036, + 0.020361734554171562, + 0.008658145554363728, + 0.007715451996773481, + 0.006656656973063946, + 0.009626499377191067, + 0.007571921683847904, + 0.02016814984381199, + 0.013064581900835037, + 0.021981792524456978, + 0.025078345090150833, + 0.013174430467188358, + 0.01258505042642355, + 0.023608313873410225, + 0.01541819330304861, + 0.02234223112463951, + 0.009665120393037796, + 0.029305757954716682, + 0.010801289230585098, + 0.006586201023310423, + 0.0107802115380764, + 0.016318822279572487, + 0.00886803213506937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.10032714158296585 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798696368932724 + ], + "opposite_logprob_tokens": [ + -1.8048696517944336 + ], + "logprob_reward": -0.1798696368932724, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798696368932724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11258119344711304, + 0.16124878823757172, + 0.06199987232685089, + 0.04738694429397583, + 0.06954950839281082, + 0.059151072055101395, + 0.14243613183498383, + 0.1090598851442337, + 0.09562340378761292, + 0.14096325635910034 + ], + "credit_map": [ + 0.016836583614349365, + 0.010368582792580128, + 0.018297454342246056, + 0.01148280594497919, + 0.009750694036483765, + 0.01270480826497078, + 0.008975561708211899, + 0.024164685979485512, + 0.030198529362678528, + 0.019651785492897034, + 0.022002773359417915, + 0.01309789065271616, + 0.007583698723465204, + 0.014001743867993355, + 0.019536565989255905, + 0.03517579287290573, + 0.004918064922094345, + 0.004283824935555458, + 0.008560244925320148, + 0.004840378183871508, + 0.004020446911454201, + 0.00870565790683031, + 0.009949062019586563, + 0.016722185537219048, + 0.00486860703676939, + 0.004267402924597263, + 0.006479978561401367, + 0.006776861380785704, + 0.00403713108971715, + 0.006754917558282614, + 0.006829963997006416, + 0.007372078951448202, + 0.004999402444809675, + 0.004990790039300919, + 0.018211420625448227, + 0.02118155173957348, + 0.0038035300094634295, + 0.005550525151193142, + 0.00550567964091897, + 0.005306597799062729, + 0.010170352645218372, + 0.0058435858227312565, + 0.010015170089900494, + 0.007155971601605415, + 0.005013425834476948, + 0.0067438166588544846, + 0.006421186029911041, + 0.0077875591814517975, + 0.027815572917461395, + 0.016044624149799347, + 0.023919833824038506, + 0.017378604039549828, + 0.01300020981580019, + 0.012249483726918697, + 0.014035375788807869, + 0.017992420122027397, + 0.020236000418663025, + 0.010198915377259254, + 0.01949760690331459, + 0.016785550862550735, + 0.007501915562897921, + 0.01075531356036663, + 0.01330319419503212, + 0.010781372897326946, + 0.015208404511213303, + 0.00955070462077856, + 0.019756710156798363, + 0.014210769906640053, + 0.00746989855542779, + 0.011938628740608692, + 0.009461300447583199, + 0.008026978932321072, + 0.023530587553977966, + 0.01000598818063736, + 0.027816100046038628, + 0.01429732982069254, + 0.00926994439214468, + 0.019004058092832565, + 0.023716330528259277, + 0.01332289818674326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.1798696368932724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, ultra-high-definition macro close-up captures a Cape Pygmy Chameleon (Bradypodion pumilum) clinging motionless to a grotesque biomechanical structure—a nightmarish amalgamation of cold, ribbed metal pipes and pulsating, translucent organic tubing slick with a thin, gooey sheen. In a biomechanical horror style similar to H.R. Giger, the creature's body is a study in absolute stillness, its elastic tail tightly wrapped around a black metal tube, its peculiar feet gripping the alien surface. As it fades into itself, a slow, mesmerizing transformation ripples through its skin; its natural dermal pattern methodically dissolves, replaced by the textures and colors of its inhabitants. The skin morphs, mimicking the polished, bronze sheen of metal and the textured, fleshy tones of the organic tubes. The changes flow like a liquid algorithm until the chameleon becomes a seamless, living extension of the eerie architecture. As its body remains completely frozen, its conical, turret-like eyes move with unsettling independence: one slowly rotates, gazing directly into the camera with a cold, intelligent gleam, while the other swivels to scan the shadowy depths of the environment. The scene is illuminated by a single, harsh, cold-toned light from above, casting deep, dramatic shadows and creating sharp, specular highlights on metallic surfaces, defining an eerie, otherworldly atmosphere. The color palette is a desaturated mix of obsidian black, cold gray, and sickly ochre, creating a mood of claustrophobia and deep unease, as if the biological and the mechanical have merged into a nightmarish new form.", + "seed": 1987, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_12/index_9321_seed_1987_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05485024619847536, + "vlm_logprob": -0.05485024619847536, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.05485024619847536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02927263081073761 + ], + "opposite_logprob_tokens": [ + -3.7792725563049316 + ], + "logprob_reward": -0.02927263081073761, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02927263081073761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10314948856830597, + 0.17092593014240265, + 0.08970987051725388, + 0.0447537936270237, + 0.08983179926872253, + 0.05519421398639679, + 0.16503949463367462, + 0.07631251960992813, + 0.15920767188072205, + 0.04587520286440849 + ], + "credit_map": [ + 0.0114136366173625, + 0.009587663225829601, + 0.01358115766197443, + 0.016976242884993553, + 0.010891025885939598, + 0.013375205919146538, + 0.011901699006557465, + 0.015422862954437733, + 0.015472901985049248, + 0.01882287487387657, + 0.014529530890285969, + 0.03193988651037216, + 0.012394391000270844, + 0.02617870271205902, + 0.02097109705209732, + 0.030616549775004387, + 0.009036287665367126, + 0.009715717285871506, + 0.00715667987242341, + 0.01246684417128563, + 0.00738633843138814, + 0.018240878358483315, + 0.01118744257837534, + 0.014519684948027134, + 0.003777341451495886, + 0.003213069401681423, + 0.0034454234410077333, + 0.006159791257232428, + 0.0033689087722450495, + 0.010446063242852688, + 0.007043503224849701, + 0.007299690507352352, + 0.006615673191845417, + 0.008282315917313099, + 0.006984154228121042, + 0.010799318552017212, + 0.007477276027202606, + 0.020555807277560234, + 0.014502812176942825, + 0.014614447019994259, + 0.005870012566447258, + 0.005067257676273584, + 0.007215654477477074, + 0.005511457100510597, + 0.005922260228544474, + 0.006058738566935062, + 0.009963328950107098, + 0.009585503488779068, + 0.011987482197582722, + 0.014905028976500034, + 0.013748629949986935, + 0.03918735310435295, + 0.015927987173199654, + 0.020198296755552292, + 0.026165012270212173, + 0.022919708862900734, + 0.012233889661729336, + 0.008334639482200146, + 0.013850286602973938, + 0.009756864048540592, + 0.00782246608287096, + 0.00857815332710743, + 0.006596272345632315, + 0.009139951318502426, + 0.016715029254555702, + 0.01449587568640709, + 0.020606454461812973, + 0.019698625430464745, + 0.015301124192774296, + 0.026657717302441597, + 0.016712237149477005, + 0.02902061492204666, + 0.00341229559853673, + 0.007978767156600952, + 0.003204797161743045, + 0.00468906294554472, + 0.006788256578147411, + 0.005359442438930273, + 0.00740411039441824, + 0.007038473151624203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.02927263081073761 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012570788152515888 + ], + "opposite_logprob_tokens": [ + -4.387570858001709 + ], + "logprob_reward": -0.012570788152515888, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012570788152515888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14748640358448029, + 0.09404411911964417, + 0.09158002585172653, + 0.058095477521419525, + 0.09502578526735306, + 0.05893880873918533, + 0.14496414363384247, + 0.12353482097387314, + 0.12452874332666397, + 0.06180160120129585 + ], + "credit_map": [ + 0.016441166400909424, + 0.017457464709877968, + 0.020104937255382538, + 0.02122218906879425, + 0.017839299514889717, + 0.017341936007142067, + 0.01452853623777628, + 0.02255089394748211, + 0.008748743683099747, + 0.009075301699340343, + 0.01025417447090149, + 0.012463843449950218, + 0.006841912865638733, + 0.016422603279352188, + 0.013302220962941647, + 0.016935326159000397, + 0.007266457192599773, + 0.010234192945063114, + 0.006511777173727751, + 0.014180716127157211, + 0.0069866422563791275, + 0.015527356415987015, + 0.015804611146450043, + 0.015068277716636658, + 0.005099269095808268, + 0.00515471026301384, + 0.005294740665704012, + 0.0075601800344884396, + 0.004506740719079971, + 0.012363622896373272, + 0.009881207719445229, + 0.008235014043748379, + 0.006853801663964987, + 0.009555011987686157, + 0.007666305638849735, + 0.011854950338602066, + 0.008496302179992199, + 0.02213813178241253, + 0.012351962737739086, + 0.01610933057963848, + 0.005439711268991232, + 0.004596409387886524, + 0.0049396394751966, + 0.005859618075191975, + 0.004833899904042482, + 0.015079373493790627, + 0.008857465349137783, + 0.00933270063251257, + 0.010247300378978252, + 0.013397622853517532, + 0.010806506499648094, + 0.03523916006088257, + 0.016736552119255066, + 0.017392408102750778, + 0.021203236654400826, + 0.019941365346312523, + 0.018748972564935684, + 0.01177042257040739, + 0.02039380557835102, + 0.014529932290315628, + 0.01058117114007473, + 0.015729175880551338, + 0.016116611659526825, + 0.015664739534258842, + 0.015993718057870865, + 0.012770936824381351, + 0.015441175550222397, + 0.015029787085950375, + 0.010264445096254349, + 0.01633562333881855, + 0.018894998356699944, + 0.019798070192337036, + 0.005421978421509266, + 0.006558213382959366, + 0.005008826497942209, + 0.006141935009509325, + 0.006295809522271156, + 0.011440940201282501, + 0.009567697532474995, + 0.011366206221282482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.012570788152515888 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062073830515146255 + ], + "opposite_logprob_tokens": [ + -2.8120739459991455 + ], + "logprob_reward": -0.062073830515146255, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062073830515146255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16944250464439392, + 0.12513600289821625, + 0.16790983080863953, + 0.1025676354765892, + 0.12277816981077194, + 0.08266604691743851, + 0.09321820735931396, + 0.02396974340081215, + 0.09853716194629669, + 0.013774785213172436 + ], + "credit_map": [ + 0.016603926196694374, + 0.0219589676707983, + 0.02380414307117462, + 0.025663351640105247, + 0.017024647444486618, + 0.018114320933818817, + 0.01769326627254486, + 0.02857985347509384, + 0.009767888113856316, + 0.014861884526908398, + 0.009482314810156822, + 0.02494528330862522, + 0.008825690485537052, + 0.017672032117843628, + 0.016188332810997963, + 0.023392563685774803, + 0.018612194806337357, + 0.01580232009291649, + 0.014686299487948418, + 0.024447033181786537, + 0.014637929387390614, + 0.032516252249479294, + 0.02173062227666378, + 0.025477156043052673, + 0.007885335013270378, + 0.010858220048248768, + 0.008318436332046986, + 0.015277852304279804, + 0.011291620321571827, + 0.01346932165324688, + 0.019450165331363678, + 0.016016673296689987, + 0.01321196649223566, + 0.013653644360601902, + 0.014072070829570293, + 0.019129585474729538, + 0.010928665287792683, + 0.014076503925025463, + 0.02042289450764656, + 0.01728283241391182, + 0.008154914714396, + 0.009606311097741127, + 0.008940009400248528, + 0.009022776037454605, + 0.008523927070200443, + 0.010842794552445412, + 0.013892335817217827, + 0.013682965189218521, + 0.010048718191683292, + 0.00851032231003046, + 0.01259576715528965, + 0.019802259281277657, + 0.0071382359601557255, + 0.009692946448922157, + 0.013854432851076126, + 0.011575515381991863, + 0.002471396466717124, + 0.003203523112460971, + 0.002960342215374112, + 0.0033685187809169292, + 0.002507236320525408, + 0.0027516603004187346, + 0.003279909724369645, + 0.003427155315876007, + 0.011284089647233486, + 0.008448163978755474, + 0.011752499267458916, + 0.01135975867509842, + 0.010897722095251083, + 0.01746341586112976, + 0.008738446980714798, + 0.01859305612742901, + 0.0011060605756938457, + 0.0017960394034162164, + 0.0009409874910488725, + 0.0012117434525862336, + 0.0013088345294818282, + 0.001911938888952136, + 0.002715391805395484, + 0.00278378720395267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.062073830515146255 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032261162996292 + ], + "opposite_logprob_tokens": [ + -2.350322723388672 + ], + "logprob_reward": -0.10032261162996292, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032261162996292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1748947650194168, + 0.10196732729673386, + 0.13367389142513275, + 0.09798014909029007, + 0.11492158472537994, + 0.061079952865839005, + 0.11415141075849533, + 0.08680552989244461, + 0.04396485909819603, + 0.0705605000257492 + ], + "credit_map": [ + 0.01513542328029871, + 0.01873827539384365, + 0.015335449017584324, + 0.028771869838237762, + 0.0216031763702631, + 0.02794310823082924, + 0.01972111500799656, + 0.027646347880363464, + 0.006615109276026487, + 0.010695184580981731, + 0.010198446922004223, + 0.013748634606599808, + 0.00948268361389637, + 0.01768450066447258, + 0.01652408577501774, + 0.01701868325471878, + 0.013089578598737717, + 0.01350055169314146, + 0.00978783704340458, + 0.018708638846874237, + 0.01107287872582674, + 0.030679205432534218, + 0.015739453956484795, + 0.02109575644135475, + 0.008439736440777779, + 0.01089469250291586, + 0.007732461206614971, + 0.015454555861651897, + 0.010915304534137249, + 0.012976465746760368, + 0.016652097925543785, + 0.01491483487188816, + 0.010203000158071518, + 0.015758920460939407, + 0.010912819765508175, + 0.01666947454214096, + 0.010244050063192844, + 0.013728484511375427, + 0.02140657976269722, + 0.015998249873518944, + 0.005940727889537811, + 0.0054305074736475945, + 0.006940542720258236, + 0.0066519975662231445, + 0.005038618575781584, + 0.011171567253768444, + 0.010795952752232552, + 0.009110040962696075, + 0.011124702170491219, + 0.00928385742008686, + 0.018303781747817993, + 0.022339334711432457, + 0.007222904823720455, + 0.01167281623929739, + 0.01935095340013504, + 0.014853058382868767, + 0.014570659026503563, + 0.005427451338618994, + 0.01621979847550392, + 0.011272105388343334, + 0.006986165884882212, + 0.010274706408381462, + 0.011606399901211262, + 0.01044823694974184, + 0.0048897177912294865, + 0.005206188187003136, + 0.006035241298377514, + 0.005675151012837887, + 0.003753010416403413, + 0.0053872475400567055, + 0.006786602083593607, + 0.0062317019328475, + 0.006832587532699108, + 0.011602738872170448, + 0.005104188807308674, + 0.005605603568255901, + 0.009258287027478218, + 0.009527621790766716, + 0.01058100163936615, + 0.012048467062413692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.10032261162996292 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001136988401413 + ], + "opposite_logprob_tokens": [ + -2.6950113773345947 + ], + "logprob_reward": -0.07001136988401413, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001136988401413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12045270949602127, + 0.18502412736415863, + 0.11349932104349136, + 0.042099516838788986, + 0.05876104533672333, + 0.03397930786013603, + 0.15849916636943817, + 0.11110115051269531, + 0.11096155643463135, + 0.06562204658985138 + ], + "credit_map": [ + 0.012105354107916355, + 0.01455786731094122, + 0.01710515283048153, + 0.01975454017519951, + 0.008246959187090397, + 0.014262367971241474, + 0.014747380279004574, + 0.01967308484017849, + 0.012356284074485302, + 0.018985455855727196, + 0.010956977494060993, + 0.022797539830207825, + 0.013246651738882065, + 0.035642899572849274, + 0.04139922559261322, + 0.029639095067977905, + 0.0071787117049098015, + 0.010353908874094486, + 0.009035377763211727, + 0.01518283225595951, + 0.008427578024566174, + 0.026307523250579834, + 0.01458373200148344, + 0.022429663687944412, + 0.0034495831932872534, + 0.005365869030356407, + 0.0031166241969913244, + 0.004819337744265795, + 0.004521734081208706, + 0.0060664936900138855, + 0.00846991315484047, + 0.006289962213486433, + 0.004623676184564829, + 0.005630295258015394, + 0.003501422004774213, + 0.006971888709813356, + 0.004737310577183962, + 0.014055580832064152, + 0.009688719175755978, + 0.00955214910209179, + 0.0027792719192802906, + 0.00391891086474061, + 0.0027603518683463335, + 0.0027644902002066374, + 0.0024846866726875305, + 0.006134835071861744, + 0.006994523573666811, + 0.006142234895378351, + 0.01208570133894682, + 0.015458270907402039, + 0.011197592131793499, + 0.035742223262786865, + 0.019832680001854897, + 0.021014660596847534, + 0.02076992578804493, + 0.02239811420440674, + 0.01824677735567093, + 0.007709234021604061, + 0.01943913847208023, + 0.014075974933803082, + 0.008934647776186466, + 0.015371903777122498, + 0.012559855356812477, + 0.014763613231480122, + 0.013825887814164162, + 0.008945576846599579, + 0.010406400077044964, + 0.009965878911316395, + 0.011714374646544456, + 0.022546371445059776, + 0.015081866644322872, + 0.01847519725561142, + 0.004812123253941536, + 0.003776522586122155, + 0.003594571491703391, + 0.005208771675825119, + 0.004474715329706669, + 0.009733238257467747, + 0.021084962412714958, + 0.012937144376337528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.07001136988401413 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, ultra-high-definition macro close-up captures a Cape Pygmy Chameleon (Bradypodion pumilum) clinging motionless to a grotesque biomechanical structure—a nightmarish amalgamation of cold, ribbed metal pipes and pulsating, translucent organic tubing slick with a thin, gooey sheen. In a biomechanical horror style similar to H.R. Giger, the creature's body is a study in absolute stillness, its elastic tail tightly wrapped around a black metal tube, its peculiar feet gripping the alien surface. As it fades into itself, a slow, mesmerizing transformation ripples through its skin; its natural dermal pattern methodically dissolves, replaced by the textures and colors of its inhabitants. The skin morphs, mimicking the polished, bronze sheen of metal and the textured, fleshy tones of the organic tubes. The changes flow like a liquid algorithm until the chameleon becomes a seamless, living extension of the eerie architecture. As its body remains completely frozen, its conical, turret-like eyes move with unsettling independence: one slowly rotates, gazing directly into the camera with a cold, intelligent gleam, while the other swivels to scan the shadowy depths of the environment. The scene is illuminated by a single, harsh, cold-toned light from above, casting deep, dramatic shadows and creating sharp, specular highlights on metallic surfaces, defining an eerie, otherworldly atmosphere. The color palette is a desaturated mix of obsidian black, cold gray, and sickly ochre, creating a mood of claustrophobia and deep unease, as if the biological and the mechanical have merged into a nightmarish new form.", + "seed": 1988, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_12/index_9321_seed_1988_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08390920646488667, + "vlm_logprob": -0.08390920646488667, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.08390920646488667, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04526485502719879 + ], + "opposite_logprob_tokens": [ + -3.170264959335327 + ], + "logprob_reward": -0.04526485502719879, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04526485502719879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2127181440591812, + 0.08817064762115479, + 0.05529375746846199, + 0.08799497783184052, + 0.05420800298452377, + 0.058171629905700684, + 0.1182074099779129, + 0.027987156063318253, + 0.07590838521718979, + 0.22133991122245789 + ], + "credit_map": [ + 0.018166599795222282, + 0.023310218006372452, + 0.028091706335544586, + 0.017221525311470032, + 0.023071452975273132, + 0.028184162452816963, + 0.034616731107234955, + 0.04005573317408562, + 0.0076819355599582195, + 0.011921987868845463, + 0.011960061267018318, + 0.009663312695920467, + 0.011381663382053375, + 0.006652828771620989, + 0.013648416846990585, + 0.015260428190231323, + 0.007327534258365631, + 0.007275968790054321, + 0.006272836588323116, + 0.005894562229514122, + 0.004741624929010868, + 0.003988147713243961, + 0.00988498330116272, + 0.00990809965878725, + 0.007614569738507271, + 0.007757595274597406, + 0.008549310266971588, + 0.00992846954613924, + 0.011265292763710022, + 0.008831936866044998, + 0.019764885306358337, + 0.01428291667252779, + 0.005179360508918762, + 0.007040448486804962, + 0.006873757112771273, + 0.006712969392538071, + 0.0062639228999614716, + 0.004893291275948286, + 0.010285128839313984, + 0.006959117483347654, + 0.0063578845001757145, + 0.01121731661260128, + 0.006270762532949448, + 0.004839739762246609, + 0.004403235390782356, + 0.006146575324237347, + 0.010146514512598515, + 0.008789599873125553, + 0.005666251294314861, + 0.005989606026560068, + 0.006739275995641947, + 0.0057290419936180115, + 0.006720203440636396, + 0.009136760607361794, + 0.04380242899060249, + 0.034423839300870895, + 0.0027106136549264193, + 0.004166326951235533, + 0.003426371840760112, + 0.0028046765364706516, + 0.002491917461156845, + 0.002169801155105233, + 0.005705518182367086, + 0.0045119295828044415, + 0.00695194024592638, + 0.00919799692928791, + 0.007725605741143227, + 0.015249763615429401, + 0.0049632214941084385, + 0.007567866705358028, + 0.013643928803503513, + 0.010608059354126453, + 0.01421611662954092, + 0.010208206251263618, + 0.01745477318763733, + 0.012397575192153454, + 0.009721136651933193, + 0.011939163319766521, + 0.0741143748164177, + 0.07128855586051941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.04526485502719879 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076529644429683685 + ], + "opposite_logprob_tokens": [ + -4.882652759552002 + ], + "logprob_reward": -0.0076529644429683685, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076529644429683685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10792534053325653, + 0.04651132598519325, + 0.24858951568603516, + 0.06974495947360992, + 0.05217234417796135, + 0.05512183532118797, + 0.15879976749420166, + 0.08014920353889465, + 0.07530681788921356, + 0.10567892342805862 + ], + "credit_map": [ + 0.011388338170945644, + 0.011397933587431908, + 0.013467260636389256, + 0.01197360921651125, + 0.010354687459766865, + 0.011882050894200802, + 0.0191604383289814, + 0.018301013857126236, + 0.006655489094555378, + 0.004415243864059448, + 0.005199771840125322, + 0.006249443627893925, + 0.00373335974290967, + 0.004230128601193428, + 0.009394088760018349, + 0.006633793469518423, + 0.02899133414030075, + 0.04554285109043121, + 0.023725152015686035, + 0.02363468147814274, + 0.035820093005895615, + 0.02175654098391533, + 0.03506725654006004, + 0.03405159339308739, + 0.005183419678360224, + 0.00596024002879858, + 0.003546578111127019, + 0.004002991132438183, + 0.004937455523759127, + 0.005559881217777729, + 0.01970032788813114, + 0.02085406333208084, + 0.0060783191584050655, + 0.0060634431429207325, + 0.006599956192076206, + 0.007920166477560997, + 0.004800604656338692, + 0.005694402381777763, + 0.009508559480309486, + 0.005506888031959534, + 0.007649888284504414, + 0.007311620283871889, + 0.006794001441448927, + 0.005643394775688648, + 0.003923109266906977, + 0.00738946720957756, + 0.009808342903852463, + 0.006602006498724222, + 0.01839960180222988, + 0.021352557465434074, + 0.024063877761363983, + 0.013795559294521809, + 0.026950186118483543, + 0.021845517680048943, + 0.01883568987250328, + 0.013556759804487228, + 0.007345433346927166, + 0.009387246333062649, + 0.011460898444056511, + 0.007685451302677393, + 0.008036843501031399, + 0.01353473961353302, + 0.013080820441246033, + 0.009617763571441174, + 0.0060172127559781075, + 0.009228900074958801, + 0.008092324249446392, + 0.019602851942181587, + 0.005780575796961784, + 0.007197236642241478, + 0.011902231723070145, + 0.007485474459826946, + 0.009042312391102314, + 0.011069511994719505, + 0.011795631609857082, + 0.016033418476581573, + 0.005426528863608837, + 0.007077649235725403, + 0.024519750848412514, + 0.020714109763503075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.0076529644429683685 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370925828814507 + ], + "opposite_logprob_tokens": [ + -3.6513710021972656 + ], + "logprob_reward": -0.026370925828814507, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370925828814507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08235170692205429, + 0.035523671656847, + 0.29273509979248047, + 0.03856813907623291, + 0.0380905345082283, + 0.05150068178772926, + 0.13027092814445496, + 0.0334954559803009, + 0.17622792720794678, + 0.12123586237430573 + ], + "credit_map": [ + 0.008626177906990051, + 0.011562218889594078, + 0.009222609922289848, + 0.006970301270484924, + 0.010754856280982494, + 0.00936741754412651, + 0.012941637076437473, + 0.01290648989379406, + 0.0028050867840647697, + 0.004302074667066336, + 0.00685270456597209, + 0.004538988694548607, + 0.004450203385204077, + 0.00263159885071218, + 0.005184922367334366, + 0.004758091177791357, + 0.03370898589491844, + 0.05078143998980522, + 0.028746267780661583, + 0.028785819187760353, + 0.043840307742357254, + 0.025314239785075188, + 0.04167696088552475, + 0.03988109156489372, + 0.0035094846971333027, + 0.004454728681594133, + 0.0033913329243659973, + 0.0032733427360653877, + 0.003115123836323619, + 0.004717057570815086, + 0.008684408850967884, + 0.007422657683491707, + 0.005304461810737848, + 0.005566322710365057, + 0.005067440215498209, + 0.004457681905478239, + 0.003873678855597973, + 0.004237381741404533, + 0.005978327244520187, + 0.003605240024626255, + 0.00696697598323226, + 0.007430989760905504, + 0.005474460776895285, + 0.004839564673602581, + 0.003522916231304407, + 0.004875654820352793, + 0.009669685736298561, + 0.00872043613344431, + 0.012987622991204262, + 0.014324790798127651, + 0.015624111518263817, + 0.011308277025818825, + 0.018553920090198517, + 0.014588627964258194, + 0.024325866252183914, + 0.0185577142983675, + 0.0032029389403760433, + 0.00472011836245656, + 0.004580266773700714, + 0.0039751469157636166, + 0.0034545203670859337, + 0.004452714696526527, + 0.005132993217557669, + 0.003976759500801563, + 0.015414450317621231, + 0.017534729093313217, + 0.01876455545425415, + 0.04504154622554779, + 0.011622617952525616, + 0.011070366948843002, + 0.031057659536600113, + 0.025722011923789978, + 0.010251273401081562, + 0.010329453274607658, + 0.01048942469060421, + 0.011608692817389965, + 0.00939172226935625, + 0.005499311722815037, + 0.0321357287466526, + 0.03153024986386299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.026370925828814507 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799083799123764 + ], + "opposite_logprob_tokens": [ + -1.8049083948135376 + ], + "logprob_reward": -0.1799083799123764, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799083799123764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11102715879678726, + 0.06341993063688278, + 0.06750869005918503, + 0.0763893574476242, + 0.07898727804422379, + 0.08490369468927383, + 0.20184531807899475, + 0.051617421209812164, + 0.07613632827997208, + 0.1881648302078247 + ], + "credit_map": [ + 0.0136372996494174, + 0.009761512279510498, + 0.012987377122044563, + 0.013389523141086102, + 0.010864732787013054, + 0.01171556580811739, + 0.022756479680538177, + 0.015914667397737503, + 0.006959976628422737, + 0.006882165092974901, + 0.008406378328800201, + 0.0068036518059670925, + 0.006265365984290838, + 0.00586652010679245, + 0.012354335747659206, + 0.009881540201604366, + 0.007350027561187744, + 0.011302034370601177, + 0.00802625622600317, + 0.0065835630521178246, + 0.009414276108145714, + 0.004366046283394098, + 0.010714776813983917, + 0.009751707315444946, + 0.006530835293233395, + 0.012252870947122574, + 0.007097681052982807, + 0.009789209812879562, + 0.013832464814186096, + 0.006154557690024376, + 0.012193780392408371, + 0.008537961170077324, + 0.012163249775767326, + 0.015144068747758865, + 0.009427781216800213, + 0.006967076100409031, + 0.008143149316310883, + 0.00754510099068284, + 0.012278323993086815, + 0.00731853349134326, + 0.011477300897240639, + 0.01244430709630251, + 0.00848359428346157, + 0.007064376957714558, + 0.0051027825102210045, + 0.00846280436962843, + 0.015816353261470795, + 0.016052179038524628, + 0.020731313154101372, + 0.028134845197200775, + 0.0268094502389431, + 0.01693827100098133, + 0.03509032726287842, + 0.026856038719415665, + 0.02854990027844906, + 0.018735164776444435, + 0.0050383140332996845, + 0.010249568149447441, + 0.005802012514322996, + 0.0059663536958396435, + 0.005514157935976982, + 0.006272441241890192, + 0.007331248838454485, + 0.00544332480058074, + 0.009878430515527725, + 0.009471888653934002, + 0.009341433644294739, + 0.008636830374598503, + 0.0050874268636107445, + 0.008268996141850948, + 0.014913903549313545, + 0.010537417605519295, + 0.009738554246723652, + 0.009878360666334629, + 0.015217983163893223, + 0.027889057993888855, + 0.010027353651821613, + 0.008607573807239532, + 0.055508434772491455, + 0.051297515630722046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.1799083799123764 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Cape Pygmy Chameleon (Bradypodion pumilum) clinging to a structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a grotesque biomechanical structure made of cold ribbed metal pipes and translucent organic tubing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the chameleon's skin undergo a transformation where its natural pattern dissolves and is replaced by the textures of the surrounding metal and organic tubes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"While the chameleon's body remains motionless, do its conical turret-like eyes move independently to scan the environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a single harsh cold-toned light from above, casting deep shadows and sharp specular highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034890711307526 + ], + "opposite_logprob_tokens": [ + -1.910348892211914 + ], + "logprob_reward": -0.16034890711307526, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034890711307526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09867607057094574, + 0.03850312530994415, + 0.031540174037218094, + 0.05459488928318024, + 0.06243826076388359, + 0.0571410097181797, + 0.2523101270198822, + 0.13694517314434052, + 0.07610291987657547, + 0.19174830615520477 + ], + "credit_map": [ + 0.011254901066422462, + 0.016871994361281395, + 0.014031131751835346, + 0.010465798899531364, + 0.009561924263834953, + 0.007455797865986824, + 0.015337292104959488, + 0.013697225600481033, + 0.005193781573325396, + 0.0039390441961586475, + 0.00475719291716814, + 0.004301035776734352, + 0.004611897747963667, + 0.0040556215681135654, + 0.005738866049796343, + 0.005905682221055031, + 0.0050456407479941845, + 0.0033756382763385773, + 0.003000988159328699, + 0.003606167621910572, + 0.004276481922715902, + 0.003483064938336611, + 0.004785069264471531, + 0.003967117518186569, + 0.005247180815786123, + 0.006364381406456232, + 0.008033241145312786, + 0.0072412146255373955, + 0.0052223242819309235, + 0.004603725858032703, + 0.01064886711537838, + 0.007233950775116682, + 0.007460559252649546, + 0.007955982349812984, + 0.0075441161170601845, + 0.007601779419928789, + 0.0065400623716413975, + 0.006282036192715168, + 0.01224018819630146, + 0.006813532672822475, + 0.00787405390292406, + 0.01039785798639059, + 0.0063104829750955105, + 0.005780796520411968, + 0.005418494809418917, + 0.007384516764432192, + 0.008496672846376896, + 0.005478130653500557, + 0.012972765602171421, + 0.012345781549811363, + 0.02452937699854374, + 0.018957816064357758, + 0.013428066857159138, + 0.02079777792096138, + 0.08579938858747482, + 0.06347911059856415, + 0.011004616506397724, + 0.01754298247396946, + 0.019147854298353195, + 0.013356879353523254, + 0.015108396299183369, + 0.024972589686512947, + 0.020141201093792915, + 0.01567063108086586, + 0.007265011314302683, + 0.008169607259333134, + 0.00824388861656189, + 0.015353837981820107, + 0.006445283070206642, + 0.004549378529191017, + 0.013294343836605549, + 0.01278156042098999, + 0.011096135713160038, + 0.011577594093978405, + 0.015290910378098488, + 0.02140776813030243, + 0.008247032761573792, + 0.01138269528746605, + 0.05951886996626854, + 0.053227297961711884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.16034890711307526 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_13/index_tensor([2349, 2349])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_13/index_tensor([2349, 2349])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b7eabaca3291e161d57b10b67f7399dba604d010 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_13/index_tensor([2349, 2349])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up captures a dramatic confrontation between two large bullfrogs, rendered in the minimalist and highly expressive style of traditional Chinese ink painting. The scene unfolds against a stark off-white background, accented by the subtle fibrous texture of rice paper, creating a sense of concentrated intensity. In a display of dominance, one bullfrog adopts a powerful, grounded pose, its form defined by bold, wet strokes of deep black ink, depicting its muscular rump. Its throat pouch, bulging in defiance, is depicted in delicate, translucent light gray ink. The other frog retreats defensively, its body twisted into a tense arc, captured in a series of rapid, energetic, dry brushstrokes that convey frantic movement and struggle. At the center of the conflict, their powerful hind legs kick up a violent cloud of ink spray, exploding across the canvas in expressive, calligraphic splashes and dots, freezing the chaotic action with artistic precision. The monochromatic palette, ranging from the deepest black to aqua grey, contrasts with the pure white of the paper, creating a high-contrast, visually appealing atmosphere of raw, elegant conflict.", + "seed": 30172, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_13/index_2349_seed_30172_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.011586230294778944, + "vlm_logprob": -0.011586230294778944, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.011586230294778944, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03893740102648735 + ], + "opposite_logprob_tokens": [ + -3.2889373302459717 + ], + "logprob_reward": -0.03893740102648735, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03893740102648735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.100577212870121, + 0.06799490749835968, + 0.07404132187366486, + 0.03118853084743023, + 0.06539666652679443, + 0.30299755930900574, + 0.1078355610370636, + 0.15933352708816528, + 0.04082249477505684, + 0.04981227591633797 + ], + "credit_map": [ + 0.017709750682115555, + 0.009845131076872349, + 0.018884330987930298, + 0.009881587699055672, + 0.006011964287608862, + 0.014439060352742672, + 0.01577341929078102, + 0.008031962439417839, + 0.00859097670763731, + 0.0131570128723979, + 0.005492606665939093, + 0.008084590546786785, + 0.008921603672206402, + 0.009324267506599426, + 0.009860783815383911, + 0.004563058726489544, + 0.014295628294348717, + 0.01249739620834589, + 0.0037671788595616817, + 0.010279931128025055, + 0.005289987660944462, + 0.011917183175683022, + 0.011345600709319115, + 0.0046484097838401794, + 0.005798347294330597, + 0.004505796357989311, + 0.0022446834482252598, + 0.004259730689227581, + 0.0027802663389593363, + 0.00445578945800662, + 0.004692006856203079, + 0.0024519087746739388, + 0.008341438136994839, + 0.010418702848255634, + 0.00501458253711462, + 0.011477898806333542, + 0.008677981793880463, + 0.008629643358290195, + 0.009152253158390522, + 0.0036841556429862976, + 0.023398879915475845, + 0.025782229378819466, + 0.03375818207859993, + 0.03396393731236458, + 0.030666178092360497, + 0.037433523684740067, + 0.09595862030982971, + 0.0220359917730093, + 0.008708471432328224, + 0.016733577474951744, + 0.01648525521159172, + 0.008909917436540127, + 0.017067361623048782, + 0.012427779845893383, + 0.014668147079646587, + 0.012835035100579262, + 0.016280295327305794, + 0.021414030343294144, + 0.02022351138293743, + 0.022780882194638252, + 0.018524158746004105, + 0.014181975275278091, + 0.030386921018362045, + 0.015541746281087399, + 0.004530931357294321, + 0.007634375710040331, + 0.003678075270727277, + 0.005218426696956158, + 0.005986825563013554, + 0.005029889754951, + 0.005293171852827072, + 0.003450794843956828, + 0.006556919310241938, + 0.008273008279502392, + 0.004918992519378662, + 0.005006512627005577, + 0.007143678143620491, + 0.007549771573394537, + 0.006478212308138609, + 0.0038851804565638304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.03893740102648735 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005959717091172934 + ], + "opposite_logprob_tokens": [ + -5.130959510803223 + ], + "logprob_reward": -0.005959717091172934, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005959717091172934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13735535740852356, + 0.08639494329690933, + 0.2403106391429901, + 0.0866064503788948, + 0.11462517082691193, + 0.10133379697799683, + 0.09397581964731216, + 0.07512009888887405, + 0.026795029640197754, + 0.03748273849487305 + ], + "credit_map": [ + 0.02672833576798439, + 0.011548449285328388, + 0.02295803464949131, + 0.015156405977904797, + 0.009729336015880108, + 0.023154577240347862, + 0.018172413110733032, + 0.009907805360853672, + 0.012468949891626835, + 0.010641755536198616, + 0.006821954157203436, + 0.009568171575665474, + 0.01053997315466404, + 0.012917068786919117, + 0.01680852845311165, + 0.006628543138504028, + 0.036568935960531235, + 0.03823740407824516, + 0.015024884603917599, + 0.03483452647924423, + 0.02765662595629692, + 0.032717783004045486, + 0.034176841378211975, + 0.021093642339110374, + 0.022042393684387207, + 0.01134750247001648, + 0.004089251160621643, + 0.007156025618314743, + 0.006170297507196665, + 0.013703025877475739, + 0.017251228913664818, + 0.004846721887588501, + 0.013905324973165989, + 0.016426777467131615, + 0.006582594010978937, + 0.01853463239967823, + 0.014857553876936436, + 0.015035534277558327, + 0.02303827553987503, + 0.006244478281587362, + 0.013658413663506508, + 0.00819438323378563, + 0.01072635967284441, + 0.01746418885886669, + 0.010979571379721165, + 0.014115964993834496, + 0.01994638890028, + 0.006248529069125652, + 0.0062043191865086555, + 0.00942502822726965, + 0.012245419435203075, + 0.010408512316644192, + 0.014238778501749039, + 0.010087218135595322, + 0.02331855520606041, + 0.008047993294894695, + 0.007910046726465225, + 0.011531858704984188, + 0.006009222008287907, + 0.009133933112025261, + 0.011057032272219658, + 0.006883170455694199, + 0.01305858138948679, + 0.009536253288388252, + 0.0038730385713279247, + 0.004906760528683662, + 0.002385633299127221, + 0.004082252271473408, + 0.0031104236841201782, + 0.0031394402030855417, + 0.0033557964488863945, + 0.0019416847499087453, + 0.003954860847443342, + 0.003727919887751341, + 0.003960594534873962, + 0.00734091317281127, + 0.0035980697721242905, + 0.005112737882882357, + 0.006101873237639666, + 0.003685771021991968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.005959717091172934 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00220104050822556 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.00220104050822556, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.00220104050822556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15016981959342957, + 0.08015305548906326, + 0.38433000445365906, + 0.05211569741368294, + 0.07096222043037415, + 0.05985834822058678, + 0.11006864905357361, + 0.025449207052588463, + 0.034127555787563324, + 0.032765500247478485 + ], + "credit_map": [ + 0.02883206307888031, + 0.01047984417527914, + 0.02319161407649517, + 0.0162932388484478, + 0.009297066368162632, + 0.02513296529650688, + 0.026221301406621933, + 0.010721704922616482, + 0.013911190442740917, + 0.011486854404211044, + 0.00799015536904335, + 0.010186081752181053, + 0.008373910561203957, + 0.012433589436113834, + 0.009967472404241562, + 0.005803795997053385, + 0.05073302984237671, + 0.05597902461886406, + 0.025150248780846596, + 0.06373806297779083, + 0.03589395806193352, + 0.05168391019105911, + 0.0692356526851654, + 0.03191608563065529, + 0.009878545999526978, + 0.005526121240109205, + 0.0041943383403122425, + 0.007388964295387268, + 0.004911092575639486, + 0.008861973881721497, + 0.008543410338461399, + 0.002811247482895851, + 0.011830448172986507, + 0.01083625853061676, + 0.005200249142944813, + 0.008170348592102528, + 0.008998806588351727, + 0.00958075001835823, + 0.012143153697252274, + 0.004202202893793583, + 0.005147612653672695, + 0.0058901868760585785, + 0.005728363059461117, + 0.01012455765157938, + 0.006094490643590689, + 0.008201910182833672, + 0.01417283434420824, + 0.004498383495956659, + 0.012909114360809326, + 0.010994961485266685, + 0.014842749573290348, + 0.010242129676043987, + 0.021273763850331306, + 0.011960817500948906, + 0.020365118980407715, + 0.007479984313249588, + 0.002744341967627406, + 0.003629297250881791, + 0.0028339081909507513, + 0.003857236821204424, + 0.004187459126114845, + 0.002785208635032177, + 0.002698851516470313, + 0.002712902147322893, + 0.0035790675319731236, + 0.005098339635878801, + 0.00308576924726367, + 0.0048596509732306, + 0.005231684539467096, + 0.0036443176213651896, + 0.006185524631291628, + 0.002443201607093215, + 0.005122065544128418, + 0.004213769920170307, + 0.004485149867832661, + 0.0034386871848255396, + 0.003211366478353739, + 0.0052261631935834885, + 0.004299725405871868, + 0.0027685705572366714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.00220104050822556 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095619544386864 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095619544386864, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095619544386864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08636493980884552, + 0.041804514825344086, + 0.19318927824497223, + 0.07510972768068314, + 0.17994186282157898, + 0.23947586119174957, + 0.06512618809938431, + 0.037637706845998764, + 0.05146760493516922, + 0.02988230064511299 + ], + "credit_map": [ + 0.013058698736131191, + 0.0069037615321576595, + 0.012834751047194004, + 0.010082539170980453, + 0.0060529024340212345, + 0.012328330427408218, + 0.01781855709850788, + 0.007285401690751314, + 0.006939318962395191, + 0.0051566436886787415, + 0.004276318941265345, + 0.0043028867803514, + 0.004627799149602652, + 0.006390917580574751, + 0.006761342752724886, + 0.0033492848742753267, + 0.04699442535638809, + 0.03066873736679554, + 0.008594295009970665, + 0.012887196615338326, + 0.013321691192686558, + 0.03538885340094566, + 0.03526737913489342, + 0.010066705755889416, + 0.019937358796596527, + 0.010264342650771141, + 0.004764105658978224, + 0.007273353170603514, + 0.0052552418783307076, + 0.012451483868062496, + 0.011281980201601982, + 0.003881857031956315, + 0.02064039744436741, + 0.026753246784210205, + 0.014374203979969025, + 0.022231006994843483, + 0.019197266548871994, + 0.021007535979151726, + 0.04361238330602646, + 0.012125828303396702, + 0.025503719225525856, + 0.017310855910182, + 0.028714487329125404, + 0.03525537997484207, + 0.029809942469000816, + 0.03044193796813488, + 0.05513066053390503, + 0.017308887094259262, + 0.005965467542409897, + 0.010948479175567627, + 0.008327974937856197, + 0.005831758491694927, + 0.012902054004371166, + 0.0067098308354616165, + 0.009341263212263584, + 0.005099363159388304, + 0.004660103935748339, + 0.005735732614994049, + 0.004873679950833321, + 0.0044376119039952755, + 0.004775614943355322, + 0.004004149232059717, + 0.004835053812712431, + 0.004315759986639023, + 0.006368094589561224, + 0.008944740518927574, + 0.00497070886194706, + 0.00733396178111434, + 0.0064270831644535065, + 0.005949144251644611, + 0.006596673280000687, + 0.004877197556197643, + 0.00424019992351532, + 0.004447974264621735, + 0.0037515985313802958, + 0.002829905366525054, + 0.004303096793591976, + 0.004223762545734644, + 0.0034625178668648005, + 0.0026232462842017412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.004095619544386864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737373303622007 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.006737373303622007, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737373303622007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09227611124515533, + 0.03616928309202194, + 0.09113036096096039, + 0.07020718604326248, + 0.34587958455085754, + 0.07835040241479874, + 0.1640990674495697, + 0.046440690755844116, + 0.04715461656451225, + 0.02829275280237198 + ], + "credit_map": [ + 0.015497714281082153, + 0.006471703294664621, + 0.01163483690470457, + 0.010235315188765526, + 0.005906731355935335, + 0.013300138525664806, + 0.02092014066874981, + 0.008309529162943363, + 0.00604903232306242, + 0.005065143574029207, + 0.0030632177367806435, + 0.003167385933920741, + 0.003915274515748024, + 0.005255109164863825, + 0.0067495135590434074, + 0.0029046041890978813, + 0.028409449383616447, + 0.009942807257175446, + 0.007010421250015497, + 0.005458977073431015, + 0.008242697454988956, + 0.021280962973833084, + 0.004905018489807844, + 0.005880020093172789, + 0.012671448290348053, + 0.011426961049437523, + 0.004059190861880779, + 0.007751445285975933, + 0.010571355000138283, + 0.01100784633308649, + 0.009030169807374477, + 0.0036887654569000006, + 0.046617940068244934, + 0.04239718243479729, + 0.021649178117513657, + 0.07222601026296616, + 0.04305557906627655, + 0.04686973989009857, + 0.04799385368824005, + 0.02507009729743004, + 0.009444544091820717, + 0.006088552065193653, + 0.008972742594778538, + 0.01739746332168579, + 0.006795869208872318, + 0.010869250632822514, + 0.013080921024084091, + 0.005701052024960518, + 0.01688775047659874, + 0.020946266129612923, + 0.018429679796099663, + 0.010649139992892742, + 0.03198985755443573, + 0.016938479617238045, + 0.034003496170043945, + 0.014254393056035042, + 0.006082257721573114, + 0.006184803321957588, + 0.006183370016515255, + 0.006362412124872208, + 0.008596406318247318, + 0.005983910523355007, + 0.004074589349329472, + 0.002972937189042568, + 0.004687061067670584, + 0.006068290211260319, + 0.004385767504572868, + 0.008147446438670158, + 0.006048852577805519, + 0.00506321107968688, + 0.009279862977564335, + 0.0034741247072815895, + 0.0036868329625576735, + 0.0042649125680327415, + 0.0030378743540495634, + 0.003984910901635885, + 0.003041550749912858, + 0.003563746577128768, + 0.0037287527229636908, + 0.002984172897413373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.006737373303622007 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures a dramatic confrontation between two large bullfrogs, rendered in the minimalist and highly expressive style of traditional Chinese ink painting. The scene unfolds against a stark off-white background, accented by the subtle fibrous texture of rice paper, creating a sense of concentrated intensity. In a display of dominance, one bullfrog adopts a powerful, grounded pose, its form defined by bold, wet strokes of deep black ink, depicting its muscular rump. Its throat pouch, bulging in defiance, is depicted in delicate, translucent light gray ink. The other frog retreats defensively, its body twisted into a tense arc, captured in a series of rapid, energetic, dry brushstrokes that convey frantic movement and struggle. At the center of the conflict, their powerful hind legs kick up a violent cloud of ink spray, exploding across the canvas in expressive, calligraphic splashes and dots, freezing the chaotic action with artistic precision. The monochromatic palette, ranging from the deepest black to aqua grey, contrasts with the pure white of the paper, creating a high-contrast, visually appealing atmosphere of raw, elegant conflict.", + "seed": 30173, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_13/index_2349_seed_30173_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.004606963833793998, + "vlm_logprob": -0.004606963833793998, + "vlm_yes_no_margin": 5.525, + "vlm_token_credit": -0.004606963833793998, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004434157628566027 + ], + "opposite_logprob_tokens": [ + -5.629434108734131 + ], + "logprob_reward": -0.004434157628566027, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.004434157628566027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1533796489238739, + 0.06260713934898376, + 0.11051518470048904, + 0.06809621304273605, + 0.20611616969108582, + 0.09952656924724579, + 0.12107779830694199, + 0.06113036721944809, + 0.03745405003428459, + 0.08009685575962067 + ], + "credit_map": [ + 0.021230120211839676, + 0.02147609554231167, + 0.015522649511694908, + 0.01828787662088871, + 0.022425813600420952, + 0.022060422226786613, + 0.023219719529151917, + 0.009156969375908375, + 0.010263606905937195, + 0.01202390342950821, + 0.006746414117515087, + 0.006237567402422428, + 0.011820572428405285, + 0.0067807347513735294, + 0.0049690427258610725, + 0.00376530853100121, + 0.013470384292304516, + 0.04286490008234978, + 0.006136530544608831, + 0.009929790161550045, + 0.00973019190132618, + 0.018095066770911217, + 0.005738358478993177, + 0.004549977835267782, + 0.010295729152858257, + 0.010855193249881268, + 0.00634472630918026, + 0.011510267853736877, + 0.006980630569159985, + 0.009708921425044537, + 0.007736369036138058, + 0.004664385225623846, + 0.038045551627874374, + 0.0597742423415184, + 0.008409527130424976, + 0.009663986973464489, + 0.02221955731511116, + 0.052449800074100494, + 0.007219476159662008, + 0.008334039710462093, + 0.008591127581894398, + 0.028386667370796204, + 0.006105612963438034, + 0.015255730599164963, + 0.01070250291377306, + 0.016541404649615288, + 0.007212810218334198, + 0.006730720400810242, + 0.019766900688409805, + 0.025671914219856262, + 0.009688559919595718, + 0.010903669521212578, + 0.010940509848296642, + 0.018727483227849007, + 0.01759590394794941, + 0.007782858796417713, + 0.008812272921204567, + 0.011632820591330528, + 0.0062907966785132885, + 0.009435482323169708, + 0.004727459512650967, + 0.009603280574083328, + 0.00519209448248148, + 0.005436166189610958, + 0.005816236604005098, + 0.008299250155687332, + 0.0028236217331141233, + 0.004698021803051233, + 0.0060602836310863495, + 0.004586195573210716, + 0.0034063232596963644, + 0.0017641204176470637, + 0.014986990951001644, + 0.02076166309416294, + 0.007153711281716824, + 0.0050893970765173435, + 0.0065262699499726295, + 0.015809979289770126, + 0.005857221316546202, + 0.003911630250513554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.004434157628566027 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953910294920206 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.005953910294920206, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953910294920206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14294348657131195, + 0.07661053538322449, + 0.09594325721263885, + 0.07767564058303833, + 0.07808198779821396, + 0.05593927949666977, + 0.06420069932937622, + 0.04630972445011139, + 0.04333699122071266, + 0.31895846128463745 + ], + "credit_map": [ + 0.022995345294475555, + 0.02818872034549713, + 0.012760261073708534, + 0.016073554754257202, + 0.014751505106687546, + 0.025841407477855682, + 0.013956360518932343, + 0.00837631057947874, + 0.010517830029129982, + 0.013187631033360958, + 0.007880410179495811, + 0.00686741853132844, + 0.012862038798630238, + 0.008649725466966629, + 0.011068502441048622, + 0.00557697331532836, + 0.013697348535060883, + 0.029341278597712517, + 0.008156987838447094, + 0.008388151414692402, + 0.008507201448082924, + 0.015621447935700417, + 0.007180434186011553, + 0.005050399340689182, + 0.012328032404184341, + 0.01510725636035204, + 0.007090662606060505, + 0.012997735291719437, + 0.007873163558542728, + 0.011776435188949108, + 0.00635752035304904, + 0.004144827835261822, + 0.01360449567437172, + 0.017565859481692314, + 0.004560409113764763, + 0.004793690983206034, + 0.009113381616771221, + 0.01843821443617344, + 0.006014219950884581, + 0.003991710487753153, + 0.006367992144078016, + 0.011964423581957817, + 0.004140819422900677, + 0.011758450418710709, + 0.003767389105632901, + 0.009834276512265205, + 0.0045073870569467545, + 0.0035985331051051617, + 0.010361192747950554, + 0.014843019656836987, + 0.003743074368685484, + 0.005164480302482843, + 0.006401720456779003, + 0.011730550788342953, + 0.007583205122500658, + 0.004373446572571993, + 0.006031581200659275, + 0.011479880660772324, + 0.00269203819334507, + 0.008174402639269829, + 0.0035842678043991327, + 0.007106347940862179, + 0.004632795695215464, + 0.0026084063574671745, + 0.006698866840451956, + 0.006529970094561577, + 0.0031937265302985907, + 0.004216190427541733, + 0.011777744628489017, + 0.0051964931190013885, + 0.0036786070559173822, + 0.0020453857723623514, + 0.05613436922430992, + 0.09271601587533951, + 0.020388366654515266, + 0.017368080094456673, + 0.0273049958050251, + 0.07151325792074203, + 0.019763780757784843, + 0.013769584707915783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.005953910294920206 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002199494047090411 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.002199494047090411, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002199494047090411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10344108939170837, + 0.11447610706090927, + 0.1417236477136612, + 0.049464765936136246, + 0.25244593620300293, + 0.056807614862918854, + 0.05460720881819725, + 0.042908165603876114, + 0.02630084566771984, + 0.15782465040683746 + ], + "credit_map": [ + 0.014876543544232845, + 0.026562893763184547, + 0.005103686824440956, + 0.009593834169209003, + 0.009568396955728531, + 0.018758950755000114, + 0.013267368078231812, + 0.005709415767341852, + 0.01624140329658985, + 0.019772890955209732, + 0.012247278355062008, + 0.011989333666861057, + 0.024054914712905884, + 0.012307732366025448, + 0.010227137245237827, + 0.007635415997356176, + 0.028256511315703392, + 0.03664673492312431, + 0.011166401207447052, + 0.013453511521220207, + 0.009803121909499168, + 0.02228311076760292, + 0.011301928199827671, + 0.008812326937913895, + 0.008294879458844662, + 0.006444595288485289, + 0.006424400955438614, + 0.007958229631185532, + 0.004028965253382921, + 0.006930557079613209, + 0.005296893417835236, + 0.004086245782673359, + 0.04600217565894127, + 0.07638156414031982, + 0.01290155854076147, + 0.012348641641438007, + 0.026315663009881973, + 0.06009487807750702, + 0.008727231994271278, + 0.009674225002527237, + 0.0047211782075464725, + 0.01473477017134428, + 0.003384533803910017, + 0.010585412383079529, + 0.005909543484449387, + 0.008646932430565357, + 0.005246540531516075, + 0.0035787043161690235, + 0.009362539276480675, + 0.01828344166278839, + 0.0030851990450173616, + 0.004024336580187082, + 0.0044512273743748665, + 0.007229303475469351, + 0.0054842927493155, + 0.0026868691202253103, + 0.005880808923393488, + 0.008673664182424545, + 0.0038961635436862707, + 0.00852997973561287, + 0.0027187755331397057, + 0.006562196649610996, + 0.004043801687657833, + 0.002602773020043969, + 0.0039976672269403934, + 0.00435807928442955, + 0.002538643078878522, + 0.0031808633357286453, + 0.004021913278847933, + 0.003399837762117386, + 0.0033381793182343245, + 0.0014656621497124434, + 0.02623778022825718, + 0.041293688118457794, + 0.01426328532397747, + 0.013056585565209389, + 0.014277727343142033, + 0.023667413741350174, + 0.017202021554112434, + 0.007826141081750393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.002199494047090411 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818899229168892 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.002818899229168892, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818899229168892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11177711188793182, + 0.04710054770112038, + 0.13773855566978455, + 0.03997797891497612, + 0.29780253767967224, + 0.03905903548002243, + 0.1334352344274521, + 0.030860522761940956, + 0.01793942227959633, + 0.14430902898311615 + ], + "credit_map": [ + 0.014515887945890427, + 0.01896027848124504, + 0.00969936978071928, + 0.013769765384495258, + 0.012039744295179844, + 0.017522184178233147, + 0.019644005224108696, + 0.005625882185995579, + 0.007243981119245291, + 0.007545842789113522, + 0.0052001578733325005, + 0.005030094180256128, + 0.008796349167823792, + 0.005626838654279709, + 0.0046392264775931835, + 0.003018062561750412, + 0.018131479620933533, + 0.050216611474752426, + 0.008652890101075172, + 0.011931737884879112, + 0.0117805115878582, + 0.022234031930565834, + 0.008924087509512901, + 0.005867204628884792, + 0.005541479680687189, + 0.0062998379580676556, + 0.004422674886882305, + 0.004624664783477783, + 0.006237579509615898, + 0.005238017532974482, + 0.005475880578160286, + 0.0021378472447395325, + 0.05528002977371216, + 0.09226908534765244, + 0.013933965004980564, + 0.01412609126418829, + 0.03394132852554321, + 0.06603983789682388, + 0.012372597120702267, + 0.009839638136327267, + 0.005501453764736652, + 0.008752834983170033, + 0.0031429214868694544, + 0.006047546397894621, + 0.0034448031801730394, + 0.006621398497372866, + 0.0031561367213726044, + 0.0023919446393847466, + 0.026516038924455643, + 0.04579603299498558, + 0.005972675047814846, + 0.008980153128504753, + 0.012927107512950897, + 0.019071267917752266, + 0.0092623895034194, + 0.0049095782451331615, + 0.004249319434165955, + 0.006816402543336153, + 0.0025720996782183647, + 0.0052124615758657455, + 0.0021711222361773252, + 0.004531266633421183, + 0.003341419156640768, + 0.0019664329010993242, + 0.0026987683959305286, + 0.0034903602208942175, + 0.0016513484297320247, + 0.0026304463390260935, + 0.002229495672509074, + 0.0020132572390139103, + 0.0019916773308068514, + 0.001234071096405387, + 0.024893715977668762, + 0.04017246514558792, + 0.0086026880890131, + 0.009420250542461872, + 0.01199386827647686, + 0.03391720727086067, + 0.009137891232967377, + 0.006170946639031172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.002818899229168892 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628357969224453 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628357969224453, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628357969224453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11798429489135742, + 0.032800108194351196, + 0.09626061469316483, + 0.036444853991270065, + 0.3804733455181122, + 0.0756433978676796, + 0.09213942289352417, + 0.02622472122311592, + 0.017479896545410156, + 0.12454935163259506 + ], + "credit_map": [ + 0.017463447526097298, + 0.026841366663575172, + 0.010457023046910763, + 0.015488461591303349, + 0.009455054067075253, + 0.019647790119051933, + 0.012114334851503372, + 0.006516812834888697, + 0.005518541671335697, + 0.0054933116771280766, + 0.003590032458305359, + 0.003154148580506444, + 0.006015073042362928, + 0.003871727269142866, + 0.003468672279268503, + 0.0016886000521481037, + 0.012017633765935898, + 0.0354413203895092, + 0.006357845384627581, + 0.009702364914119244, + 0.008050752803683281, + 0.01615835726261139, + 0.0049990154802799225, + 0.003533322596922517, + 0.005908846389502287, + 0.00750274583697319, + 0.0031665749847888947, + 0.004661551211029291, + 0.002745104720816016, + 0.005251006688922644, + 0.004753234330564737, + 0.002455787966027856, + 0.06996025145053864, + 0.11202061921358109, + 0.017458854243159294, + 0.020598389208316803, + 0.04055861756205559, + 0.0896742045879364, + 0.01547530759125948, + 0.01472706999629736, + 0.007725187111645937, + 0.020216861739754677, + 0.0054468270391225815, + 0.013103196397423744, + 0.006781115662306547, + 0.01221316959708929, + 0.005746923852711916, + 0.004410113673657179, + 0.019033323973417282, + 0.022250007838010788, + 0.0034856980200856924, + 0.005323201417922974, + 0.006795348133891821, + 0.022274088114500046, + 0.008208225481212139, + 0.004769521299749613, + 0.0031371847726404667, + 0.006278448272496462, + 0.0014852745225653052, + 0.004726610612124205, + 0.0029986996669322252, + 0.0038192267529666424, + 0.002033954719081521, + 0.0017453187610954046, + 0.002428296022117138, + 0.0034286524169147015, + 0.0016347895143553615, + 0.002382005797699094, + 0.0030723505187779665, + 0.00242850580252707, + 0.0013761698501184583, + 0.0007291267975233495, + 0.02030060812830925, + 0.027878612279891968, + 0.008239489048719406, + 0.007311220746487379, + 0.011366083286702633, + 0.03218317776918411, + 0.010371184907853603, + 0.006898961029946804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.007628357969224453 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures a dramatic confrontation between two large bullfrogs, rendered in the minimalist and highly expressive style of traditional Chinese ink painting. The scene unfolds against a stark off-white background, accented by the subtle fibrous texture of rice paper, creating a sense of concentrated intensity. In a display of dominance, one bullfrog adopts a powerful, grounded pose, its form defined by bold, wet strokes of deep black ink, depicting its muscular rump. Its throat pouch, bulging in defiance, is depicted in delicate, translucent light gray ink. The other frog retreats defensively, its body twisted into a tense arc, captured in a series of rapid, energetic, dry brushstrokes that convey frantic movement and struggle. At the center of the conflict, their powerful hind legs kick up a violent cloud of ink spray, exploding across the canvas in expressive, calligraphic splashes and dots, freezing the chaotic action with artistic precision. The monochromatic palette, ranging from the deepest black to aqua grey, contrasts with the pure white of the paper, creating a high-contrast, visually appealing atmosphere of raw, elegant conflict.", + "seed": 30174, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_13/index_2349_seed_30174_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.009792689280584455, + "vlm_logprob": -0.009792689280584455, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.009792689280584455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010967804118990898 + ], + "opposite_logprob_tokens": [ + -4.63596773147583 + ], + "logprob_reward": -0.010967804118990898, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010967804118990898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24051883816719055, + 0.09396563470363617, + 0.06219577044248581, + 0.07359099388122559, + 0.212258443236351, + 0.047795988619327545, + 0.06043357402086258, + 0.09603824466466904, + 0.05841812863945961, + 0.0547843798995018 + ], + "credit_map": [ + 0.028763167560100555, + 0.023756278678774834, + 0.029836134985089302, + 0.03161501884460449, + 0.04278989136219025, + 0.030382007360458374, + 0.033644985407590866, + 0.019731363281607628, + 0.01409292034804821, + 0.02122543752193451, + 0.009577316232025623, + 0.011363393627107143, + 0.01207472663372755, + 0.014446718618273735, + 0.005586839281022549, + 0.00559828570112586, + 0.007948486134409904, + 0.010427269153296947, + 0.007670884486287832, + 0.007322036195546389, + 0.007548450957983732, + 0.006505454424768686, + 0.008863696828484535, + 0.0059094903990626335, + 0.008850245736539364, + 0.01740766502916813, + 0.008050218224525452, + 0.010499687865376472, + 0.008402503095567226, + 0.010001947171986103, + 0.006407049018889666, + 0.003971675876528025, + 0.020612722262740135, + 0.05459078773856163, + 0.018912404775619507, + 0.03309604898095131, + 0.025007378309965134, + 0.03543112426996231, + 0.016942722722887993, + 0.007665255106985569, + 0.005421922076493502, + 0.007733782287687063, + 0.004044596571475267, + 0.0061197965405881405, + 0.007500726729631424, + 0.007431091275066137, + 0.006096986588090658, + 0.0034470863174647093, + 0.00870552472770214, + 0.01612497679889202, + 0.007129507604986429, + 0.006305133458226919, + 0.007874730043113232, + 0.005315261427313089, + 0.005759601015597582, + 0.0032188387122005224, + 0.013557279482483864, + 0.02001825161278248, + 0.011829863302409649, + 0.013770455494523048, + 0.012927752919495106, + 0.010492640547454357, + 0.007815245538949966, + 0.005626760423183441, + 0.010135015472769737, + 0.011098896153271198, + 0.007417941465973854, + 0.005629223771393299, + 0.0075021893717348576, + 0.00929221324622631, + 0.00391614343971014, + 0.003426506882533431, + 0.008643140085041523, + 0.010849115438759327, + 0.003777199424803257, + 0.003602350363507867, + 0.007822827436029911, + 0.010625570081174374, + 0.004768022336065769, + 0.00469615263864398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.010967804118990898 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864465907216072 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.00864465907216072, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864465907216072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1437220722436905, + 0.0512530617415905, + 0.08468955755233765, + 0.08532960712909698, + 0.15014676749706268, + 0.07835011929273605, + 0.06093689054250717, + 0.1225307285785675, + 0.10551399737596512, + 0.11752718687057495 + ], + "credit_map": [ + 0.02512526884675026, + 0.01583017408847809, + 0.016237612813711166, + 0.024187682196497917, + 0.018008999526500702, + 0.018070975318551064, + 0.01731080934405327, + 0.008950551971793175, + 0.006601706147193909, + 0.007998628541827202, + 0.005968489684164524, + 0.00956424418836832, + 0.0066841705702245235, + 0.007011326961219311, + 0.003946057520806789, + 0.0034784392919391394, + 0.01231854222714901, + 0.02537357620894909, + 0.004940900951623917, + 0.00661950558423996, + 0.011359235271811485, + 0.00864786934107542, + 0.008956997655332088, + 0.006472927052527666, + 0.010712363757193089, + 0.018798064440488815, + 0.00885444413870573, + 0.017916640266776085, + 0.00898479763418436, + 0.010997713543474674, + 0.005436866078525782, + 0.0036287151742726564, + 0.0123869888484478, + 0.04806395247578621, + 0.014309054240584373, + 0.01970840059220791, + 0.01780410297214985, + 0.024351008236408234, + 0.00853049661964178, + 0.004992768634110689, + 0.007032856345176697, + 0.013802861794829369, + 0.007079399190843105, + 0.009463410824537277, + 0.013157577253878117, + 0.011660893447697163, + 0.011639624834060669, + 0.004513496998697519, + 0.009583770297467709, + 0.018265241757035255, + 0.005871487315744162, + 0.005958758760243654, + 0.007454222068190575, + 0.006196001544594765, + 0.00454024039208889, + 0.0030671695712953806, + 0.026663053780794144, + 0.02525792084634304, + 0.012367960065603256, + 0.013094727881252766, + 0.013380621559917927, + 0.015011254698038101, + 0.009275302290916443, + 0.007479888387024403, + 0.024731414392590523, + 0.012683597393333912, + 0.007837384007871151, + 0.011607781052589417, + 0.010190477594733238, + 0.02683381736278534, + 0.00623492943122983, + 0.00539459427818656, + 0.021764397621154785, + 0.02372787706553936, + 0.007458025589585304, + 0.011041739955544472, + 0.013895807787775993, + 0.021573912352323532, + 0.007702269591391087, + 0.010363162495195866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.00864465907216072 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028217521030455828 + ], + "opposite_logprob_tokens": [ + -5.877821922302246 + ], + "logprob_reward": -0.0028217521030455828, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028217521030455828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20709237456321716, + 0.03586422652006149, + 0.05609988048672676, + 0.12168961763381958, + 0.15657998621463776, + 0.060437723994255066, + 0.05619410052895546, + 0.04025351628661156, + 0.06806008517742157, + 0.1977284848690033 + ], + "credit_map": [ + 0.023287495598196983, + 0.02190525084733963, + 0.02495851367712021, + 0.031868599355220795, + 0.019951194524765015, + 0.026099372655153275, + 0.032227035611867905, + 0.026794912293553352, + 0.004983869846910238, + 0.007235880941152573, + 0.0038199268747121096, + 0.004030258860439062, + 0.0034636601340025663, + 0.005156791768968105, + 0.002829340985044837, + 0.004344498738646507, + 0.0064160022884607315, + 0.012512427754700184, + 0.008886177092790604, + 0.006263012997806072, + 0.005472892429679632, + 0.004681610036641359, + 0.007537403143942356, + 0.004330356139689684, + 0.014172924682497978, + 0.03799962252378464, + 0.0062919664196670055, + 0.021780261769890785, + 0.013506636954843998, + 0.017970098182559013, + 0.005675945430994034, + 0.004292170517146587, + 0.014713280834257603, + 0.04551588371396065, + 0.01735205203294754, + 0.023705486208200455, + 0.01585898920893669, + 0.024490373209118843, + 0.008870991878211498, + 0.006072937045246363, + 0.006399661768227816, + 0.013311987742781639, + 0.0053130812011659145, + 0.006068413611501455, + 0.007794840261340141, + 0.00922746304422617, + 0.00904591754078865, + 0.003276358125731349, + 0.007628017570823431, + 0.015093831345438957, + 0.005398667883127928, + 0.007825345732271671, + 0.007370030973106623, + 0.00593508081510663, + 0.0044397697784006596, + 0.002503358293324709, + 0.004660527687519789, + 0.00659678690135479, + 0.004639473278075457, + 0.0070582362823188305, + 0.007209652569144964, + 0.0048171174712479115, + 0.00283702346496284, + 0.0024346972350031137, + 0.009939439594745636, + 0.012245767749845982, + 0.005869164597243071, + 0.0090402290225029, + 0.009924632497131824, + 0.008407451212406158, + 0.00846625491976738, + 0.004167144652456045, + 0.03652699291706085, + 0.059721384197473526, + 0.012253904715180397, + 0.01608833484351635, + 0.02732611447572708, + 0.024731168523430824, + 0.013190691359341145, + 0.007889903150498867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.0028217521030455828 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059472741559147835 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.0059472741559147835, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059472741559147835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1878252774477005, + 0.05541327968239784, + 0.08826566487550735, + 0.13374225795269012, + 0.0729483887553215, + 0.08284028619527817, + 0.08229929208755493, + 0.0656396821141243, + 0.15605497360229492, + 0.07497093826532364 + ], + "credit_map": [ + 0.035924408584833145, + 0.020688433200120926, + 0.019027018919587135, + 0.022493721917271614, + 0.021159734576940536, + 0.03071230836212635, + 0.020030854269862175, + 0.017788777127861977, + 0.0070264884270727634, + 0.00949654821306467, + 0.006264062598347664, + 0.006382439285516739, + 0.007212891709059477, + 0.0076600657775998116, + 0.005859991069883108, + 0.005510792601853609, + 0.009670116938650608, + 0.01717410609126091, + 0.013574867509305477, + 0.010735121555626392, + 0.00905604474246502, + 0.007749052252620459, + 0.012997292913496494, + 0.007309055421501398, + 0.017878852784633636, + 0.026390090584754944, + 0.012500438839197159, + 0.033840570598840714, + 0.01257774606347084, + 0.015587280504405499, + 0.008192609995603561, + 0.0067746699787676334, + 0.00625555869191885, + 0.014267221093177795, + 0.007793905679136515, + 0.016821788623929024, + 0.007705533877015114, + 0.010893391445279121, + 0.004629989620298147, + 0.004580999258905649, + 0.0078079355880618095, + 0.01606486365199089, + 0.007364237681031227, + 0.008127066306769848, + 0.015378846786916256, + 0.01223759911954403, + 0.010378031060099602, + 0.0054816994816064835, + 0.011928318999707699, + 0.020748523995280266, + 0.006937146186828613, + 0.011650546453893185, + 0.011119424365460873, + 0.009195868857204914, + 0.006692745257169008, + 0.0040267277508974075, + 0.010712753981351852, + 0.012396672740578651, + 0.007732657715678215, + 0.00766671821475029, + 0.010503098368644714, + 0.008114463649690151, + 0.004596223589032888, + 0.003917100373655558, + 0.028521927073597908, + 0.027109449729323387, + 0.012358182109892368, + 0.021042272448539734, + 0.015587625093758106, + 0.02192598767578602, + 0.02106538601219654, + 0.008444137871265411, + 0.013407503254711628, + 0.020545784384012222, + 0.004245080519467592, + 0.006107289344072342, + 0.010684776119887829, + 0.01109694316983223, + 0.0044654132798314095, + 0.004418147727847099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.0059472741559147835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581956952810287 + ], + "opposite_logprob_tokens": [ + -3.8955819606781006 + ], + "logprob_reward": -0.020581956952810287, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581956952810287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11594005674123764, + 0.09124938398599625, + 0.13469387590885162, + 0.07806936651468277, + 0.07763411104679108, + 0.1030283197760582, + 0.1004379466176033, + 0.16349095106124878, + 0.05370430275797844, + 0.08175163716077805 + ], + "credit_map": [ + 0.021074587479233742, + 0.01190579030662775, + 0.011209726333618164, + 0.018039969727396965, + 0.017164235934615135, + 0.017828071489930153, + 0.010984880849719048, + 0.0077327946200966835, + 0.01340118981897831, + 0.02222614362835884, + 0.008647664450109005, + 0.01039468590170145, + 0.007151393219828606, + 0.015340371057391167, + 0.0063379015773534775, + 0.007750033866614103, + 0.021937187761068344, + 0.03734318166971207, + 0.010041274130344391, + 0.011430886574089527, + 0.01289665699005127, + 0.01566772535443306, + 0.01573663391172886, + 0.009640336968004704, + 0.008354557678103447, + 0.02183513157069683, + 0.0066424584947526455, + 0.009565969929099083, + 0.009273441508412361, + 0.01155160367488861, + 0.0066367550753057, + 0.004209449980407953, + 0.006812651641666889, + 0.016757972538471222, + 0.009658927097916603, + 0.009820155799388885, + 0.010446688160300255, + 0.0130218006670475, + 0.006355236750096083, + 0.004760675597935915, + 0.010048117488622665, + 0.025175433605909348, + 0.007012620102614164, + 0.01163414865732193, + 0.01571752317249775, + 0.01692153513431549, + 0.011026706546545029, + 0.005492235533893108, + 0.015412870794534683, + 0.0337686650454998, + 0.010357157327234745, + 0.00880615133792162, + 0.01072327047586441, + 0.009149006567895412, + 0.00827090721577406, + 0.0039499239064753056, + 0.0396837517619133, + 0.024352993816137314, + 0.01582547463476658, + 0.026356495916843414, + 0.012530811131000519, + 0.023393871262669563, + 0.009505211375653744, + 0.011842354200780392, + 0.010198593139648438, + 0.007396569475531578, + 0.004392951261252165, + 0.006167997606098652, + 0.006189208943396807, + 0.010890730656683445, + 0.005741134285926819, + 0.002727117855101824, + 0.01481273490935564, + 0.02045479416847229, + 0.004500350449234247, + 0.006594989914447069, + 0.009448324330151081, + 0.014092842116951942, + 0.005173764657229185, + 0.006673840340226889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.020581956952810287 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures a dramatic confrontation between two large bullfrogs, rendered in the minimalist and highly expressive style of traditional Chinese ink painting. The scene unfolds against a stark off-white background, accented by the subtle fibrous texture of rice paper, creating a sense of concentrated intensity. In a display of dominance, one bullfrog adopts a powerful, grounded pose, its form defined by bold, wet strokes of deep black ink, depicting its muscular rump. Its throat pouch, bulging in defiance, is depicted in delicate, translucent light gray ink. The other frog retreats defensively, its body twisted into a tense arc, captured in a series of rapid, energetic, dry brushstrokes that convey frantic movement and struggle. At the center of the conflict, their powerful hind legs kick up a violent cloud of ink spray, exploding across the canvas in expressive, calligraphic splashes and dots, freezing the chaotic action with artistic precision. The monochromatic palette, ranging from the deepest black to aqua grey, contrasts with the pure white of the paper, creating a high-contrast, visually appealing atmosphere of raw, elegant conflict.", + "seed": 30175, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_13/index_2349_seed_30175_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010866677179001272, + "vlm_logprob": -0.010866677179001272, + "vlm_yes_no_margin": 4.849999952316284, + "vlm_token_credit": -0.010866677179001272, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02151273377239704 + ], + "opposite_logprob_tokens": [ + -3.896512746810913 + ], + "logprob_reward": -0.02151273377239704, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02151273377239704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15470142662525177, + 0.2406778633594513, + 0.01352222915738821, + 0.047882575541734695, + 0.01587696187198162, + 0.47883760929107666, + 0.003198757069185376, + 0.02120056562125683, + 0.005134067963808775, + 0.018967870622873306 + ], + "credit_map": [ + 0.05782531574368477, + 0.008145877160131931, + 0.012712543830275536, + 0.0037843731697648764, + 0.0036344442050904036, + 0.06174764782190323, + 0.0039827218279242516, + 0.0028685315046459436, + 0.06686624884605408, + 0.04382607713341713, + 0.008937676437199116, + 0.0046869986690580845, + 0.0075081102550029755, + 0.1004018783569336, + 0.00375840300694108, + 0.0046925051137804985, + 0.0023480961099267006, + 0.0020180915016680956, + 0.0011818689526990056, + 0.0008402324165217578, + 0.0022504897788167, + 0.0020690131932497025, + 0.0016000736504793167, + 0.0012143641943112016, + 0.01679241470992565, + 0.00283957296051085, + 0.0034552807919681072, + 0.0013462645001709461, + 0.0020899332594126463, + 0.018419120460748672, + 0.0016566230915486813, + 0.0012833725195378065, + 0.002144890371710062, + 0.0021159243769943714, + 0.0027292328886687756, + 0.0012216134928166866, + 0.001774126896634698, + 0.0018259396310895681, + 0.0023067689035087824, + 0.0017584663582965732, + 0.06872767210006714, + 0.08529927581548691, + 0.03087773360311985, + 0.019241781905293465, + 0.15481054782867432, + 0.06087924912571907, + 0.03536860644817352, + 0.02363278716802597, + 0.00048614456318318844, + 0.0005719860200770199, + 0.00032461644150316715, + 0.00020192578085698187, + 0.000545890477951616, + 0.0003773686767090112, + 0.00035943201510235667, + 0.0003313935303594917, + 0.007299533113837242, + 0.0023456166964024305, + 0.000974484661128372, + 0.0006858273991383612, + 0.0010567938443273306, + 0.007333718240261078, + 0.0008119088597595692, + 0.0006926854257471859, + 0.0008659482118673623, + 0.0008464322309009731, + 0.000600297236815095, + 0.000551197852473706, + 0.0005868691368959844, + 0.0005470709293149412, + 0.0006769367028027773, + 0.00045931656495667994, + 0.0037676040083169937, + 0.0021034665405750275, + 0.0015403712168335915, + 0.0008002924150787294, + 0.0032037270721048117, + 0.004015037324279547, + 0.0021291542798280716, + 0.0014082189882174134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.02151273377239704 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743649020791054 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.006743649020791054, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743649020791054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05394838750362396, + 0.4007835388183594, + 0.02440478466451168, + 0.044669121503829956, + 0.04470481723546982, + 0.20362307131290436, + 0.02286587283015251, + 0.02956993691623211, + 0.02822805568575859, + 0.14720243215560913 + ], + "credit_map": [ + 0.014323567040264606, + 0.006730049382895231, + 0.0053354110568761826, + 0.002372817136347294, + 0.004963228013366461, + 0.012942234054207802, + 0.004296130035072565, + 0.002984949853271246, + 0.1078098937869072, + 0.06629520654678345, + 0.017759382724761963, + 0.009243782609701157, + 0.020089272409677505, + 0.1563212275505066, + 0.011853923089802265, + 0.011410859413444996, + 0.0038703265599906445, + 0.0028626685962080956, + 0.003024138743057847, + 0.0017709056846797466, + 0.0035605558659881353, + 0.00422251783311367, + 0.0025919286999851465, + 0.0025017415173351765, + 0.010583005845546722, + 0.007049651350826025, + 0.0034967174287885427, + 0.002138731302693486, + 0.005354555323719978, + 0.010396570898592472, + 0.0035907793790102005, + 0.002059111837297678, + 0.010508091188967228, + 0.007942669093608856, + 0.009260192513465881, + 0.002351752482354641, + 0.0029097292572259903, + 0.005358509253710508, + 0.002638395642861724, + 0.003735480597242713, + 0.029909661039710045, + 0.036539651453495026, + 0.013001993298530579, + 0.008536607027053833, + 0.06339292973279953, + 0.02543724700808525, + 0.015320259146392345, + 0.01148472260683775, + 0.0035538505762815475, + 0.002680219244211912, + 0.0037889874074608088, + 0.0011730828555300832, + 0.003088124329224229, + 0.003027188591659069, + 0.002969225635752082, + 0.0025851933751255274, + 0.005748117808252573, + 0.005768523085862398, + 0.002338046906515956, + 0.001803852734155953, + 0.003739419160410762, + 0.005488493479788303, + 0.0025370479561388493, + 0.0021464370656758547, + 0.0044981949031353, + 0.004521805327385664, + 0.003231948707252741, + 0.0028162410017102957, + 0.00511946389451623, + 0.0030206928495317698, + 0.002831862773746252, + 0.0021878464613109827, + 0.019913190975785255, + 0.019620811566710472, + 0.012200753204524517, + 0.011088310740888119, + 0.0241807010024786, + 0.020955640822649002, + 0.019874773919582367, + 0.019368240609765053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.006743649020791054 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001943010720424354 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.001943010720424354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001943010720424354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.265495240688324, + 0.37186428904533386, + 0.020893698558211327, + 0.03117547556757927, + 0.05061924457550049, + 0.052920009940862656, + 0.032395824790000916, + 0.08318378776311874, + 0.010268760845065117, + 0.08118361979722977 + ], + "credit_map": [ + 0.09333773702383041, + 0.015022195875644684, + 0.021676789969205856, + 0.006547397933900356, + 0.011692723259329796, + 0.10405919700860977, + 0.007691832259297371, + 0.005467379465699196, + 0.10535625368356705, + 0.06414929032325745, + 0.014042352326214314, + 0.007197779603302479, + 0.01300225779414177, + 0.15226736664772034, + 0.007542935200035572, + 0.00830606184899807, + 0.0033862439449876547, + 0.0029762103222310543, + 0.0032068337313830853, + 0.0013069435954093933, + 0.004498443566262722, + 0.002706593833863735, + 0.0014937813393771648, + 0.0013186500873416662, + 0.00495149614289403, + 0.00505004171282053, + 0.0028570580761879683, + 0.002509928308427334, + 0.003659179899841547, + 0.004647890571504831, + 0.005231755319982767, + 0.0022681267000734806, + 0.012541862204670906, + 0.008945937268435955, + 0.009302468970417976, + 0.002932879375293851, + 0.003759875427931547, + 0.005441784858703613, + 0.003187899012118578, + 0.004506535828113556, + 0.0076164971105754375, + 0.009626311250030994, + 0.003592562163248658, + 0.0023646613117307425, + 0.01641281507909298, + 0.006776000373065472, + 0.00376023817807436, + 0.0027709216810762882, + 0.006168061401695013, + 0.004217370878905058, + 0.004551585763692856, + 0.0024498600978404284, + 0.003953457809984684, + 0.005050544627010822, + 0.0029663504101336002, + 0.003038594266399741, + 0.020474523305892944, + 0.021521124988794327, + 0.005984631832689047, + 0.002404425758868456, + 0.006319808308035135, + 0.020689688622951508, + 0.0028557623736560345, + 0.002933815587311983, + 0.0016126803820952773, + 0.0019958161283284426, + 0.0010946516413241625, + 0.0009928405052050948, + 0.0013894649455323815, + 0.0011397864436730742, + 0.0011569573543965816, + 0.0008865633863024414, + 0.01286401879042387, + 0.010806703940033913, + 0.006625024601817131, + 0.005564403720200062, + 0.013303137384355068, + 0.01456258725374937, + 0.008846690878272057, + 0.008611057884991169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.001943010720424354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059472741559147835 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.0059472741559147835, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059472741559147835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07130616158246994, + 0.1890152096748352, + 0.03710921108722687, + 0.14246121048927307, + 0.03251836076378822, + 0.3017394244670868, + 0.04734702780842781, + 0.07620221376419067, + 0.02478937804698944, + 0.07751183956861496 + ], + "credit_map": [ + 0.016102930530905724, + 0.014891743659973145, + 0.005552247166633606, + 0.0034977993927896023, + 0.009224443696439266, + 0.010933510959148407, + 0.0068261390551924706, + 0.004277339670807123, + 0.042640551924705505, + 0.030946528539061546, + 0.012686105445027351, + 0.007778756320476532, + 0.015561635605990887, + 0.059853289276361465, + 0.009309589862823486, + 0.010238738730549812, + 0.005998071748763323, + 0.004600148182362318, + 0.006274688523262739, + 0.002240272704511881, + 0.006392020732164383, + 0.005544576793909073, + 0.003046134952455759, + 0.0030132958199828863, + 0.046460725367069244, + 0.009915080852806568, + 0.010205515660345554, + 0.004963210318237543, + 0.00934614147990942, + 0.04958145692944527, + 0.007072559557855129, + 0.00491650914773345, + 0.006555676460266113, + 0.0046952227130532265, + 0.007131765130907297, + 0.001767375972121954, + 0.0029843852389603853, + 0.0038275225088000298, + 0.0023503960110247135, + 0.003206014633178711, + 0.04349769279360771, + 0.05400929972529411, + 0.018241189420223236, + 0.011699235066771507, + 0.09908124804496765, + 0.03982445225119591, + 0.021810635924339294, + 0.01357566099613905, + 0.0072432877495884895, + 0.006125225685536861, + 0.007916521281003952, + 0.0025222718250006437, + 0.0073061613366007805, + 0.006317892577499151, + 0.005466741044074297, + 0.004448921885341406, + 0.01611112616956234, + 0.018362965434789658, + 0.0068222698755562305, + 0.0033804529812186956, + 0.006116956006735563, + 0.013802245259284973, + 0.0049674720503389835, + 0.006638721097260714, + 0.0038754623383283615, + 0.003701427485793829, + 0.00332472613081336, + 0.0023502979893237352, + 0.003937803208827972, + 0.003082234412431717, + 0.002444646554067731, + 0.0020727759692817926, + 0.014530934393405914, + 0.00830616895109415, + 0.006090526469051838, + 0.0031904641073197126, + 0.013632295653223991, + 0.016368495300412178, + 0.008580719120800495, + 0.006812223233282566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.0059472741559147835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two large bullfrogs visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Chinese ink painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background stark off-white with a subtle fibrous rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the bullfrogs' hind legs kick up a cloud of ink spray with calligraphic splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color palette monochromatic, ranging from deep black to aqua grey?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186718225479126 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.018186718225479126, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018186718225479126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13446564972400665, + 0.3620886504650116, + 0.07127886265516281, + 0.05945713445544243, + 0.05603926256299019, + 0.09788960218429565, + 0.025680338963866234, + 0.080958291888237, + 0.06221656873822212, + 0.04992563650012016 + ], + "credit_map": [ + 0.04352376610040665, + 0.015880044549703598, + 0.010962851345539093, + 0.004682433791458607, + 0.008755750954151154, + 0.03959310054779053, + 0.006411056034266949, + 0.0046566566452383995, + 0.09596126526594162, + 0.05922948569059372, + 0.026447486132383347, + 0.015848394483327866, + 0.022326214239001274, + 0.11300811916589737, + 0.013168102130293846, + 0.01609957404434681, + 0.011593851260840893, + 0.007456696592271328, + 0.00836270209401846, + 0.004255311097949743, + 0.017429055646061897, + 0.011122772470116615, + 0.006247621960937977, + 0.0048108454793691635, + 0.009294611401855946, + 0.013520576059818268, + 0.00635302159935236, + 0.003562632016837597, + 0.00820140726864338, + 0.00789092481136322, + 0.006232280284166336, + 0.00440168147906661, + 0.010204128921031952, + 0.010496346279978752, + 0.013564801774919033, + 0.004577333107590675, + 0.003158960957080126, + 0.005222659092396498, + 0.004169480409473181, + 0.004645553883165121, + 0.01431781891733408, + 0.017344072461128235, + 0.007524610031396151, + 0.004071858245879412, + 0.030261848121881485, + 0.012679792940616608, + 0.006731078960001469, + 0.00495852530002594, + 0.004674422554671764, + 0.003509040456265211, + 0.004345047287642956, + 0.0014565413584932685, + 0.0036258974578231573, + 0.0031886117067188025, + 0.002838052809238434, + 0.0020427240524441004, + 0.02106057107448578, + 0.01555325835943222, + 0.005055414512753487, + 0.0034527399111539125, + 0.005990167148411274, + 0.020717287436127663, + 0.004992372822016478, + 0.004136484116315842, + 0.009242710657417774, + 0.007738409098237753, + 0.006572789046913385, + 0.005687561351805925, + 0.01093219593167305, + 0.008280911482870579, + 0.007917826995253563, + 0.0058441609144210815, + 0.007715312764048576, + 0.006243956740945578, + 0.005911280866712332, + 0.004795336630195379, + 0.008081012405455112, + 0.006712546106427908, + 0.005278053693473339, + 0.005188138224184513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.018186718225479126 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_14/index_tensor([3262, 3262])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_14/index_tensor([3262, 3262])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..64dfe350ce6494d80d03f67e144b3e9ad0944089 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_14/index_tensor([3262, 3262])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot captures a scene rendered in a somber, painterly style characterized by thick, visible impasto and heavy, dark pigment. At the center of the painting, a gnarled, twisted tree root, like a small, emaciated human figure, is half-swallowed by an oppressive, inky shadow. A cold, directional light sweeps across its upper portion, highlighting the sheen of oil on its fibrous, bark-like texture and the intricate knots that suggest rough joints and a bowed head. The figure's lower half disappears into an impenetrable darkness, creating a dramatic effect of chiaroscuro. The root form is not static; it subtly contracts and oscillates with a slow, rhythmic pulse, its form viscerally trembling, as if breathing. This tiny, steady movement—a slight inward pull followed by a gentle sway—transforms the still life into a deeply unsettling and melancholic spectacle, a living sculpture that seems to breathe within the suffocating painted darkness.", + "seed": 20823, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_14/index_3262_seed_20823_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02074846103787422, + "vlm_logprob": -0.02074846103787422, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.02074846103787422, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03235916048288345 + ], + "opposite_logprob_tokens": [ + -4.2823591232299805 + ], + "logprob_reward": -0.03235916048288345, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.03235916048288345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10280589014291763, + 0.3232993185520172, + 0.06630761176347733, + 0.03180146589875221, + 0.03819317743182182, + 0.05675491318106651, + 0.20245686173439026, + 0.03889509662985802, + 0.025910215452313423, + 0.11357542872428894 + ], + "credit_map": [ + 0.010770188644528389, + 0.005439062137156725, + 0.008666429668664932, + 0.013592333532869816, + 0.015802906826138496, + 0.01761500909924507, + 0.01240542996674776, + 0.01851453073322773, + 0.02502521313726902, + 0.021266430616378784, + 0.03369053825736046, + 0.03257855772972107, + 0.058189891278743744, + 0.062444038689136505, + 0.031043771654367447, + 0.05906087905168533, + 0.005112210754305124, + 0.0038945528212934732, + 0.006439042277634144, + 0.006694776471704245, + 0.012684923596680164, + 0.013955739326775074, + 0.006813996005803347, + 0.01071237027645111, + 0.0027347940485924482, + 0.0025061217602342367, + 0.0036360363010317087, + 0.003614225657656789, + 0.00430815014988184, + 0.005346398800611496, + 0.004032680299133062, + 0.005623058881610632, + 0.004033129196614027, + 0.002339349826797843, + 0.003373888786882162, + 0.007392883766442537, + 0.004731024615466595, + 0.00552083645015955, + 0.003904959885403514, + 0.006897106766700745, + 0.005941716022789478, + 0.002265679184347391, + 0.004808053374290466, + 0.007420605979859829, + 0.006046503782272339, + 0.01081757340580225, + 0.0038867248222231865, + 0.01556805893778801, + 0.0170219074934721, + 0.011548737995326519, + 0.021518250927329063, + 0.019939683377742767, + 0.03497551754117012, + 0.04558086767792702, + 0.016583435237407684, + 0.03528846055269241, + 0.004119839984923601, + 0.0022567755077034235, + 0.003696280065923929, + 0.005359974689781666, + 0.005827685352414846, + 0.007325196638703346, + 0.004092232324182987, + 0.006217113230377436, + 0.0016251159831881523, + 0.0011669319355860353, + 0.0029966398142278194, + 0.005430352874100208, + 0.003965058363974094, + 0.00368904578499496, + 0.003772362135350704, + 0.003264708211645484, + 0.010284950025379658, + 0.006045246962457895, + 0.009388147853314877, + 0.011195586062967777, + 0.016871923580765724, + 0.02335832267999649, + 0.00846273172646761, + 0.027968518435955048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.03235916048288345 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233188197016716 + ], + "opposite_logprob_tokens": [ + -3.7733187675476074 + ], + "logprob_reward": -0.0233188197016716, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233188197016716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22679468989372253, + 0.225955069065094, + 0.06835530698299408, + 0.031059274449944496, + 0.07047133892774582, + 0.05678558722138405, + 0.04650074988603592, + 0.16518044471740723, + 0.06737233698368073, + 0.041525162756443024 + ], + "credit_map": [ + 0.020443949848413467, + 0.01424553245306015, + 0.028897332027554512, + 0.03339068219065666, + 0.03519801050424576, + 0.03589522838592529, + 0.029862234368920326, + 0.028861738741397858, + 0.024677377194166183, + 0.019233133643865585, + 0.031428366899490356, + 0.03127961978316307, + 0.0330323688685894, + 0.033946968615055084, + 0.021145692095160484, + 0.03121154010295868, + 0.006465272046625614, + 0.004346283618360758, + 0.006132497917860746, + 0.006387440022081137, + 0.011453909799456596, + 0.014243309386074543, + 0.00807943008840084, + 0.011247165501117706, + 0.0026842679362744093, + 0.0017430431907996535, + 0.004196352791041136, + 0.0035164004657417536, + 0.00399486580863595, + 0.0047960225492715836, + 0.004138082265853882, + 0.005990243051201105, + 0.004967179615050554, + 0.004069971386343241, + 0.006377135869115591, + 0.014339742250740528, + 0.008751049637794495, + 0.010612367652356625, + 0.006166333332657814, + 0.015187562443315983, + 0.004924936685711145, + 0.002250431338325143, + 0.006839974317699671, + 0.006929386407136917, + 0.007907594554126263, + 0.01038421131670475, + 0.005718281492590904, + 0.011830773204565048, + 0.0036395720671862364, + 0.00289380201138556, + 0.004670850001275539, + 0.0033496487885713577, + 0.009736877866089344, + 0.009507364593446255, + 0.005302036181092262, + 0.0074005997739732265, + 0.01271973829716444, + 0.009444665163755417, + 0.014025372453033924, + 0.01976088434457779, + 0.027127429842948914, + 0.031233036890625954, + 0.015473629347980022, + 0.035395704209804535, + 0.007335928734391928, + 0.003567075589671731, + 0.005083570256829262, + 0.005209009628742933, + 0.0085715651512146, + 0.011606684885919094, + 0.007486804388463497, + 0.018511688336730003, + 0.0037133018486201763, + 0.0031240596435964108, + 0.005874462425708771, + 0.0057359724305570126, + 0.006935785990208387, + 0.00646595936268568, + 0.004819123074412346, + 0.004856498911976814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0233188197016716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022031813859939575 + ], + "opposite_logprob_tokens": [ + -6.127202987670898 + ], + "logprob_reward": -0.0022031813859939575, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022031813859939575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1683468222618103, + 0.3517361581325531, + 0.14037610590457916, + 0.03735075145959854, + 0.10400209575891495, + 0.05857374519109726, + 0.06136596202850342, + 0.030490895733237267, + 0.021710040047764778, + 0.026047438383102417 + ], + "credit_map": [ + 0.012685408815741539, + 0.014449741691350937, + 0.021090349182486534, + 0.02811313606798649, + 0.028311125934123993, + 0.025583766400814056, + 0.022054996341466904, + 0.016058295965194702, + 0.028821207582950592, + 0.027890346944332123, + 0.04713192954659462, + 0.036556877195835114, + 0.06667176634073257, + 0.06507201492786407, + 0.02998967096209526, + 0.049602314829826355, + 0.010883196257054806, + 0.008607755415141582, + 0.013037005439400673, + 0.01408330723643303, + 0.025166692212224007, + 0.028293700888752937, + 0.012760953046381474, + 0.027543487027287483, + 0.0036963550373911858, + 0.0015660099452361465, + 0.0032847649417817593, + 0.004858754575252533, + 0.0049097430892288685, + 0.007895098067820072, + 0.004092717077583075, + 0.007047307211905718, + 0.009312709793448448, + 0.0059546539559960365, + 0.01010733563452959, + 0.012070986442267895, + 0.012660881504416466, + 0.017376212403178215, + 0.009679761715233326, + 0.026839550584554672, + 0.005197602789849043, + 0.0044692521914839745, + 0.009413659572601318, + 0.011081735603511333, + 0.009223567321896553, + 0.007621357217431068, + 0.004955886863172054, + 0.0066106850281357765, + 0.00520450621843338, + 0.0032086020801216364, + 0.006414462346583605, + 0.00811149924993515, + 0.007305166218429804, + 0.010671858675777912, + 0.005788594484329224, + 0.014661273919045925, + 0.002386592561379075, + 0.0018939203582704067, + 0.003248803550377488, + 0.0030738492496311665, + 0.006680862512439489, + 0.006565061863511801, + 0.003041474847123027, + 0.0036003286950290203, + 0.0020754134748131037, + 0.0014390507712960243, + 0.003166902344673872, + 0.0027789378073066473, + 0.00343831954523921, + 0.0038508796133100986, + 0.0025669902097433805, + 0.002393545815721154, + 0.0022774061653763056, + 0.002019454725086689, + 0.003317573806270957, + 0.0037033644039183855, + 0.004717562347650528, + 0.004296204540878534, + 0.0023438516072928905, + 0.003372020088136196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0022031813859939575 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160689540207386 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.0160689540207386, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160689540207386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13507427275180817, + 0.3988789916038513, + 0.12223787605762482, + 0.026651259511709213, + 0.048015791922807693, + 0.06015671044588089, + 0.049618661403656006, + 0.06328808516263962, + 0.05519425868988037, + 0.04088404402136803 + ], + "credit_map": [ + 0.01174979843199253, + 0.008020703680813313, + 0.013861686922609806, + 0.019987722858786583, + 0.018997738137841225, + 0.021059907972812653, + 0.017613627016544342, + 0.023783095180988312, + 0.02897653728723526, + 0.02693335898220539, + 0.04127730429172516, + 0.0337158665060997, + 0.07926108688116074, + 0.08297786116600037, + 0.03594932705163956, + 0.06978767365217209, + 0.012632871977984905, + 0.00816249754279852, + 0.008634441532194614, + 0.012383973225951195, + 0.015741292387247086, + 0.02172614261507988, + 0.009136272594332695, + 0.033820394426584244, + 0.002521642716601491, + 0.001934138941578567, + 0.0026158506516367197, + 0.005190474912524223, + 0.004179771989583969, + 0.0037178483325988054, + 0.0036587899085134268, + 0.002832740778103471, + 0.0037236048374325037, + 0.0023099014069885015, + 0.0050841704942286015, + 0.0067817093804478645, + 0.007240114267915487, + 0.007903694175183773, + 0.004572369623929262, + 0.010400224477052689, + 0.005665208678692579, + 0.0036899768747389317, + 0.005801314953714609, + 0.00888978410512209, + 0.008854522369801998, + 0.009327784180641174, + 0.0047494033351540565, + 0.013178711757063866, + 0.004480795469135046, + 0.002513716695830226, + 0.004873552359640598, + 0.004617222584784031, + 0.007522079162299633, + 0.012150109745562077, + 0.003992103971540928, + 0.009469082579016685, + 0.004156310111284256, + 0.003791307331994176, + 0.00442076800391078, + 0.005995648447424173, + 0.012904834002256393, + 0.015276042744517326, + 0.007017683237791061, + 0.009725493378937244, + 0.005857196636497974, + 0.003031931584700942, + 0.004070870578289032, + 0.004692425020039082, + 0.006780144292861223, + 0.009604478254914284, + 0.0059587848372757435, + 0.015198426321148872, + 0.00339390616863966, + 0.0033520220313221216, + 0.004282232839614153, + 0.007906071841716766, + 0.005316365044564009, + 0.005634210538119078, + 0.003647135803475976, + 0.007352098356932402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0160689540207386 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029792189598083496 + ], + "opposite_logprob_tokens": [ + -3.529792308807373 + ], + "logprob_reward": -0.029792189598083496, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029792189598083496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06112176924943924, + 0.12623271346092224, + 0.17305110394954681, + 0.06052852421998978, + 0.06314840167760849, + 0.0781724825501442, + 0.1793888509273529, + 0.0637187585234642, + 0.110756054520607, + 0.08388134837150574 + ], + "credit_map": [ + 0.004085540305823088, + 0.0031501897610723972, + 0.006100027821958065, + 0.009172080084681511, + 0.010210247710347176, + 0.009749404154717922, + 0.00819080974906683, + 0.010463470593094826, + 0.009165173396468163, + 0.009486719965934753, + 0.01169962901622057, + 0.01299536507576704, + 0.022199099883437157, + 0.02859976328909397, + 0.010127000510692596, + 0.021959960460662842, + 0.016151724383234978, + 0.011707688681781292, + 0.013590275309979916, + 0.017710022628307343, + 0.026065973564982414, + 0.031819917261600494, + 0.014086311683058739, + 0.04191918671131134, + 0.003436621278524399, + 0.003226309549063444, + 0.005228075198829174, + 0.011519019491970539, + 0.008457972668111324, + 0.011192001402378082, + 0.007968543097376823, + 0.009499984793365002, + 0.0074985446408391, + 0.0030316829215735197, + 0.004966563545167446, + 0.005817885510623455, + 0.00787927582859993, + 0.01372550055384636, + 0.005161798559129238, + 0.015067152678966522, + 0.007215770427137613, + 0.005091216415166855, + 0.012096577323973179, + 0.009026275016367435, + 0.012438731268048286, + 0.010257097892463207, + 0.008992944844067097, + 0.01305386982858181, + 0.013783147558569908, + 0.010793851688504219, + 0.0178777314722538, + 0.018712496384978294, + 0.029386315494775772, + 0.03970465436577797, + 0.014937052503228188, + 0.03419359028339386, + 0.004997133743017912, + 0.0038528465665876865, + 0.0052671851590275764, + 0.006464194506406784, + 0.012571852654218674, + 0.012442262843251228, + 0.006774933077394962, + 0.011348352767527103, + 0.01285307202488184, + 0.005305842962116003, + 0.0075109186582267284, + 0.009014611132442951, + 0.013381372205913067, + 0.020451268181204796, + 0.011083238758146763, + 0.031155727803707123, + 0.007609041407704353, + 0.004442374687641859, + 0.008005773648619652, + 0.008882589638233185, + 0.013117069378495216, + 0.01573188789188862, + 0.006714648101478815, + 0.01937796175479889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.029792189598083496 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot captures a scene rendered in a somber, painterly style characterized by thick, visible impasto and heavy, dark pigment. At the center of the painting, a gnarled, twisted tree root, like a small, emaciated human figure, is half-swallowed by an oppressive, inky shadow. A cold, directional light sweeps across its upper portion, highlighting the sheen of oil on its fibrous, bark-like texture and the intricate knots that suggest rough joints and a bowed head. The figure's lower half disappears into an impenetrable darkness, creating a dramatic effect of chiaroscuro. The root form is not static; it subtly contracts and oscillates with a slow, rhythmic pulse, its form viscerally trembling, as if breathing. This tiny, steady movement—a slight inward pull followed by a gentle sway—transforms the still life into a deeply unsettling and melancholic spectacle, a living sculpture that seems to breathe within the suffocating painted darkness.", + "seed": 20824, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_14/index_3262_seed_20824_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.042878275364637376, + "vlm_logprob": -0.042878275364637376, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.042878275364637376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03308369591832161 + ], + "opposite_logprob_tokens": [ + -3.78308367729187 + ], + "logprob_reward": -0.03308369591832161, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03308369591832161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2826085388660431, + 0.06441711634397507, + 0.08551070094108582, + 0.11365529894828796, + 0.028092799708247185, + 0.09363897144794464, + 0.0903845950961113, + 0.04864848777651787, + 0.06721515953540802, + 0.1258283257484436 + ], + "credit_map": [ + 0.032899074256420135, + 0.014903347007930279, + 0.0999065637588501, + 0.021383902058005333, + 0.020196767523884773, + 0.04012587293982506, + 0.028331704437732697, + 0.02486129105091095, + 0.005075389053672552, + 0.00695708068087697, + 0.013449648395180702, + 0.0076093426905572414, + 0.008066210895776749, + 0.008662577718496323, + 0.009635633789002895, + 0.004961230326443911, + 0.00771694490686059, + 0.008733021095395088, + 0.01643499545753002, + 0.010621055029332638, + 0.012673771008849144, + 0.011364184319972992, + 0.012101455591619015, + 0.005865276325494051, + 0.011527888476848602, + 0.008079349994659424, + 0.019675543531775475, + 0.017142362892627716, + 0.010768976993858814, + 0.016072573140263557, + 0.01802312396466732, + 0.012365476228296757, + 0.003938276786357164, + 0.002588793635368347, + 0.0058999136090278625, + 0.0030418357346206903, + 0.0035644513554871082, + 0.0032965606078505516, + 0.003221451072022319, + 0.002541515976190567, + 0.010553655214607716, + 0.004634146112948656, + 0.028049994260072708, + 0.00842691957950592, + 0.00665371585637331, + 0.014527068473398685, + 0.011536103673279285, + 0.009257368743419647, + 0.009093796834349632, + 0.004510688595473766, + 0.028721395879983902, + 0.0100983502343297, + 0.007509151939302683, + 0.01357128657400608, + 0.009796561673283577, + 0.007083356846123934, + 0.005223706364631653, + 0.002768842503428459, + 0.009389467537403107, + 0.00523670157417655, + 0.005105612799525261, + 0.006420888006687164, + 0.009759112261235714, + 0.004744154401123524, + 0.008728071115911007, + 0.004109594970941544, + 0.015158295631408691, + 0.006376242730766535, + 0.007830891758203506, + 0.008793340064585209, + 0.009414452128112316, + 0.006804266944527626, + 0.013489410281181335, + 0.008389933966100216, + 0.029929736629128456, + 0.011219379492104053, + 0.010428914800286293, + 0.023768646642565727, + 0.014396771788597107, + 0.014205520041286945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.03308369591832161 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296918988227844 + ], + "opposite_logprob_tokens": [ + -2.237969160079956 + ], + "logprob_reward": -0.11296918988227844, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296918988227844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2490486204624176, + 0.14281629025936127, + 0.10317563265562057, + 0.05249666050076485, + 0.06779260188341141, + 0.07700701057910919, + 0.06647462397813797, + 0.09414087235927582, + 0.08703409135341644, + 0.06001356989145279 + ], + "credit_map": [ + 0.03638352453708649, + 0.01890057511627674, + 0.03965936228632927, + 0.02643313631415367, + 0.02756662108004093, + 0.036519017070531845, + 0.037515778094530106, + 0.026070620864629745, + 0.015785617753863335, + 0.010869995690882206, + 0.018151747062802315, + 0.013644807040691376, + 0.016538990661501884, + 0.030203871428966522, + 0.017115434631705284, + 0.020505841821432114, + 0.010455464944243431, + 0.007315303198993206, + 0.015408157370984554, + 0.012014375999569893, + 0.013141386210918427, + 0.019561775028705597, + 0.013255256228148937, + 0.012023922987282276, + 0.005068450700491667, + 0.0039129056967794895, + 0.013022244907915592, + 0.006368646863847971, + 0.006040399894118309, + 0.007242364808917046, + 0.006417504046112299, + 0.004424147307872772, + 0.00958285853266716, + 0.005013741087168455, + 0.012416298501193523, + 0.007112607825547457, + 0.00867049302905798, + 0.010330190882086754, + 0.007564531173557043, + 0.007101885508745909, + 0.010006317868828773, + 0.0027726010885089636, + 0.026501523330807686, + 0.005134872626513243, + 0.0045128390192985535, + 0.011588509194552898, + 0.009734627790749073, + 0.00675572594627738, + 0.008147470653057098, + 0.005073284264653921, + 0.013031400740146637, + 0.006070733070373535, + 0.006062311585992575, + 0.013304735533893108, + 0.006115313619375229, + 0.008669381961226463, + 0.010436215437948704, + 0.0030616107396781445, + 0.03610757738351822, + 0.006156256888061762, + 0.004548529163002968, + 0.01550478208810091, + 0.008737497963011265, + 0.009588408283889294, + 0.010915866121649742, + 0.00349796237424016, + 0.026151668280363083, + 0.005371193867176771, + 0.004559464752674103, + 0.016428496688604355, + 0.009897533804178238, + 0.010211917571723461, + 0.010016284883022308, + 0.0030374673660844564, + 0.015677940100431442, + 0.004403917118906975, + 0.0059958104975521564, + 0.008133722469210625, + 0.007033270318061113, + 0.005715161096304655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.11296918988227844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206700026988983 + ], + "opposite_logprob_tokens": [ + -4.264206886291504 + ], + "logprob_reward": -0.014206700026988983, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206700026988983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3220715820789337, + 0.13579195737838745, + 0.11269056051969528, + 0.07073559612035751, + 0.03670450672507286, + 0.08863426744937897, + 0.09495995938777924, + 0.0557587705552578, + 0.02687784470617771, + 0.05577489361166954 + ], + "credit_map": [ + 0.029591728001832962, + 0.029392974451184273, + 0.04137180373072624, + 0.037402134388685226, + 0.0444340743124485, + 0.0460469126701355, + 0.05829613283276558, + 0.035535845905542374, + 0.010006560944020748, + 0.013308478519320488, + 0.015173145569860935, + 0.017345476895570755, + 0.019682960584759712, + 0.026890233159065247, + 0.01771647296845913, + 0.015668638050556183, + 0.01011164765805006, + 0.01137588918209076, + 0.017445603385567665, + 0.013036180287599564, + 0.016073783859610558, + 0.016496850177645683, + 0.016364727169275284, + 0.011785889975726604, + 0.004305769689381123, + 0.0050033023580908775, + 0.014330482110381126, + 0.0080097746104002, + 0.008617915213108063, + 0.013446256518363953, + 0.009217297658324242, + 0.0078048063442111015, + 0.0037366405595093966, + 0.002995590213686228, + 0.005555362440645695, + 0.004333071410655975, + 0.005241040606051683, + 0.005797331687062979, + 0.005027268547564745, + 0.004018204752355814, + 0.010270765982568264, + 0.004128476604819298, + 0.02778559923171997, + 0.007107987999916077, + 0.007054296787828207, + 0.01290837861597538, + 0.01165861263871193, + 0.007720160763710737, + 0.009115015156567097, + 0.008448053151369095, + 0.010086100548505783, + 0.009478716179728508, + 0.011103419587016106, + 0.0193697027862072, + 0.013861823827028275, + 0.01349713932722807, + 0.006299667991697788, + 0.0031099668703973293, + 0.014036661013960838, + 0.0045570977963507175, + 0.00390347046777606, + 0.011064590886235237, + 0.005832008086144924, + 0.006955310236662626, + 0.0033395388163626194, + 0.0021083897445350885, + 0.004307421389967203, + 0.00344844744540751, + 0.0034035281278192997, + 0.00411947350949049, + 0.0036552755627781153, + 0.0024957722052931786, + 0.0068497308529913425, + 0.004139828495681286, + 0.011107089929282665, + 0.005067009944468737, + 0.005232789553701878, + 0.010555467568337917, + 0.006123372353613377, + 0.006699613761156797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014206700026988983 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076156049966812 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011076156049966812, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076156049966812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16345171630382538, + 0.2415798008441925, + 0.23273006081581116, + 0.08183866739273071, + 0.058187928050756454, + 0.04195196554064751, + 0.053940050303936005, + 0.04621487855911255, + 0.044946517795324326, + 0.035158444195985794 + ], + "credit_map": [ + 0.020106540992856026, + 0.013638502918183804, + 0.030854299664497375, + 0.018846409395337105, + 0.020587056875228882, + 0.024057922884821892, + 0.018163928762078285, + 0.01719706691801548, + 0.03135359287261963, + 0.0173527542501688, + 0.04890401288866997, + 0.02644447609782219, + 0.026732200756669044, + 0.03995683044195175, + 0.025724932551383972, + 0.02511100098490715, + 0.02606181427836418, + 0.015616510063409805, + 0.05338320881128311, + 0.027200140058994293, + 0.025831853970885277, + 0.037038493901491165, + 0.025711486116051674, + 0.02188655361533165, + 0.007764487527310848, + 0.0052911266684532166, + 0.019401663914322853, + 0.011294011026620865, + 0.008410792797803879, + 0.011620067059993744, + 0.012032225728034973, + 0.0060242945328354836, + 0.008659454993903637, + 0.005075768101960421, + 0.010903280228376389, + 0.007099109701812267, + 0.006937755271792412, + 0.006600069813430309, + 0.00824683252722025, + 0.004665655549615622, + 0.006026189774274826, + 0.0031634110491722822, + 0.009243057109415531, + 0.004439521115273237, + 0.004218480549752712, + 0.006535394582897425, + 0.004030308686196804, + 0.004295602906495333, + 0.007442089729011059, + 0.003510442329570651, + 0.0069046118296682835, + 0.008332652039825916, + 0.0052291047759354115, + 0.00830861460417509, + 0.00806008093059063, + 0.006152451504021883, + 0.006184614263474941, + 0.0029170375782996416, + 0.007667116355150938, + 0.004483328666538, + 0.0041281115263700485, + 0.00854426808655262, + 0.006343310233205557, + 0.00594708975404501, + 0.003816310316324234, + 0.002621373161673546, + 0.011089487932622433, + 0.004832753911614418, + 0.004431047476828098, + 0.008144094608724117, + 0.0050385938957333565, + 0.004972857888787985, + 0.003895736765116453, + 0.0029360547196120024, + 0.006598784122616053, + 0.0036300106439739466, + 0.00452444888651371, + 0.0046735843643546104, + 0.006011371500790119, + 0.0028884524945169687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.011076156049966812 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305563494563103 + ], + "opposite_logprob_tokens": [ + -3.168055534362793 + ], + "logprob_reward": -0.04305563494563103, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305563494563103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14705103635787964, + 0.11286963522434235, + 0.09299679100513458, + 0.1470961719751358, + 0.12215825170278549, + 0.0975365936756134, + 0.08892256766557693, + 0.09201109409332275, + 0.051090165972709656, + 0.048267707228660583 + ], + "credit_map": [ + 0.0203485619276762, + 0.008249263279139996, + 0.038636501878499985, + 0.012622135691344738, + 0.015360556542873383, + 0.02115805074572563, + 0.016432398930191994, + 0.01424356084316969, + 0.012673215940594673, + 0.006732996553182602, + 0.028414186090230942, + 0.01094820536673069, + 0.011452714912593365, + 0.01747426763176918, + 0.012953324243426323, + 0.012220715172588825, + 0.009672939777374268, + 0.007552039343863726, + 0.02039288729429245, + 0.01009781938046217, + 0.011880400590598583, + 0.012776453047990799, + 0.012907135300338268, + 0.00771710928529501, + 0.013206306844949722, + 0.008361872285604477, + 0.03605257347226143, + 0.021678604185581207, + 0.015091806650161743, + 0.022951751947402954, + 0.01845410279929638, + 0.011299132369458675, + 0.010252058506011963, + 0.00833116751164198, + 0.0305967777967453, + 0.01224641315639019, + 0.013874310068786144, + 0.020212693139910698, + 0.01496199332177639, + 0.011682827956974506, + 0.008401099592447281, + 0.009277844801545143, + 0.017300715669989586, + 0.01351896207779646, + 0.012938346713781357, + 0.01153414323925972, + 0.01474372111260891, + 0.009821760468184948, + 0.012402723543345928, + 0.006685197353363037, + 0.012951450422406197, + 0.010396460071206093, + 0.010545995086431503, + 0.014374648220837116, + 0.010494817048311234, + 0.011071271263062954, + 0.010478176176548004, + 0.004576842300593853, + 0.028201423585414886, + 0.00832031387835741, + 0.007679361384361982, + 0.01580548658967018, + 0.007307899184525013, + 0.00964159145951271, + 0.006274102721363306, + 0.002946582855656743, + 0.013825582340359688, + 0.005937650334089994, + 0.004922090098261833, + 0.007462614215910435, + 0.005076518282294273, + 0.00464502302929759, + 0.005872970912605524, + 0.0029098696541041136, + 0.012797393836081028, + 0.0038640096317976713, + 0.00458819605410099, + 0.007570117246359587, + 0.006331481039524078, + 0.00433367071673274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.04305563494563103 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot captures a scene rendered in a somber, painterly style characterized by thick, visible impasto and heavy, dark pigment. At the center of the painting, a gnarled, twisted tree root, like a small, emaciated human figure, is half-swallowed by an oppressive, inky shadow. A cold, directional light sweeps across its upper portion, highlighting the sheen of oil on its fibrous, bark-like texture and the intricate knots that suggest rough joints and a bowed head. The figure's lower half disappears into an impenetrable darkness, creating a dramatic effect of chiaroscuro. The root form is not static; it subtly contracts and oscillates with a slow, rhythmic pulse, its form viscerally trembling, as if breathing. This tiny, steady movement—a slight inward pull followed by a gentle sway—transforms the still life into a deeply unsettling and melancholic spectacle, a living sculpture that seems to breathe within the suffocating painted darkness.", + "seed": 20825, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_14/index_3262_seed_20825_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03572074170224369, + "vlm_logprob": -0.03572074170224369, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.03572074170224369, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09751130640506744 + ], + "opposite_logprob_tokens": [ + -2.4725112915039062 + ], + "logprob_reward": -0.09751130640506744, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09751130640506744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12966187298297882, + 0.09502309560775757, + 0.07502161711454391, + 0.08789059519767761, + 0.29674282670021057, + 0.0675840675830841, + 0.042530667036771774, + 0.055369794368743896, + 0.0721912607550621, + 0.07798424363136292 + ], + "credit_map": [ + 0.011198426596820354, + 0.014365289360284805, + 0.018946101889014244, + 0.012952497228980064, + 0.026508359238505363, + 0.014783166348934174, + 0.01974296197295189, + 0.01116506103426218, + 0.008662523701786995, + 0.012108489871025085, + 0.012799656018614769, + 0.01259838417172432, + 0.015948930755257607, + 0.01044468954205513, + 0.014641419053077698, + 0.007818999700248241, + 0.006045765243470669, + 0.008333412930369377, + 0.012248220853507519, + 0.0074694654904305935, + 0.01138447318226099, + 0.007091568782925606, + 0.013172741048038006, + 0.009275970049202442, + 0.008596999570727348, + 0.011591662652790546, + 0.013587098568677902, + 0.01221813727170229, + 0.012373194098472595, + 0.007346154656261206, + 0.013888498768210411, + 0.008288845419883728, + 0.02204078435897827, + 0.033285293728113174, + 0.049917418509721756, + 0.038369644433259964, + 0.03693855181336403, + 0.02906184270977974, + 0.05059448257088661, + 0.03653479740023613, + 0.007205730304121971, + 0.00756260147318244, + 0.009463159367442131, + 0.006618219893425703, + 0.008318719454109669, + 0.010228224098682404, + 0.010232829488813877, + 0.007954584434628487, + 0.0037888854276388884, + 0.005623441189527512, + 0.006633105222135782, + 0.004142619203776121, + 0.005898577626794577, + 0.005997738800942898, + 0.0057918536476790905, + 0.004654444754123688, + 0.004314820747822523, + 0.00609853770583868, + 0.009867285378277302, + 0.007719563320279121, + 0.00903958547860384, + 0.005642348900437355, + 0.0058735390193760395, + 0.0068141138181090355, + 0.005683272611349821, + 0.006757176481187344, + 0.01652001589536667, + 0.011009171605110168, + 0.008531337603926659, + 0.006223292555660009, + 0.006947462446987629, + 0.010519525036215782, + 0.006708901841193438, + 0.009857863187789917, + 0.015895692631602287, + 0.007504517678171396, + 0.011323071084916592, + 0.006912441458553076, + 0.008770580403506756, + 0.011011174879968166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.09751130640506744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033730026334524155 + ], + "opposite_logprob_tokens": [ + -3.4087300300598145 + ], + "logprob_reward": -0.033730026334524155, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033730026334524155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3035738468170166, + 0.15105579793453217, + 0.08759462088346481, + 0.06823999434709549, + 0.08805584907531738, + 0.06468918919563293, + 0.06548751145601273, + 0.039911702275276184, + 0.04389774799346924, + 0.08749375492334366 + ], + "credit_map": [ + 0.019031953066587448, + 0.03499376028776169, + 0.049654509872198105, + 0.03494558110833168, + 0.04833930358290672, + 0.033263470977544785, + 0.053116749972105026, + 0.0302285049110651, + 0.01117841899394989, + 0.016725871711969376, + 0.02657546103000641, + 0.020774099975824356, + 0.021663516759872437, + 0.015951327979564667, + 0.020648662000894547, + 0.017538443207740784, + 0.00596311641857028, + 0.010234905406832695, + 0.014357853680849075, + 0.01067289151251316, + 0.014250366017222404, + 0.009766989387571812, + 0.01170339249074459, + 0.010645106434822083, + 0.0050833867862820625, + 0.008914189413189888, + 0.008846988901495934, + 0.008648911491036415, + 0.01225336641073227, + 0.007921621203422546, + 0.010575949214398861, + 0.005995577666908503, + 0.006928845774382353, + 0.011521494016051292, + 0.012131892144680023, + 0.013438425958156586, + 0.01376073993742466, + 0.009867853485047817, + 0.01281092781573534, + 0.007595671806484461, + 0.006720652803778648, + 0.007216761354357004, + 0.011103151366114616, + 0.0076384046114981174, + 0.008463037200272083, + 0.00714368000626564, + 0.008589707314968109, + 0.00781379546970129, + 0.00862354040145874, + 0.006859733257442713, + 0.010968024842441082, + 0.008733035065233707, + 0.007785053923726082, + 0.005904154386371374, + 0.008455537259578705, + 0.008158428594470024, + 0.0035836840979754925, + 0.004436099901795387, + 0.006932035554200411, + 0.004409332759678364, + 0.00490647042170167, + 0.004241630434989929, + 0.00482872175052762, + 0.006573728285729885, + 0.0031150311697274446, + 0.0051467991434037685, + 0.007471808232367039, + 0.005582510959357023, + 0.006548509933054447, + 0.004250827245414257, + 0.006761356256902218, + 0.005020902492105961, + 0.010742573998868465, + 0.013115169480443, + 0.012543556280434132, + 0.007487800903618336, + 0.011372468434274197, + 0.011658997274935246, + 0.01026956457644701, + 0.010303625836968422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.033730026334524155 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004406196530908346 + ], + "opposite_logprob_tokens": [ + -7.75044059753418 + ], + "logprob_reward": -0.0004406196530908346, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004406196530908346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22022369503974915, + 0.0632646307349205, + 0.05379496142268181, + 0.07007744163274765, + 0.21091879904270172, + 0.1943935751914978, + 0.056456755846738815, + 0.0451466403901577, + 0.04625020921230316, + 0.03947330266237259 + ], + "credit_map": [ + 0.01806926541030407, + 0.025920698419213295, + 0.03946230560541153, + 0.03126807138323784, + 0.02826584130525589, + 0.019192853942513466, + 0.03541465476155281, + 0.022630000486969948, + 0.007042644079774618, + 0.00609760545194149, + 0.010984216816723347, + 0.006710007321089506, + 0.007991808466613293, + 0.006497517228126526, + 0.010522614233195782, + 0.007418219931423664, + 0.0043855831027030945, + 0.006291279569268227, + 0.010161101818084717, + 0.006046624388545752, + 0.005909338593482971, + 0.004718671087175608, + 0.009960350580513477, + 0.006322014145553112, + 0.006099430378526449, + 0.00884546060115099, + 0.011125832796096802, + 0.010290742851793766, + 0.009312896057963371, + 0.007541228551417589, + 0.010417230427265167, + 0.00644462276250124, + 0.017126288264989853, + 0.03016667440533638, + 0.029163241386413574, + 0.027821367606520653, + 0.03307675942778587, + 0.02004729025065899, + 0.0355665348470211, + 0.01795063354074955, + 0.015115337446331978, + 0.025258971378207207, + 0.03010295331478119, + 0.02552679181098938, + 0.027077415958046913, + 0.01767645962536335, + 0.032624803483486176, + 0.021010834723711014, + 0.0046379524283111095, + 0.007161400280892849, + 0.007959334179759026, + 0.0064592366106808186, + 0.0098141273483634, + 0.007103875279426575, + 0.00784771703183651, + 0.005473110359162092, + 0.00404756935313344, + 0.005574190989136696, + 0.006201721727848053, + 0.005492068361490965, + 0.007953206077218056, + 0.005869382061064243, + 0.005009884014725685, + 0.004998620133846998, + 0.003678535344079137, + 0.007134817540645599, + 0.006813282147049904, + 0.0056915138848125935, + 0.007284980267286301, + 0.003643856616690755, + 0.007185544818639755, + 0.004817679524421692, + 0.003920403774827719, + 0.00538094574585557, + 0.005123585928231478, + 0.005788316950201988, + 0.005919336806982756, + 0.004182822071015835, + 0.005567567422986031, + 0.003590327687561512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0004406196530908346 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058020606637001 + ], + "opposite_logprob_tokens": [ + -3.895580291748047 + ], + "logprob_reward": -0.02058020606637001, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058020606637001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20474599301815033, + 0.09834174811840057, + 0.07385537773370743, + 0.10412833839654922, + 0.104650117456913, + 0.06472166627645493, + 0.05673033744096756, + 0.0543486550450325, + 0.04933897778391838, + 0.18913882970809937 + ], + "credit_map": [ + 0.01622307114303112, + 0.025037867948412895, + 0.02906130440533161, + 0.025398598983883858, + 0.02767624892294407, + 0.02057069167494774, + 0.036600593477487564, + 0.02417762391269207, + 0.010223131626844406, + 0.011793626472353935, + 0.01433074101805687, + 0.01190732978284359, + 0.013660264201462269, + 0.010888970457017422, + 0.015101267956197262, + 0.01043640822172165, + 0.006777327042073011, + 0.00607365183532238, + 0.014794404618442059, + 0.0073109096847474575, + 0.00916146207600832, + 0.008080900646746159, + 0.009401940740644932, + 0.012254776433110237, + 0.011907647363841534, + 0.011337396688759327, + 0.015561336651444435, + 0.012563816271722317, + 0.012279731221497059, + 0.015123103745281696, + 0.012923009693622589, + 0.01243229117244482, + 0.008307560347020626, + 0.013540889136493206, + 0.014499383047223091, + 0.01521273422986269, + 0.014618109911680222, + 0.013724427670240402, + 0.01437684241682291, + 0.01037017535418272, + 0.005347255151718855, + 0.006558180786669254, + 0.009447863325476646, + 0.008447918109595776, + 0.010232904925942421, + 0.0073332469910383224, + 0.009888780303299427, + 0.007465512957423925, + 0.006722500082105398, + 0.006154599599540234, + 0.010850745253264904, + 0.006779934279620647, + 0.00596016738563776, + 0.0050341407768428326, + 0.007072627078741789, + 0.008155622519552708, + 0.006068486254662275, + 0.006390681955963373, + 0.008109880611300468, + 0.0062184506095945835, + 0.0051943156868219376, + 0.008311731740832329, + 0.006269919220358133, + 0.007785187568515539, + 0.004120348021388054, + 0.005682643037289381, + 0.007352293003350496, + 0.005338623188436031, + 0.008588068187236786, + 0.005528151523321867, + 0.006385969929397106, + 0.0063428799621760845, + 0.015299888327717781, + 0.024490276351571083, + 0.02389857918024063, + 0.026844125241041183, + 0.020912913605570793, + 0.023862255737185478, + 0.030368909239768982, + 0.023461896926164627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.02058020606637001 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341550052165985 + ], + "opposite_logprob_tokens": [ + -3.651341438293457 + ], + "logprob_reward": -0.026341550052165985, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341550052165985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17273513972759247, + 0.08249415457248688, + 0.09312688559293747, + 0.044384073466062546, + 0.11600453406572342, + 0.10934186726808548, + 0.061716556549072266, + 0.12248183786869049, + 0.13656863570213318, + 0.061146292835474014 + ], + "credit_map": [ + 0.019723132252693176, + 0.016135452315211296, + 0.025944054126739502, + 0.01993715949356556, + 0.021799221634864807, + 0.019373465329408646, + 0.032540131360292435, + 0.01728251576423645, + 0.007806341163814068, + 0.009120944887399673, + 0.012883400544524193, + 0.00947419460862875, + 0.012875931337475777, + 0.008526211604475975, + 0.01200309582054615, + 0.009804039262235165, + 0.007306545041501522, + 0.012322584167122841, + 0.014253783039748669, + 0.010207419283688068, + 0.013942204415798187, + 0.008280792273581028, + 0.014483862556517124, + 0.012329699471592903, + 0.004486579913645983, + 0.005103497300297022, + 0.006795749068260193, + 0.005176158621907234, + 0.0077067487873137, + 0.004667994566261768, + 0.0064351740293204784, + 0.004012176766991615, + 0.009718452580273151, + 0.014199843630194664, + 0.018485430628061295, + 0.01346749346703291, + 0.016621094197034836, + 0.01387883722782135, + 0.013699854724109173, + 0.015933537855744362, + 0.008891569450497627, + 0.01110618282109499, + 0.017723767086863518, + 0.013481191359460354, + 0.013669390231370926, + 0.012232176959514618, + 0.01773044466972351, + 0.014507153071463108, + 0.006443793419748545, + 0.0074609313160181046, + 0.010984922759234905, + 0.007525233086198568, + 0.009286003187298775, + 0.006260768510401249, + 0.006670494098216295, + 0.007084416225552559, + 0.009106487967073917, + 0.01126440055668354, + 0.022792214527726173, + 0.015192556194961071, + 0.01902955025434494, + 0.01379839051514864, + 0.014303677715361118, + 0.016994565725326538, + 0.010927768424153328, + 0.015314151532948017, + 0.022919373586773872, + 0.013749105855822563, + 0.023160047829151154, + 0.014868037775158882, + 0.020531127229332924, + 0.01509903371334076, + 0.006354367360472679, + 0.008348704315721989, + 0.007255499716848135, + 0.010161896236240864, + 0.005944231525063515, + 0.00733993062749505, + 0.00982208177447319, + 0.00591958686709404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.026341550052165985 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot captures a scene rendered in a somber, painterly style characterized by thick, visible impasto and heavy, dark pigment. At the center of the painting, a gnarled, twisted tree root, like a small, emaciated human figure, is half-swallowed by an oppressive, inky shadow. A cold, directional light sweeps across its upper portion, highlighting the sheen of oil on its fibrous, bark-like texture and the intricate knots that suggest rough joints and a bowed head. The figure's lower half disappears into an impenetrable darkness, creating a dramatic effect of chiaroscuro. The root form is not static; it subtly contracts and oscillates with a slow, rhythmic pulse, its form viscerally trembling, as if breathing. This tiny, steady movement—a slight inward pull followed by a gentle sway—transforms the still life into a deeply unsettling and melancholic spectacle, a living sculpture that seems to breathe within the suffocating painted darkness.", + "seed": 20826, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_14/index_3262_seed_20826_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08040541873779147, + "vlm_logprob": -0.08040541873779147, + "vlm_yes_no_margin": 3.200000023841858, + "vlm_token_credit": -0.08040541873779147, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09565606713294983 + ], + "opposite_logprob_tokens": [ + -2.470656156539917 + ], + "logprob_reward": -0.09565606713294983, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09565606713294983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18882402777671814, + 0.06592047959566116, + 0.05850709229707718, + 0.06181929633021355, + 0.08116620779037476, + 0.09603141993284225, + 0.12192681431770325, + 0.1625952124595642, + 0.07623744755983353, + 0.08697204291820526 + ], + "credit_map": [ + 0.027055980637669563, + 0.0201611015945673, + 0.02099468745291233, + 0.020125089213252068, + 0.018631717190146446, + 0.030229683965444565, + 0.016620153561234474, + 0.035005610436201096, + 0.008050727657973766, + 0.005311785731464624, + 0.011072960682213306, + 0.006355448625981808, + 0.007258243393152952, + 0.010131717659533024, + 0.005314147565513849, + 0.012425447814166546, + 0.007906810380518436, + 0.006900581996887922, + 0.00760357640683651, + 0.007065407931804657, + 0.0053517455235123634, + 0.009654075838625431, + 0.004894477780908346, + 0.009130416437983513, + 0.005508530419319868, + 0.003271239111199975, + 0.005349517799913883, + 0.003990648314356804, + 0.004775621462613344, + 0.019564388319849968, + 0.004557501524686813, + 0.01480184681713581, + 0.009776047430932522, + 0.00767949502915144, + 0.009021361358463764, + 0.007582321297377348, + 0.013502025976777077, + 0.01377307903021574, + 0.008811108767986298, + 0.011020765639841557, + 0.011391904205083847, + 0.010306176729500294, + 0.012068428099155426, + 0.012928485870361328, + 0.01200317032635212, + 0.013510462827980518, + 0.009340438060462475, + 0.014482361264526844, + 0.016356151551008224, + 0.00791226141154766, + 0.012035086750984192, + 0.00642768107354641, + 0.008995241485536098, + 0.031665876507759094, + 0.0063869766891002655, + 0.03214755654335022, + 0.021309249103069305, + 0.008669606409966946, + 0.016268840059638023, + 0.008547641336917877, + 0.009288052096962929, + 0.048017580062150955, + 0.010507536120712757, + 0.039986707270145416, + 0.012592156417667866, + 0.006502695847302675, + 0.010022721253335476, + 0.0068595875054597855, + 0.004966660402715206, + 0.012937785126268864, + 0.00545199541375041, + 0.016903843730688095, + 0.007871875539422035, + 0.007841524668037891, + 0.010136909782886505, + 0.008186819031834602, + 0.006147542968392372, + 0.02338159643113613, + 0.006466163322329521, + 0.016939613968133926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.09565606713294983 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034558176994324 + ], + "opposite_logprob_tokens": [ + -2.3503456115722656 + ], + "logprob_reward": -0.10034558176994324, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034558176994324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30678504705429077, + 0.15921960771083832, + 0.06251831352710724, + 0.0810050442814827, + 0.08300691097974777, + 0.03854810446500778, + 0.0539393275976181, + 0.031211741268634796, + 0.12813298404216766, + 0.05563290789723396 + ], + "credit_map": [ + 0.04950461909174919, + 0.026245133951306343, + 0.04596288129687309, + 0.02809320203959942, + 0.025229940190911293, + 0.05424315482378006, + 0.02385428547859192, + 0.0536518320441246, + 0.01909862458705902, + 0.016310378909111023, + 0.020662305876612663, + 0.015488868579268456, + 0.015133805572986603, + 0.0355074517428875, + 0.011112104170024395, + 0.025906069204211235, + 0.008420106023550034, + 0.005324751138687134, + 0.008076067082583904, + 0.00527619943022728, + 0.006694879848510027, + 0.013712911866605282, + 0.005276782903820276, + 0.009736616164445877, + 0.008595429360866547, + 0.006028852891176939, + 0.009168165735900402, + 0.008112356998026371, + 0.00919879786670208, + 0.014652455225586891, + 0.011539214290678501, + 0.013709770515561104, + 0.011391593143343925, + 0.006753724068403244, + 0.011701606214046478, + 0.010139363817870617, + 0.007934930734336376, + 0.011145398020744324, + 0.013128470629453659, + 0.010811827145516872, + 0.004271687939763069, + 0.004368214402347803, + 0.005038601811975241, + 0.004550259560346603, + 0.0043608457781374454, + 0.006414856296032667, + 0.0038168386090546846, + 0.0057267979718744755, + 0.006460176780819893, + 0.0032535542268306017, + 0.004687584936618805, + 0.0033327729906886816, + 0.003110397607088089, + 0.016131486743688583, + 0.002720274031162262, + 0.014243080280721188, + 0.005122998263686895, + 0.00302443141117692, + 0.00397511525079608, + 0.0036824322305619717, + 0.002860972424969077, + 0.005065473727881908, + 0.0029263056349009275, + 0.004554012790322304, + 0.019701242446899414, + 0.00841548852622509, + 0.01081634871661663, + 0.009194701910018921, + 0.006188498809933662, + 0.03769442066550255, + 0.006041694898158312, + 0.03008059598505497, + 0.0058314916677773, + 0.003794001415371895, + 0.0043571870774030685, + 0.003427824005484581, + 0.003279689699411392, + 0.018467091023921967, + 0.002822671551257372, + 0.013652952387928963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.10034558176994324 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011857867939397693 + ], + "opposite_logprob_tokens": [ + -6.751185894012451 + ], + "logprob_reward": -0.0011857867939397693, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011857867939397693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22789542376995087, + 0.16883446276187897, + 0.1191190853714943, + 0.10155323147773743, + 0.048865076154470444, + 0.07289339601993561, + 0.07682795077562332, + 0.05083697661757469, + 0.07733774930238724, + 0.05583663657307625 + ], + "credit_map": [ + 0.02582239732146263, + 0.022594530135393143, + 0.02542612887918949, + 0.023448368534445763, + 0.022766167297959328, + 0.048865363001823425, + 0.01804075948894024, + 0.040931716561317444, + 0.01608903706073761, + 0.012727359309792519, + 0.014698858372867107, + 0.012463376857340336, + 0.014454071410000324, + 0.055349163711071014, + 0.010734000243246555, + 0.0323186069726944, + 0.011023678816854954, + 0.007862886413931847, + 0.00986011978238821, + 0.008653960190713406, + 0.009457537904381752, + 0.035777248442173004, + 0.01225916389375925, + 0.024224484339356422, + 0.01137586124241352, + 0.006512352731078863, + 0.009282113052904606, + 0.007238862570375204, + 0.00799153745174408, + 0.031365785747766495, + 0.0054381354711949825, + 0.022348573431372643, + 0.005849979817867279, + 0.003467794042080641, + 0.005490290466696024, + 0.0048331208527088165, + 0.004557918757200241, + 0.01065953727811575, + 0.004854355473071337, + 0.009152079001069069, + 0.007146782707422972, + 0.005913941189646721, + 0.008516097441315651, + 0.007633475121110678, + 0.006956844590604305, + 0.015314729884266853, + 0.009436246939003468, + 0.01197528000921011, + 0.008588036522269249, + 0.005110518541187048, + 0.007324363570660353, + 0.004391662776470184, + 0.004979296587407589, + 0.01966588944196701, + 0.006264680530875921, + 0.02050350233912468, + 0.007002910599112511, + 0.003488963469862938, + 0.006213741842657328, + 0.0033723236992955208, + 0.0034765370655804873, + 0.012075082398951054, + 0.0033850623294711113, + 0.011822354979813099, + 0.010574489831924438, + 0.004551686346530914, + 0.007622994016855955, + 0.0041206213645637035, + 0.004237369634211063, + 0.02163589373230934, + 0.004158869385719299, + 0.02043583057820797, + 0.0067639434710145, + 0.005818455014377832, + 0.005646978504955769, + 0.006107877939939499, + 0.006434457376599312, + 0.009511944837868214, + 0.005854668095707893, + 0.009698311798274517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0011857867939397693 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279787242412567 + ], + "opposite_logprob_tokens": [ + -2.0177979469299316 + ], + "logprob_reward": -0.14279787242412567, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14279787242412567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2765345275402069, + 0.09065847843885422, + 0.06779114156961441, + 0.15592604875564575, + 0.15801095962524414, + 0.07516267150640488, + 0.04161818325519562, + 0.02514761872589588, + 0.021031254902482033, + 0.08811917901039124 + ], + "credit_map": [ + 0.03589963540434837, + 0.018723446875810623, + 0.024189257994294167, + 0.018717942759394646, + 0.01687461882829666, + 0.08363493531942368, + 0.013459348119795322, + 0.06503532826900482, + 0.010192610323429108, + 0.007721625734120607, + 0.008409052155911922, + 0.006477107759565115, + 0.006980938836932182, + 0.026726074516773224, + 0.0050599053502082825, + 0.019091162830591202, + 0.008627749979496002, + 0.005983763374388218, + 0.010806413367390633, + 0.006446425803005695, + 0.006873969919979572, + 0.01122349314391613, + 0.006138702388852835, + 0.011690620332956314, + 0.023425979539752007, + 0.006949012633413076, + 0.014044109731912613, + 0.007575927302241325, + 0.00800319667905569, + 0.04997514188289642, + 0.0053926329128444195, + 0.040560029447078705, + 0.01821315661072731, + 0.008893152698874474, + 0.02160794474184513, + 0.01123050320893526, + 0.010747904889285564, + 0.048782069236040115, + 0.006044704932719469, + 0.03249150887131691, + 0.008147643879055977, + 0.004771269392222166, + 0.008121080696582794, + 0.005472853314131498, + 0.005381413735449314, + 0.02375083602964878, + 0.004821522627025843, + 0.014696048572659492, + 0.0067240954376757145, + 0.0028106733225286007, + 0.004416509531438351, + 0.002440408803522587, + 0.003141432534903288, + 0.0091642364859581, + 0.002550221746787429, + 0.010370601899921894, + 0.0033703246153891087, + 0.0019370673689991236, + 0.0033152129035443068, + 0.002613408723846078, + 0.0023268007207661867, + 0.0038342317566275597, + 0.002969471039250493, + 0.004781098570674658, + 0.002485652221366763, + 0.0022036638110876083, + 0.0024632876738905907, + 0.002646056702360511, + 0.002444859826937318, + 0.002920755883678794, + 0.0019141922239214182, + 0.003952784929424524, + 0.009341122582554817, + 0.004162243101745844, + 0.007199724670499563, + 0.0051666609942913055, + 0.0042741489596664906, + 0.02946525253355503, + 0.005215908866375685, + 0.0232941173017025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.14279787242412567 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style painterly with thick, visible impasto and heavy dark pigment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject a gnarled, twisted tree root resembling an emaciated human figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a cold, directional light sweeping across the subject to highlight oil sheen and bark texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the root form exhibit a slow, rhythmic pulse with subtle inward pulling and swaying motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062041785567998886 + ], + "opposite_logprob_tokens": [ + -2.812041759490967 + ], + "logprob_reward": -0.062041785567998886, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062041785567998886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29058343172073364, + 0.11561550945043564, + 0.0949912741780281, + 0.04728538542985916, + 0.103647880256176, + 0.04761992394924164, + 0.03181901201605797, + 0.14608387649059296, + 0.05968143418431282, + 0.06267228722572327 + ], + "credit_map": [ + 0.041449666023254395, + 0.01570996269583702, + 0.028332972899079323, + 0.01732484996318817, + 0.022746602073311806, + 0.07916062325239182, + 0.015530252829194069, + 0.07032851874828339, + 0.016979705542325974, + 0.007494381163269281, + 0.01663641631603241, + 0.01074269413948059, + 0.012937105260789394, + 0.018975457176566124, + 0.009879318997263908, + 0.02197043038904667, + 0.013409926556050777, + 0.007648965809494257, + 0.01529326755553484, + 0.00808072742074728, + 0.009954063221812248, + 0.01624193973839283, + 0.007807121612131596, + 0.016555264592170715, + 0.007670981343835592, + 0.00294347177259624, + 0.006625201087445021, + 0.0033741500228643417, + 0.0035099696833640337, + 0.008253725245594978, + 0.003848825581371784, + 0.011059058830142021, + 0.016184071078896523, + 0.008243112824857235, + 0.01801813393831253, + 0.010403384454548359, + 0.009501561522483826, + 0.018038470298051834, + 0.009889199398458004, + 0.013369942083954811, + 0.005091470200568438, + 0.002873236546292901, + 0.004629375413060188, + 0.00373106449842453, + 0.003340252209454775, + 0.014707349240779877, + 0.0030953078530728817, + 0.010151871480047703, + 0.005371564067900181, + 0.002790850354358554, + 0.00486787548288703, + 0.0023803734220564365, + 0.003464812645688653, + 0.005354196298867464, + 0.0020388339180499315, + 0.005550506990402937, + 0.015036290511488914, + 0.007339883595705032, + 0.013721155934035778, + 0.010278336703777313, + 0.008329413831233978, + 0.04355539008975029, + 0.015597783029079437, + 0.03222561627626419, + 0.007429060526192188, + 0.005386412609368563, + 0.0077831014059484005, + 0.006445963867008686, + 0.005700970534235239, + 0.007946625351905823, + 0.00738809397444129, + 0.011601203121244907, + 0.008633323013782501, + 0.005229711998254061, + 0.0073988065123558044, + 0.00617036921903491, + 0.005928901955485344, + 0.009409275837242603, + 0.006609688047319651, + 0.013292208313941956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.062041785567998886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_15/index_tensor([626, 626])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_15/index_tensor([626, 626])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5aec601ce6e80efcb758b591671a99dec97125b6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_15/index_tensor([626, 626])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up, with a very shallow depth of field, initially holds razor-sharp focus on a single, magnificent sunflower in the immediate foreground. Its vibrant golden-yellow petals, each with a delicate, paper-like texture gently curling at the tip, frame a dark, velvety brown center composed of intricate, spiraling seeds. The entire flower head is bathed in the warm, soft afternoon sun, the fine hairs on its dense green stem shimmering in the sunlight. Then, in a smooth and deliberate rack-focus effect, the sharpness of the point gracefully transitions; as the focus pulls back, the sunflower dissolves into a soft, luminous sphere of color, neatly revealing a weathered wooden painter's easel standing a few feet behind it. The easel holds a canvas, its surface awash in bold, impressionistic brushstrokes. Thick, textured cadmium yellow, emerald green, and azure blue artfully mimic the scene, capturing its essence. Throughout the shot, the distant background remains a dreamy, pastel color – the endless field of sunflowers is rendered in a creamy bokeh of yellows and greens, meeting the pale blue sky at the horizon, creating an extremely serene and contemplative atmosphere, capturing a quiet moment of artistic interpretation in nature.", + "seed": 13191, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_15/index_626_seed_13191_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.025323358457535507, + "vlm_logprob": -0.025323358457535507, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.025323358457535507, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017045896500349045 + ], + "opposite_logprob_tokens": [ + -4.517045974731445 + ], + "logprob_reward": -0.017045896500349045, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.017045896500349045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06600237637758255, + 0.04758945852518082, + 0.082734115421772, + 0.24288973212242126, + 0.05969209223985672, + 0.09289062023162842, + 0.0263013057410717, + 0.06773009151220322, + 0.029660334810614586, + 0.2845098674297333 + ], + "credit_map": [ + 0.013890113681554794, + 0.009136246517300606, + 0.009442021138966084, + 0.01011477131396532, + 0.006929718889296055, + 0.007773785386234522, + 0.0036132975947111845, + 0.0051024192944169044, + 0.008660424500703812, + 0.005077902693301439, + 0.013378460891544819, + 0.008064552210271358, + 0.0030729020945727825, + 0.002897750586271286, + 0.0024691082071512938, + 0.00396835757419467, + 0.015692245215177536, + 0.011678504757583141, + 0.014936641789972782, + 0.01254882849752903, + 0.007606396917253733, + 0.005452721379697323, + 0.0062826466746628284, + 0.008536133915185928, + 0.031335148960351944, + 0.03723251074552536, + 0.047691989690065384, + 0.03523828089237213, + 0.02069501206278801, + 0.03965446725487709, + 0.010408680886030197, + 0.020633649080991745, + 0.00988251343369484, + 0.015513352118432522, + 0.007738651242107153, + 0.009649247862398624, + 0.004101167432963848, + 0.004595552571117878, + 0.003230499569326639, + 0.004981111269444227, + 0.010289574973285198, + 0.015081127174198627, + 0.015474633313715458, + 0.021737275645136833, + 0.007104689721018076, + 0.004908686503767967, + 0.006555694155395031, + 0.011738939210772514, + 0.002756366040557623, + 0.004080463200807571, + 0.004523587878793478, + 0.004133441019803286, + 0.0027364727575331926, + 0.0015604729996994138, + 0.003083358285948634, + 0.0034271443728357553, + 0.012857849709689617, + 0.008947663940489292, + 0.011650844477117062, + 0.007889007218182087, + 0.0052976845763623714, + 0.006324613932520151, + 0.007308354601264, + 0.007454075384885073, + 0.005904638208448887, + 0.004966691602021456, + 0.004821851383894682, + 0.003923898097127676, + 0.002423147903755307, + 0.0019741670694202185, + 0.0026184648741036654, + 0.0030274763703346252, + 0.06468255817890167, + 0.05558668076992035, + 0.022851211950182915, + 0.04437601938843727, + 0.01596507802605629, + 0.014967972412705421, + 0.03481953591108322, + 0.031260792165994644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.017045896500349045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371758013963699 + ], + "opposite_logprob_tokens": [ + -3.408717632293701 + ], + "logprob_reward": -0.03371758013963699, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371758013963699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06374891102313995, + 0.06698911637067795, + 0.0849442109465599, + 0.09808237850666046, + 0.1359439194202423, + 0.17235247790813446, + 0.040255699306726456, + 0.07893206924200058, + 0.04265453666448593, + 0.2160966545343399 + ], + "credit_map": [ + 0.013213586993515491, + 0.007862059399485588, + 0.007928719744086266, + 0.008645744062960148, + 0.005830519367009401, + 0.007225649897009134, + 0.003925502300262451, + 0.009117127396166325, + 0.012553836219012737, + 0.00721001485362649, + 0.015295881778001785, + 0.011157330125570297, + 0.0049315085634589195, + 0.005367509555071592, + 0.003912038169801235, + 0.006560998037457466, + 0.016921263188123703, + 0.013702116906642914, + 0.018496766686439514, + 0.013851974159479141, + 0.005990545265376568, + 0.0047363643534481525, + 0.004871815908700228, + 0.006373364478349686, + 0.01125936210155487, + 0.013701488263905048, + 0.019682858139276505, + 0.014971609227359295, + 0.008108179084956646, + 0.016980327665805817, + 0.004805759992450476, + 0.008572790771722794, + 0.021038465201854706, + 0.04275001212954521, + 0.019478006288409233, + 0.02380295656621456, + 0.007406787946820259, + 0.007703427225351334, + 0.005791701376438141, + 0.007972564548254013, + 0.012773030437529087, + 0.012919043190777302, + 0.03949670121073723, + 0.04751245677471161, + 0.009119756519794464, + 0.008802099153399467, + 0.007662010844796896, + 0.03406738489866257, + 0.010778246447443962, + 0.005796869285404682, + 0.00605026725679636, + 0.006389514077454805, + 0.002663139021024108, + 0.0021018851548433304, + 0.002869277261197567, + 0.003606498707085848, + 0.020688988268375397, + 0.011081174947321415, + 0.0148516446352005, + 0.007483907975256443, + 0.004447427112609148, + 0.006974800024181604, + 0.00661730021238327, + 0.006786826997995377, + 0.008861989714205265, + 0.00781552866101265, + 0.00911115575581789, + 0.006066580768674612, + 0.0029044817201793194, + 0.0021560522727668285, + 0.0026293734554201365, + 0.0031093740835785866, + 0.05182097479701042, + 0.033126961439847946, + 0.018300576135516167, + 0.034619756042957306, + 0.01328858733177185, + 0.011821763589978218, + 0.028360435739159584, + 0.024757610633969307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.03371758013963699 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002915441989899 + ], + "opposite_logprob_tokens": [ + -2.6950292587280273 + ], + "logprob_reward": -0.07002915441989899, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002915441989899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05191866308450699, + 0.026308437809348106, + 0.04984967038035393, + 0.0761846974492073, + 0.0533941276371479, + 0.08738542348146439, + 0.028693903237581253, + 0.22992746531963348, + 0.05928951874375343, + 0.33704814314842224 + ], + "credit_map": [ + 0.012710284441709518, + 0.004762095864862204, + 0.006545127835124731, + 0.00825637299567461, + 0.004124408587813377, + 0.004766668193042278, + 0.0032887125853449106, + 0.007464990951120853, + 0.004964714869856834, + 0.002360017504543066, + 0.006361253093928099, + 0.004298977553844452, + 0.001988084986805916, + 0.0017138604307547212, + 0.0014970634365454316, + 0.0031244626734405756, + 0.012157795950770378, + 0.007032322231680155, + 0.011809093877673149, + 0.00820887740701437, + 0.002888850634917617, + 0.0024928185157477856, + 0.0018622798379510641, + 0.0033976295962929726, + 0.008008521050214767, + 0.010541178286075592, + 0.013706322759389877, + 0.011336803436279297, + 0.0069597945548594, + 0.014163863845169544, + 0.0038468537386506796, + 0.007621347904205322, + 0.006319616921246052, + 0.012810856103897095, + 0.010975819081068039, + 0.00841380562633276, + 0.003245027270168066, + 0.003545931540429592, + 0.003395106177777052, + 0.004687954671680927, + 0.006107785273343325, + 0.007288720924407244, + 0.02371109090745449, + 0.02091626450419426, + 0.00414924556389451, + 0.0036985627375543118, + 0.004202426411211491, + 0.017311325296759605, + 0.005312828812748194, + 0.005455809645354748, + 0.005209393333643675, + 0.004121310543268919, + 0.0019066659733653069, + 0.0014589839847758412, + 0.0018031602958217263, + 0.0034257471561431885, + 0.07253992557525635, + 0.03211628645658493, + 0.03501791134476662, + 0.026138899847865105, + 0.013791399076581001, + 0.012209075503051281, + 0.019567588344216347, + 0.018546344712376595, + 0.009466778486967087, + 0.007602189667522907, + 0.011729858815670013, + 0.011827217414975166, + 0.004142743535339832, + 0.0027485298924148083, + 0.005690212361514568, + 0.006081981584429741, + 0.06330130249261856, + 0.08646360784769058, + 0.02710028551518917, + 0.06354708969593048, + 0.016648508608341217, + 0.015137563459575176, + 0.03364149481058121, + 0.03120826557278633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.07002915441989899 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642422776669264 + ], + "opposite_logprob_tokens": [ + -5.379642486572266 + ], + "logprob_reward": -0.004642422776669264, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642422776669264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07184299826622009, + 0.06411682814359665, + 0.13612066209316254, + 0.10720869898796082, + 0.2083127647638321, + 0.06406163424253464, + 0.06617175787687302, + 0.15249288082122803, + 0.034965306520462036, + 0.09470639377832413 + ], + "credit_map": [ + 0.013921268284320831, + 0.011629922315478325, + 0.009436914697289467, + 0.009106927551329136, + 0.006409407593309879, + 0.011045184917747974, + 0.0029873372986912727, + 0.007306037936359644, + 0.014115762896835804, + 0.006222554016858339, + 0.01355775073170662, + 0.010289044119417667, + 0.00604505930095911, + 0.004362781066447496, + 0.0036092642694711685, + 0.005914607085287571, + 0.031685970723629, + 0.02627359703183174, + 0.027193620800971985, + 0.02106746844947338, + 0.007396422792226076, + 0.008656324818730354, + 0.00567627651616931, + 0.008170979097485542, + 0.013619537465274334, + 0.018492070958018303, + 0.02377508580684662, + 0.016850313171744347, + 0.008154138922691345, + 0.014696236699819565, + 0.003961459267884493, + 0.00765985669568181, + 0.02643263339996338, + 0.04457239806652069, + 0.05126914754509926, + 0.03214913606643677, + 0.013332773931324482, + 0.011837917380034924, + 0.012252315878868103, + 0.016466442495584488, + 0.0037949744146317244, + 0.0060940636321902275, + 0.01595119759440422, + 0.012565411627292633, + 0.003993595950305462, + 0.004517773631960154, + 0.004147047642618418, + 0.012997576035559177, + 0.014342961832880974, + 0.00903522502630949, + 0.01137295737862587, + 0.009340324439108372, + 0.004377196542918682, + 0.004350480157881975, + 0.005468337330967188, + 0.007884274236857891, + 0.047560326755046844, + 0.022946840152144432, + 0.024766679853200912, + 0.01514444313943386, + 0.009366748854517937, + 0.007962901145219803, + 0.012208692729473114, + 0.012536244466900826, + 0.005296250339597464, + 0.0037262437399476767, + 0.005586694926023483, + 0.007722571957856417, + 0.00345406006090343, + 0.002215893240645528, + 0.003536751726642251, + 0.00342684262432158, + 0.019344467669725418, + 0.019779464229941368, + 0.010296918451786041, + 0.012348372489213943, + 0.007984654977917671, + 0.005125218071043491, + 0.010370263829827309, + 0.00945703312754631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 2 + ], + "selected_reward": -0.004642422776669264 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011817384511232376 + ], + "opposite_logprob_tokens": [ + -6.751181602478027 + ], + "logprob_reward": -0.0011817384511232376, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011817384511232376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15361888706684113, + 0.1728653609752655, + 0.06919469684362411, + 0.08381014317274094, + 0.15537478029727936, + 0.0803803876042366, + 0.05775978043675423, + 0.12945987284183502, + 0.02912595123052597, + 0.06841017305850983 + ], + "credit_map": [ + 0.027857493609189987, + 0.013243593275547028, + 0.01990680582821369, + 0.02438349835574627, + 0.015689047053456306, + 0.016340741887688637, + 0.012802883982658386, + 0.023394828662276268, + 0.02349264547228813, + 0.015950223430991173, + 0.05583376809954643, + 0.030833596363663673, + 0.012442763894796371, + 0.010319718159735203, + 0.009506914764642715, + 0.014485718682408333, + 0.010247691534459591, + 0.009966252371668816, + 0.016364024952054024, + 0.010837108828127384, + 0.004660202190279961, + 0.005794856697320938, + 0.00519300764426589, + 0.006131553556770086, + 0.009201143868267536, + 0.013598667457699776, + 0.019976573064923286, + 0.01270794216543436, + 0.005735112354159355, + 0.012097981758415699, + 0.0036321443039923906, + 0.006860575173050165, + 0.024721523746848106, + 0.04621795937418938, + 0.020103733986616135, + 0.02608884498476982, + 0.012181367725133896, + 0.007738720625638962, + 0.008054304867982864, + 0.010268323123455048, + 0.007544329855591059, + 0.009665443561971188, + 0.021464642137289047, + 0.016158344224095345, + 0.0051587666384875774, + 0.005082559771835804, + 0.005369297228753567, + 0.009937010705471039, + 0.006061561871320009, + 0.009131710045039654, + 0.008658192120492458, + 0.01003269013017416, + 0.004845192190259695, + 0.003664691699668765, + 0.006954162381589413, + 0.008411583490669727, + 0.035935577005147934, + 0.017527364194393158, + 0.024412164464592934, + 0.015010352246463299, + 0.0074194204062223434, + 0.006790731567889452, + 0.010545213706791401, + 0.011819044128060341, + 0.005259481258690357, + 0.0035254976246505976, + 0.003935693763196468, + 0.004402894992381334, + 0.0030025106389075518, + 0.00223733251914382, + 0.002973078517243266, + 0.00378946284763515, + 0.012089165858924389, + 0.011605771258473396, + 0.0059357378631830215, + 0.011182194575667381, + 0.004386520013213158, + 0.005873389542102814, + 0.008612778037786484, + 0.008724617771804333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0011817384511232376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up, with a very shallow depth of field, initially holds razor-sharp focus on a single, magnificent sunflower in the immediate foreground. Its vibrant golden-yellow petals, each with a delicate, paper-like texture gently curling at the tip, frame a dark, velvety brown center composed of intricate, spiraling seeds. The entire flower head is bathed in the warm, soft afternoon sun, the fine hairs on its dense green stem shimmering in the sunlight. Then, in a smooth and deliberate rack-focus effect, the sharpness of the point gracefully transitions; as the focus pulls back, the sunflower dissolves into a soft, luminous sphere of color, neatly revealing a weathered wooden painter's easel standing a few feet behind it. The easel holds a canvas, its surface awash in bold, impressionistic brushstrokes. Thick, textured cadmium yellow, emerald green, and azure blue artfully mimic the scene, capturing its essence. Throughout the shot, the distant background remains a dreamy, pastel color – the endless field of sunflowers is rendered in a creamy bokeh of yellows and greens, meeting the pale blue sky at the horizon, creating an extremely serene and contemplative atmosphere, capturing a quiet moment of artistic interpretation in nature.", + "seed": 13192, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_15/index_626_seed_13192_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.015661000180989502, + "vlm_logprob": -0.015661000180989502, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.015661000180989502, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016709022223949432 + ], + "opposite_logprob_tokens": [ + -4.266708850860596 + ], + "logprob_reward": -0.016709022223949432, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016709022223949432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08887159079313278, + 0.21972337365150452, + 0.18647946417331696, + 0.12607350945472717, + 0.1265307366847992, + 0.07198939472436905, + 0.05922268331050873, + 0.049400340765714645, + 0.037381481379270554, + 0.03432741016149521 + ], + "credit_map": [ + 0.007195279002189636, + 0.019561009481549263, + 0.010462570935487747, + 0.011209363117814064, + 0.0070986817590892315, + 0.010067927651107311, + 0.008665929548442364, + 0.014610831625759602, + 0.024738524109125137, + 0.025525430217385292, + 0.03035619482398033, + 0.02182677388191223, + 0.021610353142023087, + 0.03661526367068291, + 0.019628195092082024, + 0.03942263871431351, + 0.013393108732998371, + 0.01735096238553524, + 0.048536866903305054, + 0.010123665444552898, + 0.011225671507418156, + 0.01751973107457161, + 0.008868122473359108, + 0.059461332857608795, + 0.010278934612870216, + 0.015602911822497845, + 0.014711224474012852, + 0.01877516508102417, + 0.016512669622898102, + 0.012956630438566208, + 0.018584907054901123, + 0.018651071935892105, + 0.008980602957308292, + 0.008014025166630745, + 0.03432777151465416, + 0.006137370131909847, + 0.007413724437355995, + 0.009298431687057018, + 0.00785010401159525, + 0.04450869560241699, + 0.005377146881073713, + 0.004849323537200689, + 0.019211767241358757, + 0.0050308723002672195, + 0.004732918925583363, + 0.005695052910596132, + 0.0043698581866919994, + 0.022722458466887474, + 0.00445016473531723, + 0.005384915973991156, + 0.009170711040496826, + 0.01359287928789854, + 0.003841434372588992, + 0.004322320222854614, + 0.010350468568503857, + 0.008109785616397858, + 0.00727634783834219, + 0.004410690627992153, + 0.005941217765212059, + 0.008636253885924816, + 0.004124518949538469, + 0.005406274925917387, + 0.006174277979880571, + 0.007430759724229574, + 0.0049857813864946365, + 0.002740895375609398, + 0.005150807090103626, + 0.003787512192502618, + 0.0032408491242676973, + 0.005741190630942583, + 0.004724207799881697, + 0.007010234519839287, + 0.005391768645495176, + 0.0031409640796482563, + 0.00552621902897954, + 0.0027623004280030727, + 0.00300621404312551, + 0.005081134382635355, + 0.004166493192315102, + 0.005252316128462553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.016709022223949432 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594803616404533 + ], + "opposite_logprob_tokens": [ + -3.895594835281372 + ], + "logprob_reward": -0.020594803616404533, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594803616404533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23666952550411224, + 0.18239372968673706, + 0.1562606692314148, + 0.023793933913111687, + 0.11227616667747498, + 0.06132039427757263, + 0.04800860956311226, + 0.0609777458012104, + 0.06117711216211319, + 0.057122111320495605 + ], + "credit_map": [ + 0.021334746852517128, + 0.06767989695072174, + 0.022189510986208916, + 0.026043934747576714, + 0.022325970232486725, + 0.02821921929717064, + 0.01527335587888956, + 0.03360288217663765, + 0.0161838810890913, + 0.01953674480319023, + 0.026433993130922318, + 0.022201674059033394, + 0.019200647249817848, + 0.02780883200466633, + 0.019432706758379936, + 0.03159525617957115, + 0.009095914661884308, + 0.012666800059378147, + 0.04426111280918121, + 0.008829308673739433, + 0.007559207733720541, + 0.011819710023701191, + 0.008648998104035854, + 0.053379613906145096, + 0.0023293455597013235, + 0.0019305768655613065, + 0.0028302797582000494, + 0.0035544175188988447, + 0.0038953665643930435, + 0.0028099287301301956, + 0.00327565660700202, + 0.003168361494317651, + 0.008097514510154724, + 0.007672950625419617, + 0.028438931331038475, + 0.007988175377249718, + 0.006950414273887873, + 0.006781889125704765, + 0.010260192677378654, + 0.036086101084947586, + 0.005547726526856422, + 0.004714347887784243, + 0.016396082937717438, + 0.003577212104573846, + 0.002359559293836355, + 0.005319242365658283, + 0.0035820018965750933, + 0.01982422173023224, + 0.004286563955247402, + 0.00279467785730958, + 0.006829432677477598, + 0.005678561516106129, + 0.0039178128354251385, + 0.006086645647883415, + 0.008458658121526241, + 0.009956253692507744, + 0.005105686839669943, + 0.005466506350785494, + 0.008314970880746841, + 0.009120438247919083, + 0.005495223216712475, + 0.00430018175393343, + 0.010374659672379494, + 0.012800081633031368, + 0.005081382114440203, + 0.0043742721900343895, + 0.00768783176317811, + 0.009332295507192612, + 0.006707820110023022, + 0.007294206414371729, + 0.0094938138499856, + 0.011205491609871387, + 0.008463270030915737, + 0.005266865715384483, + 0.010947128757834435, + 0.004439252428710461, + 0.003729144809767604, + 0.006305815652012825, + 0.004575551487505436, + 0.013395078480243683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.020594803616404533 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370801359415054 + ], + "opposite_logprob_tokens": [ + -3.408708095550537 + ], + "logprob_reward": -0.03370801359415054, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370801359415054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0991169735789299, + 0.23050306737422943, + 0.40326014161109924, + 0.04851064458489418, + 0.025192750617861748, + 0.073792465031147, + 0.021274279803037643, + 0.040452226996421814, + 0.04471553862094879, + 0.01318181399255991 + ], + "credit_map": [ + 0.008475009351968765, + 0.030549418181180954, + 0.007052684668451548, + 0.013597245328128338, + 0.011418337002396584, + 0.009087897837162018, + 0.008787872269749641, + 0.01014851313084364, + 0.021204106509685516, + 0.03486170619726181, + 0.030254412442445755, + 0.031849220395088196, + 0.024558095261454582, + 0.03485587611794472, + 0.02534441649913788, + 0.027575219050049782, + 0.022693581879138947, + 0.022605594247579575, + 0.1289215385913849, + 0.01753774657845497, + 0.015629760921001434, + 0.02190430276095867, + 0.01492974255234003, + 0.15903791785240173, + 0.003392589744180441, + 0.0039981394074857235, + 0.0058822776190936565, + 0.0076830023899674416, + 0.005248652305454016, + 0.004167073871940374, + 0.009870002046227455, + 0.008268908597528934, + 0.002392612863332033, + 0.001710819429717958, + 0.0038608326576650143, + 0.0025664023123681545, + 0.0021881856955587864, + 0.0048542991280555725, + 0.0027820118702948093, + 0.004837585613131523, + 0.007340691052377224, + 0.005070678889751434, + 0.017617324367165565, + 0.005777850281447172, + 0.00482944305986166, + 0.006498595234006643, + 0.005409932229667902, + 0.021247945725917816, + 0.002235128078609705, + 0.0020846424158662558, + 0.003115416970103979, + 0.002736992435529828, + 0.0019168790895491838, + 0.002400970784947276, + 0.002697422867640853, + 0.0040868292562663555, + 0.004615561105310917, + 0.002932370873168111, + 0.005749422125518322, + 0.007158507127314806, + 0.003617881564423442, + 0.0035984048154205084, + 0.0045629763044416904, + 0.008217108435928822, + 0.003590679494664073, + 0.0034179226495325565, + 0.005476437509059906, + 0.0077765085734426975, + 0.005015768576413393, + 0.0034749323967844248, + 0.008249391801655293, + 0.007713898550719023, + 0.0016006329096853733, + 0.0011938584502786398, + 0.002288747113198042, + 0.001237768679857254, + 0.0011765292147174478, + 0.0015399940311908722, + 0.001060029841028154, + 0.003084252355620265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.03370801359415054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954503081738949 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.005954503081738949, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954503081738949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20103251934051514, + 0.0314396433532238, + 0.28088298439979553, + 0.026697687804698944, + 0.3070512115955353, + 0.030016345903277397, + 0.014834223315119743, + 0.01672850362956524, + 0.03312963247299194, + 0.05818729102611542 + ], + "credit_map": [ + 0.01736804097890854, + 0.05085323005914688, + 0.020140105858445168, + 0.024705130606889725, + 0.021037230268120766, + 0.026605835184454918, + 0.014247122220695019, + 0.026075828820466995, + 0.002824238268658519, + 0.006346503272652626, + 0.0043428810313344, + 0.0039322953671216965, + 0.0025438815355300903, + 0.004155878908932209, + 0.0029955755453556776, + 0.004298385698348284, + 0.016046689823269844, + 0.016789736226201057, + 0.08741899579763412, + 0.014131220988929272, + 0.011632589623332024, + 0.0170974712818861, + 0.010578407905995846, + 0.10718785971403122, + 0.0019753878004848957, + 0.002538966480642557, + 0.0037446734495460987, + 0.003848125459626317, + 0.0021688933484256268, + 0.00237606605514884, + 0.005457595456391573, + 0.004587976727634668, + 0.02432490698993206, + 0.019614312797784805, + 0.07946354150772095, + 0.01795348711311817, + 0.018851930275559425, + 0.023402215912938118, + 0.019627174362540245, + 0.10381361097097397, + 0.003026345046237111, + 0.002379032550379634, + 0.006260695867240429, + 0.002283086534589529, + 0.00282670883461833, + 0.0025396968703716993, + 0.0023377540055662394, + 0.008363023400306702, + 0.0017232424579560757, + 0.0010127718560397625, + 0.002568701049312949, + 0.002076955046504736, + 0.001032434869557619, + 0.0018557667499408126, + 0.0021918886341154575, + 0.002372462535277009, + 0.0019520525820553303, + 0.0010907721007242799, + 0.0021309524308890104, + 0.0022332947701215744, + 0.0014596019173040986, + 0.002274417784065008, + 0.002475205110386014, + 0.0031122060026973486, + 0.003327347571030259, + 0.0030673518776893616, + 0.0049369619227945805, + 0.004753587767481804, + 0.002475201152265072, + 0.0028777027036994696, + 0.00586147652938962, + 0.005830000154674053, + 0.009692679159343243, + 0.005109318532049656, + 0.010738612152636051, + 0.00457254983484745, + 0.0034455556888133287, + 0.0062893712893128395, + 0.005414706189185381, + 0.012924493290483952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.005954503081738949 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013386583887040615 + ], + "opposite_logprob_tokens": [ + -6.626338481903076 + ], + "logprob_reward": -0.0013386583887040615, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013386583887040615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1791079193353653, + 0.24201463162899017, + 0.08534255623817444, + 0.09983040392398834, + 0.155891552567482, + 0.07751937955617905, + 0.06039953604340553, + 0.047163620591163635, + 0.03334127366542816, + 0.01938912458717823 + ], + "credit_map": [ + 0.017072396352887154, + 0.049445345997810364, + 0.014977254904806614, + 0.02337728813290596, + 0.01941313035786152, + 0.019077947363257408, + 0.015561233274638653, + 0.02018331177532673, + 0.02980859950184822, + 0.037199269980192184, + 0.030271882191300392, + 0.024578340351581573, + 0.02561318688094616, + 0.04205499216914177, + 0.023863188922405243, + 0.028625205159187317, + 0.008203621953725815, + 0.010053928010165691, + 0.012434103526175022, + 0.007268030196428299, + 0.009063380770385265, + 0.01168033666908741, + 0.01079147681593895, + 0.015847675502300262, + 0.007523472420871258, + 0.01181921735405922, + 0.016260258853435516, + 0.014387834817171097, + 0.009314040653407574, + 0.008158056065440178, + 0.01359413843601942, + 0.0187733955681324, + 0.011926112696528435, + 0.006555995438247919, + 0.03588356822729111, + 0.013355939649045467, + 0.012486031278967857, + 0.020684456452727318, + 0.010545386001467705, + 0.044454071670770645, + 0.00595425721257925, + 0.0052468860521912575, + 0.01906609907746315, + 0.006562268827110529, + 0.005166932940483093, + 0.0075865937396883965, + 0.007120277266949415, + 0.02081606723368168, + 0.004838653840124607, + 0.005583065561950207, + 0.009040715172886848, + 0.012354833073914051, + 0.004013019613921642, + 0.004851396195590496, + 0.01115013100206852, + 0.008567724376916885, + 0.004912217613309622, + 0.004098020028322935, + 0.005751889664679766, + 0.007283133454620838, + 0.004057579208165407, + 0.0051775164902210236, + 0.006385534070432186, + 0.009497735649347305, + 0.0028399897273629904, + 0.002810965059325099, + 0.00526232086122036, + 0.00453009782359004, + 0.002290017670020461, + 0.0032266974449157715, + 0.006081821396946907, + 0.006299363914877176, + 0.0026919725351035595, + 0.0018602204509079456, + 0.003047981997951865, + 0.0025339205749332905, + 0.00159132806584239, + 0.0019529879791662097, + 0.002052614465355873, + 0.003658097703009844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0013386583887040615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up, with a very shallow depth of field, initially holds razor-sharp focus on a single, magnificent sunflower in the immediate foreground. Its vibrant golden-yellow petals, each with a delicate, paper-like texture gently curling at the tip, frame a dark, velvety brown center composed of intricate, spiraling seeds. The entire flower head is bathed in the warm, soft afternoon sun, the fine hairs on its dense green stem shimmering in the sunlight. Then, in a smooth and deliberate rack-focus effect, the sharpness of the point gracefully transitions; as the focus pulls back, the sunflower dissolves into a soft, luminous sphere of color, neatly revealing a weathered wooden painter's easel standing a few feet behind it. The easel holds a canvas, its surface awash in bold, impressionistic brushstrokes. Thick, textured cadmium yellow, emerald green, and azure blue artfully mimic the scene, capturing its essence. Throughout the shot, the distant background remains a dreamy, pastel color – the endless field of sunflowers is rendered in a creamy bokeh of yellows and greens, meeting the pale blue sky at the horizon, creating an extremely serene and contemplative atmosphere, capturing a quiet moment of artistic interpretation in nature.", + "seed": 13193, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_15/index_626_seed_13193_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026676541811320932, + "vlm_logprob": -0.026676541811320932, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.026676541811320932, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018521105870604515 + ], + "opposite_logprob_tokens": [ + -4.393521308898926 + ], + "logprob_reward": -0.018521105870604515, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.018521105870604515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31780725717544556, + 0.07966255396604538, + 0.15971307456493378, + 0.11638329923152924, + 0.058850694447755814, + 0.1010848805308342, + 0.07355180382728577, + 0.030428417026996613, + 0.03209266811609268, + 0.030425313860177994 + ], + "credit_map": [ + 0.050030384212732315, + 0.027335498481988907, + 0.0832299217581749, + 0.028363652527332306, + 0.025321613997220993, + 0.028321506455540657, + 0.03008834645152092, + 0.045116350054740906, + 0.016216428950428963, + 0.005894326604902744, + 0.02126428484916687, + 0.005620915908366442, + 0.004490406718105078, + 0.006653099786490202, + 0.005451983772218227, + 0.014071117155253887, + 0.014977024868130684, + 0.017186470329761505, + 0.03067917749285698, + 0.015641439706087112, + 0.011369619518518448, + 0.028239281848073006, + 0.014934699051082134, + 0.026685383170843124, + 0.011356789618730545, + 0.013146758079528809, + 0.025418788194656372, + 0.013635839335620403, + 0.01372470147907734, + 0.013774369843304157, + 0.008342321030795574, + 0.01698373816907406, + 0.007709840778261423, + 0.004881029482930899, + 0.008673912845551968, + 0.006305009592324495, + 0.006588883697986603, + 0.010344340465962887, + 0.007174482569098473, + 0.007173194084316492, + 0.007191868498921394, + 0.0088841887190938, + 0.022977126762270927, + 0.010241862386465073, + 0.011865725740790367, + 0.01543360948562622, + 0.009914788417518139, + 0.014575718902051449, + 0.007100061979144812, + 0.007386424113065004, + 0.01557846087962389, + 0.00671120872721076, + 0.008501032367348671, + 0.010470752604305744, + 0.0056350138038396835, + 0.012168857268989086, + 0.004367946647107601, + 0.0023539667017757893, + 0.005672220606356859, + 0.003171849763020873, + 0.0031616217456758022, + 0.00518077565357089, + 0.002175207482650876, + 0.004344832617789507, + 0.003655890468508005, + 0.006165810860693455, + 0.005163881927728653, + 0.0037298111710697412, + 0.0035535991191864014, + 0.0029618539847433567, + 0.002452474320307374, + 0.004409350920468569, + 0.004047730006277561, + 0.003973241429775953, + 0.005322035867720842, + 0.0034259429667145014, + 0.002428839448839426, + 0.0043945047073066235, + 0.002174004912376404, + 0.004659018479287624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.018521105870604515 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207282096147537 + ], + "opposite_logprob_tokens": [ + -2.81207275390625 + ], + "logprob_reward": -0.06207282096147537, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207282096147537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21705761551856995, + 0.09381898492574692, + 0.10400527715682983, + 0.05909710004925728, + 0.06987708061933517, + 0.052918776869773865, + 0.0918646976351738, + 0.0612865574657917, + 0.1078743189573288, + 0.1421995908021927 + ], + "credit_map": [ + 0.03490917384624481, + 0.019616376608610153, + 0.05009008198976517, + 0.01995830237865448, + 0.01994294486939907, + 0.018774790689349174, + 0.022556878626346588, + 0.031209072098135948, + 0.018192492425441742, + 0.010067565366625786, + 0.021611394360661507, + 0.007205754052847624, + 0.0059598335064947605, + 0.008151100017130375, + 0.00728600611910224, + 0.015344840474426746, + 0.011617968790233135, + 0.010306384414434433, + 0.028144292533397675, + 0.01111350953578949, + 0.00859618280082941, + 0.008009696379303932, + 0.008735247887670994, + 0.01748199388384819, + 0.005458900239318609, + 0.004871224518865347, + 0.014236615970730782, + 0.004587384406477213, + 0.007457924075424671, + 0.010316148400306702, + 0.003796803764998913, + 0.008372103795409203, + 0.005436280742287636, + 0.006507156416773796, + 0.013135481625795364, + 0.011122940108180046, + 0.00981995090842247, + 0.008518775925040245, + 0.005889615043997765, + 0.009446884505450726, + 0.005821187049150467, + 0.004195875022560358, + 0.012414735741913319, + 0.005620745941996574, + 0.004372879862785339, + 0.00811096653342247, + 0.003785555250942707, + 0.008596828207373619, + 0.009166932664811611, + 0.008552810177206993, + 0.017403291538357735, + 0.011957496404647827, + 0.007571196649223566, + 0.01795508898794651, + 0.007094902917742729, + 0.01216298621147871, + 0.007029218133538961, + 0.005973401945084333, + 0.012673740275204182, + 0.008999450132250786, + 0.006047731265425682, + 0.007080797106027603, + 0.005379785317927599, + 0.00810242909938097, + 0.011177596636116505, + 0.022755280137062073, + 0.016156546771526337, + 0.012948673218488693, + 0.01155540719628334, + 0.010844874195754528, + 0.007045506499707699, + 0.015390435233712196, + 0.01671545021235943, + 0.025667762383818626, + 0.01963120326399803, + 0.01619561016559601, + 0.0129953408613801, + 0.021529631689190865, + 0.00939089898020029, + 0.020073700696229935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.06207282096147537 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866647347807884 + ], + "opposite_logprob_tokens": [ + -3.048666477203369 + ], + "logprob_reward": -0.04866647347807884, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866647347807884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23322100937366486, + 0.0660526230931282, + 0.06519980728626251, + 0.0770675390958786, + 0.05525188148021698, + 0.05322245880961418, + 0.10580373555421829, + 0.08569277077913284, + 0.08628329634666443, + 0.17220492660999298 + ], + "credit_map": [ + 0.030425269156694412, + 0.02104383148252964, + 0.0587431937456131, + 0.02288658544421196, + 0.021465804427862167, + 0.02048666402697563, + 0.023993534967303276, + 0.03417612612247467, + 0.011682077310979366, + 0.006597424857318401, + 0.015297837555408478, + 0.005817111115902662, + 0.004437208641320467, + 0.00577196991071105, + 0.005983835551887751, + 0.010465162806212902, + 0.007195405196398497, + 0.005260626319795847, + 0.011986816301941872, + 0.0057044667191803455, + 0.0057911379262804985, + 0.013996802270412445, + 0.005528089590370655, + 0.009736464358866215, + 0.006705597974359989, + 0.007881342433393002, + 0.01594751700758934, + 0.006963513325899839, + 0.011610141955316067, + 0.01171186100691557, + 0.006166831590235233, + 0.010080734267830849, + 0.008398354984819889, + 0.003680566558614373, + 0.012451346963644028, + 0.006826487369835377, + 0.004863391630351543, + 0.007989324629306793, + 0.002948981011286378, + 0.008093424141407013, + 0.007722203619778156, + 0.004665466956794262, + 0.008092746138572693, + 0.005030862521380186, + 0.006297473330050707, + 0.01052902266383171, + 0.004140043165534735, + 0.0067446413449943066, + 0.010694949887692928, + 0.010728418827056885, + 0.02317199856042862, + 0.013546742498874664, + 0.012156065553426743, + 0.012455560266971588, + 0.007056003902107477, + 0.0159939993172884, + 0.01015661284327507, + 0.009176578372716904, + 0.014856877736747265, + 0.011793040670454502, + 0.009358745068311691, + 0.010316415689885616, + 0.007437604945152998, + 0.012596889398992062, + 0.009312694892287254, + 0.017156464979052544, + 0.012670169584453106, + 0.008801757358014584, + 0.009227712638676167, + 0.01027965173125267, + 0.0056460946798324585, + 0.013188745826482773, + 0.021269910037517548, + 0.02517182193696499, + 0.026177015155553818, + 0.021141381934285164, + 0.014298026449978352, + 0.028442498296499252, + 0.011119567789137363, + 0.024584710597991943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.04866647347807884 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032001733779907227 + ], + "opposite_logprob_tokens": [ + -5.753200054168701 + ], + "logprob_reward": -0.0032001733779907227, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032001733779907227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12385090440511703, + 0.0939699113368988, + 0.08425882458686829, + 0.17335940897464752, + 0.07982325553894043, + 0.12602394819259644, + 0.042596299201250076, + 0.07240486890077591, + 0.08235325664281845, + 0.12135936319828033 + ], + "credit_map": [ + 0.014275837689638138, + 0.012356862425804138, + 0.027452968060970306, + 0.013840817846357822, + 0.012902617454528809, + 0.010774267837405205, + 0.014808734878897667, + 0.01743878237903118, + 0.017408303916454315, + 0.00997140258550644, + 0.02252652309834957, + 0.008659887127578259, + 0.005728247109800577, + 0.008015752770006657, + 0.007029047701507807, + 0.014630735851824284, + 0.013003130443394184, + 0.006330014672130346, + 0.024819716811180115, + 0.006255097687244415, + 0.004605199676007032, + 0.009552286006510258, + 0.00502276374027133, + 0.014670604839920998, + 0.01884550414979458, + 0.011471092700958252, + 0.05543795973062515, + 0.013892843388020992, + 0.011203540489077568, + 0.021784234791994095, + 0.010947844944894314, + 0.029776372015476227, + 0.00792196486145258, + 0.00825253315269947, + 0.017934052273631096, + 0.010474270209670067, + 0.007528470363467932, + 0.010223166085779667, + 0.006608784664422274, + 0.010880008339881897, + 0.01512213982641697, + 0.009365157224237919, + 0.03208687901496887, + 0.00969703495502472, + 0.0058917272835969925, + 0.02781238593161106, + 0.005733700934797525, + 0.020314902067184448, + 0.004023366142064333, + 0.005421077366918325, + 0.00667330389842391, + 0.005330371204763651, + 0.005423600785434246, + 0.0071089016273617744, + 0.0028249267488718033, + 0.0057907430455088615, + 0.011361481621861458, + 0.006165459752082825, + 0.013047548942267895, + 0.008716760203242302, + 0.00664298003539443, + 0.011602123267948627, + 0.0053459592163562775, + 0.009522547945380211, + 0.013713249936699867, + 0.007899648509919643, + 0.011830293573439121, + 0.011130934581160545, + 0.006004297640174627, + 0.014927421696484089, + 0.005023439414799213, + 0.011823968030512333, + 0.0186189953237772, + 0.013088387437164783, + 0.02299456112086773, + 0.015758750960230827, + 0.008716939017176628, + 0.015592841431498528, + 0.008704107254743576, + 0.017884759232401848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.0032001733779907227 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009221353684552014 + ], + "opposite_logprob_tokens": [ + -7.000922203063965 + ], + "logprob_reward": -0.0009221353684552014, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009221353684552014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2593761682510376, + 0.25894978642463684, + 0.11137587577104568, + 0.07278924435377121, + 0.07968124747276306, + 0.0660640299320221, + 0.056946080178022385, + 0.028729861602187157, + 0.02750539593398571, + 0.0385822132229805 + ], + "credit_map": [ + 0.04176776856184006, + 0.025395013391971588, + 0.07077470421791077, + 0.018869340419769287, + 0.019362101331353188, + 0.023971768096089363, + 0.019958432763814926, + 0.039277080446481705, + 0.03882778808474541, + 0.02983608841896057, + 0.05070577561855316, + 0.030341556295752525, + 0.023109255358576775, + 0.021012071520090103, + 0.02695004642009735, + 0.038167234510183334, + 0.010456071235239506, + 0.012631329707801342, + 0.020293498411774635, + 0.010463612154126167, + 0.008262756280601025, + 0.021862532943487167, + 0.009671200066804886, + 0.017734883353114128, + 0.006351981777697802, + 0.00608859583735466, + 0.01845822110772133, + 0.0071523962542414665, + 0.00890975072979927, + 0.010704182088375092, + 0.00502926018089056, + 0.01009486522525549, + 0.007133688777685165, + 0.007414544932544231, + 0.015601866878569126, + 0.0073246113024652, + 0.009185009635984898, + 0.014593999832868576, + 0.008238658308982849, + 0.010188884101808071, + 0.007840145379304886, + 0.007165703922510147, + 0.01148287858814001, + 0.007685309741646051, + 0.00512111559510231, + 0.012998892925679684, + 0.004774231463670731, + 0.00899576023221016, + 0.004878764506429434, + 0.007147395517677069, + 0.009270005859434605, + 0.006722209043800831, + 0.008068893104791641, + 0.008737736381590366, + 0.0038937812205404043, + 0.00822729617357254, + 0.004646527115255594, + 0.002418900141492486, + 0.004333776421844959, + 0.0036608900409191847, + 0.0022313096560537815, + 0.00529150664806366, + 0.002293104538694024, + 0.003853850532323122, + 0.0032473658211529255, + 0.005921286530792713, + 0.0036359464284032583, + 0.0027893532533198595, + 0.003328857012093067, + 0.0026769759133458138, + 0.0021999692544341087, + 0.003705642418935895, + 0.005384030286222696, + 0.0036676477175205946, + 0.00881091970950365, + 0.0054151928052306175, + 0.0033427930902689695, + 0.004082907922565937, + 0.002674089279025793, + 0.005204635672271252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0009221353684552014 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up, with a very shallow depth of field, initially holds razor-sharp focus on a single, magnificent sunflower in the immediate foreground. Its vibrant golden-yellow petals, each with a delicate, paper-like texture gently curling at the tip, frame a dark, velvety brown center composed of intricate, spiraling seeds. The entire flower head is bathed in the warm, soft afternoon sun, the fine hairs on its dense green stem shimmering in the sunlight. Then, in a smooth and deliberate rack-focus effect, the sharpness of the point gracefully transitions; as the focus pulls back, the sunflower dissolves into a soft, luminous sphere of color, neatly revealing a weathered wooden painter's easel standing a few feet behind it. The easel holds a canvas, its surface awash in bold, impressionistic brushstrokes. Thick, textured cadmium yellow, emerald green, and azure blue artfully mimic the scene, capturing its essence. Throughout the shot, the distant background remains a dreamy, pastel color – the endless field of sunflowers is rendered in a creamy bokeh of yellows and greens, meeting the pale blue sky at the horizon, creating an extremely serene and contemplative atmosphere, capturing a quiet moment of artistic interpretation in nature.", + "seed": 13194, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_15/index_626_seed_13194_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.140706880367361, + "vlm_logprob": -0.140706880367361, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.140706880367361, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03168884664773941 + ], + "opposite_logprob_tokens": [ + -3.656688928604126 + ], + "logprob_reward": -0.03168884664773941, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03168884664773941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06128435954451561, + 0.2658241391181946, + 0.15195979177951813, + 0.06204400956630707, + 0.10991958528757095, + 0.07251448929309845, + 0.08467191457748413, + 0.06909167021512985, + 0.09306158870458603, + 0.02962842583656311 + ], + "credit_map": [ + 0.006973040755838156, + 0.005012498702853918, + 0.010194855742156506, + 0.008803515695035458, + 0.005383015144616365, + 0.009618558920919895, + 0.006890513934195042, + 0.00840835738927126, + 0.0177738256752491, + 0.007817559875547886, + 0.03004077821969986, + 0.017854392528533936, + 0.011106768622994423, + 0.05418267101049423, + 0.03200913220643997, + 0.09503902494907379, + 0.015145853161811829, + 0.010538921691477299, + 0.027217652648687363, + 0.014070381410419941, + 0.008666262030601501, + 0.021691368892788887, + 0.013599333353340626, + 0.04103000834584236, + 0.009731980040669441, + 0.0064765289425849915, + 0.006381804123520851, + 0.007861152291297913, + 0.006852500140666962, + 0.00689592445269227, + 0.00845533050596714, + 0.009388793259859085, + 0.023372702300548553, + 0.008379528298974037, + 0.021064287051558495, + 0.007832618430256844, + 0.006805329583585262, + 0.017177456989884377, + 0.009740875102579594, + 0.01554677914828062, + 0.007966438308358192, + 0.005832462105900049, + 0.01416698843240738, + 0.008625700138509274, + 0.007278505712747574, + 0.0066794054582715034, + 0.008987465873360634, + 0.01297752559185028, + 0.00757125997915864, + 0.008010273799300194, + 0.011194171383976936, + 0.01198555063456297, + 0.006233620923012495, + 0.0069887880235910416, + 0.01509737130254507, + 0.017590882256627083, + 0.00833260640501976, + 0.005914939101785421, + 0.008672017604112625, + 0.009022761136293411, + 0.004362362436950207, + 0.006575455889105797, + 0.009694130159914494, + 0.01651740074157715, + 0.014983129687607288, + 0.00742421206086874, + 0.01372027862817049, + 0.012089395895600319, + 0.006532147526741028, + 0.010493678972125053, + 0.014161109924316406, + 0.01365763321518898, + 0.0034153286833316088, + 0.0021048204507678747, + 0.0041693695820868015, + 0.004193567670881748, + 0.0021290539298206568, + 0.003039498580619693, + 0.005403623916208744, + 0.005173163488507271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.03168884664773941 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428389996290207 + ], + "opposite_logprob_tokens": [ + -2.017838954925537 + ], + "logprob_reward": -0.1428389996290207, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428389996290207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06350676715373993, + 0.2254912406206131, + 0.06461045145988464, + 0.09726857393980026, + 0.15712174773216248, + 0.03982507437467575, + 0.08244102448225021, + 0.06503552943468094, + 0.10868973284959793, + 0.09600989520549774 + ], + "credit_map": [ + 0.007417396176606417, + 0.0043356698006391525, + 0.005939343478530645, + 0.012382254004478455, + 0.005322894547134638, + 0.00808837078511715, + 0.01052432507276535, + 0.00949651189148426, + 0.012577418237924576, + 0.004975246265530586, + 0.03495994955301285, + 0.01100187562406063, + 0.006263197399675846, + 0.03999357298016548, + 0.02450869418680668, + 0.09121126681566238, + 0.007472100667655468, + 0.006847040262073278, + 0.010367227718234062, + 0.007648354861885309, + 0.005598639603704214, + 0.008131798356771469, + 0.005509513895958662, + 0.01303577795624733, + 0.015176642686128616, + 0.008771023713052273, + 0.009617534466087818, + 0.015422044321894646, + 0.007355870679020882, + 0.012749253772199154, + 0.01291080191731453, + 0.015265405178070068, + 0.014181707985699177, + 0.015810837969183922, + 0.024198494851589203, + 0.02172812819480896, + 0.025341631844639778, + 0.014184323139488697, + 0.024190517142415047, + 0.017486106604337692, + 0.0046474444679915905, + 0.003159307176247239, + 0.006235989276319742, + 0.005752429831773043, + 0.004995421040803194, + 0.0035213776864111423, + 0.005385405849665403, + 0.006127700209617615, + 0.008554850704967976, + 0.007331775967031717, + 0.010825207456946373, + 0.009888487868010998, + 0.005195138044655323, + 0.008134384639561176, + 0.016315223649144173, + 0.016195958480238914, + 0.006611742544919252, + 0.005070697050541639, + 0.008973472751677036, + 0.0079319067299366, + 0.004318186081945896, + 0.0054701464250683784, + 0.010480044409632683, + 0.016179336234927177, + 0.01696629449725151, + 0.006633371114730835, + 0.0140768988057971, + 0.013517915271222591, + 0.006807141471654177, + 0.014163406565785408, + 0.02008327841758728, + 0.016441423445940018, + 0.011017107404768467, + 0.008093048818409443, + 0.014606937766075134, + 0.011531425639986992, + 0.009165138006210327, + 0.008640842512249947, + 0.013953598216176033, + 0.019001804292201996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 8 + ], + "selected_reward": -0.1428389996290207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233873128890991 + ], + "opposite_logprob_tokens": [ + -0.8983873128890991 + ], + "logprob_reward": -0.5233873128890991, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233873128890991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04033118113875389, + 0.07319872826337814, + 0.11567141115665436, + 0.09563419222831726, + 0.1278049498796463, + 0.09552615880966187, + 0.09135318547487259, + 0.0672026053071022, + 0.1240292340517044, + 0.16924841701984406 + ], + "credit_map": [ + 0.003921120893210173, + 0.0032357252202928066, + 0.005720167886465788, + 0.00620837090536952, + 0.0025889533571898937, + 0.00704298447817564, + 0.004896967671811581, + 0.006716890726238489, + 0.004462433513253927, + 0.0014643638860434294, + 0.010953708551824093, + 0.003619240131229162, + 0.002060063648968935, + 0.01340238656848669, + 0.007902718149125576, + 0.029333803802728653, + 0.013715403154492378, + 0.009027605876326561, + 0.017995445057749748, + 0.009459210559725761, + 0.008235754445195198, + 0.018509147688746452, + 0.009707754477858543, + 0.02902108244597912, + 0.011484000831842422, + 0.010151919908821583, + 0.011809011921286583, + 0.01086938101798296, + 0.007809456903487444, + 0.01236659288406372, + 0.011863733641803265, + 0.019280094653367996, + 0.02658352628350258, + 0.010880879126489162, + 0.016481028869748116, + 0.012411156669259071, + 0.014635535888373852, + 0.01564974896609783, + 0.015417546033859253, + 0.015745505690574646, + 0.009422785602509975, + 0.010815303772687912, + 0.017596395686268806, + 0.009558500722050667, + 0.014576919376850128, + 0.008696770295500755, + 0.01280872244387865, + 0.012050753459334373, + 0.010079480707645416, + 0.0070785838179290295, + 0.013782964088022709, + 0.01067961473017931, + 0.0060506281442940235, + 0.008418768644332886, + 0.01680467650294304, + 0.018458455801010132, + 0.0062140305526554585, + 0.00543011212721467, + 0.009671930223703384, + 0.006397607736289501, + 0.004396297503262758, + 0.0052713374607264996, + 0.01349689718335867, + 0.016324389725923538, + 0.013750397600233555, + 0.009295829571783543, + 0.016574088484048843, + 0.015928229317069054, + 0.009270947426557541, + 0.017202727496623993, + 0.02219628356397152, + 0.019810719415545464, + 0.01985582895576954, + 0.014704132452607155, + 0.024264687672257423, + 0.019328927621245384, + 0.018971679732203484, + 0.013914025388658047, + 0.023528842255473137, + 0.03468028083443642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.5233873128890991 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104048945009708 + ], + "opposite_logprob_tokens": [ + -5.504104137420654 + ], + "logprob_reward": -0.004104048945009708, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104048945009708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15954339504241943, + 0.060207728296518326, + 0.07213056832551956, + 0.019014853984117508, + 0.06842220574617386, + 0.12746968865394592, + 0.05479730665683746, + 0.05850088223814964, + 0.08428450673818588, + 0.2956289052963257 + ], + "credit_map": [ + 0.016581669449806213, + 0.016668468713760376, + 0.02855030819773674, + 0.020987968891859055, + 0.014318585395812988, + 0.01594400219619274, + 0.018781891092658043, + 0.027710506692528725, + 0.00419317651540041, + 0.0035417962353676558, + 0.00822308287024498, + 0.004157917574048042, + 0.0034145840909332037, + 0.010429502464830875, + 0.006537569221109152, + 0.019710101187229156, + 0.007518433034420013, + 0.0033633371349424124, + 0.01056822668761015, + 0.0048173838295042515, + 0.0027606997173279524, + 0.014026413671672344, + 0.00663487846031785, + 0.02244119718670845, + 0.00268124183639884, + 0.002161037176847458, + 0.0027609807439148426, + 0.0027217238675802946, + 0.0019511214923113585, + 0.0020923535339534283, + 0.0020003379322588444, + 0.002646057400852442, + 0.013273876160383224, + 0.0050307209603488445, + 0.013400708325207233, + 0.005690128076821566, + 0.005602700635790825, + 0.008677331730723381, + 0.0066632903181016445, + 0.010083446279168129, + 0.011554094031453133, + 0.0146991778165102, + 0.024602826684713364, + 0.011171949096024036, + 0.020265478640794754, + 0.011588041670620441, + 0.016418332234025, + 0.017169786617159843, + 0.006978660821914673, + 0.00324298907071352, + 0.009661865420639515, + 0.006039580795913935, + 0.0038155806250870228, + 0.005163981579244137, + 0.010492175817489624, + 0.009402468800544739, + 0.005659592337906361, + 0.005451949313282967, + 0.00882963091135025, + 0.007597772404551506, + 0.0043750787153840065, + 0.00569713581353426, + 0.007022561505436897, + 0.01386716216802597, + 0.00980233121663332, + 0.006547552067786455, + 0.012392875738441944, + 0.010173655115067959, + 0.006775699555873871, + 0.01163909025490284, + 0.012714823707938194, + 0.014238479547202587, + 0.03491557016968727, + 0.025075335055589676, + 0.04847678169608116, + 0.025348884984850883, + 0.040929507464170456, + 0.032895952463150024, + 0.02889307588338852, + 0.059093791991472244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.004104048945009708 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a static, eye-level close-up of a single sunflower in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video feature a smooth rack-focus transition where the sunflower dissolves into a soft sphere of color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weathered wooden painter's easel with a canvas holding impressionistic brushstrokes visible after the focus shift?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a creamy bokeh of yellow and green sunflowers meeting a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sunflower depicted with vibrant golden-yellow petals and a dark, velvety brown center?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015151937259361148 + ], + "opposite_logprob_tokens": [ + -6.5015153884887695 + ], + "logprob_reward": -0.0015151937259361148, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015151937259361148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12989254295825958, + 0.43004104495048523, + 0.029183058068156242, + 0.05114806815981865, + 0.09705683588981628, + 0.06625102460384369, + 0.03751390054821968, + 0.036765627562999725, + 0.05259670689702034, + 0.06955117732286453 + ], + "credit_map": [ + 0.011574133299291134, + 0.012304969131946564, + 0.01812703162431717, + 0.019519595429301262, + 0.011879566125571728, + 0.01622886210680008, + 0.016800591722130775, + 0.023457791656255722, + 0.029355525970458984, + 0.011343492195010185, + 0.05135190486907959, + 0.028063269332051277, + 0.017174288630485535, + 0.08625071495771408, + 0.048867568373680115, + 0.15763428807258606, + 0.003664093092083931, + 0.0017021159874275327, + 0.004786433186382055, + 0.003421617206186056, + 0.0015879477141425014, + 0.0038725046906620264, + 0.003024677513167262, + 0.007123669143766165, + 0.007859994657337666, + 0.005658862646669149, + 0.00482909195125103, + 0.005930914543569088, + 0.004654743708670139, + 0.00704288762062788, + 0.00609592255204916, + 0.009075653739273548, + 0.01500843558460474, + 0.009751624427735806, + 0.01407140213996172, + 0.01068542618304491, + 0.015118125826120377, + 0.008555547334253788, + 0.01302255317568779, + 0.010843714699149132, + 0.006377174519002438, + 0.008223329670727253, + 0.012968940660357475, + 0.005795348901301622, + 0.010187280364334583, + 0.0062706670723855495, + 0.007662253919988871, + 0.008766024373471737, + 0.004458643961697817, + 0.004043187014758587, + 0.005459148902446032, + 0.004573477432131767, + 0.0030447342433035374, + 0.003891188418492675, + 0.006098234094679356, + 0.005945285316556692, + 0.004658619407564402, + 0.0026663572061806917, + 0.005628284532576799, + 0.004150460474193096, + 0.0024569116067141294, + 0.0032602795399725437, + 0.005915545392781496, + 0.008029170334339142, + 0.006237829104065895, + 0.00423874007537961, + 0.006467489991337061, + 0.007552741095423698, + 0.003645321587100625, + 0.007109420839697123, + 0.008950994350016117, + 0.008394170552492142, + 0.008920310996472836, + 0.005781999323517084, + 0.011167784221470356, + 0.006113767623901367, + 0.010174775496125221, + 0.007600992452353239, + 0.006205476820468903, + 0.013586072251200676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0015151937259361148 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_2/index_tensor([5690, 5690])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_2/index_tensor([5690, 5690])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b6f2685a56e6111a3d78c8d4f26be4fbc869e066 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_2/index_tensor([5690, 5690])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium close-up captures an intimate and tense moment in a luxurious dressing room. A tall, statuesque woman with her sleek, raven-black hair pulled into a tight, high ponytail stands beside a shorter woman seated on a velvet stool. The tall woman, clad in an exquisite sleeveless black satin gown, exudes a calm authority in her posture. She is completely focused, one hand firmly and deliberately cupping the shorter woman's chin while the other carefully applies vibrant pale red lipstick to her lips. The shorter woman, with soft, wavy blonde hair and a simple cream silk blouse, looks up with a mixture of fear and resignation, her body tense. The background softly blurs, revealing a large vanity mirror framed by warm, glowing Hollywood-style bulbs, casting a soft golden light across the scene, creating delicate highlights on their skin and the glossy tube of lipstick. After completing the final stroke, the tall woman steps back, her expression shifting from concentration to calm scrutiny. She gave a curt, critical nod, then dismissed the smaller woman with a swift, decisive wave of her hand, a flick of her wrist that was both elegant and final, breaking eye contact as she shifted her focus, leaving behind a thick atmosphere of unspoken hierarchy and aloof control.", + "seed": 6957, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_2/index_5690_seed_6957_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11017836518585682, + "vlm_logprob": -0.11017836518585682, + "vlm_yes_no_margin": 3.5249999523162843, + "vlm_token_credit": -0.11017836518585682, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1442396491765976 + ], + "opposite_logprob_tokens": [ + -4.14423942565918 + ], + "logprob_reward": -0.1442396491765976, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.1442396491765976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15080705285072327, + 0.08621232211589813, + 0.039522577077150345, + 0.08825217932462692, + 0.05454978719353676, + 0.12047972530126572, + 0.11492890864610672, + 0.15244102478027344, + 0.10263079404830933, + 0.09017565101385117 + ], + "credit_map": [ + 0.019214021041989326, + 0.022917425259947777, + 0.02108275331556797, + 0.019144520163536072, + 0.021040374413132668, + 0.011914603412151337, + 0.021324196830391884, + 0.014169151894748211, + 0.010371843352913857, + 0.014939977787435055, + 0.012049568817019463, + 0.011847159825265408, + 0.010049900971353054, + 0.0054455772042274475, + 0.013918882235884666, + 0.007589413318783045, + 0.005753455683588982, + 0.007449732162058353, + 0.005010525695979595, + 0.006454776972532272, + 0.003953613806515932, + 0.002760289004072547, + 0.003741397289559245, + 0.004398783668875694, + 0.013990817591547966, + 0.020427726209163666, + 0.011089681647717953, + 0.011801130138337612, + 0.010232177563011646, + 0.004520352929830551, + 0.009614347480237484, + 0.00657594483345747, + 0.00716797448694706, + 0.009760220535099506, + 0.005738428793847561, + 0.012200366705656052, + 0.0057568736374378204, + 0.0032840236090123653, + 0.005360237788408995, + 0.005281660705804825, + 0.017772013321518898, + 0.028914043679833412, + 0.011808004230260849, + 0.018252767622470856, + 0.016208728775382042, + 0.006471159402281046, + 0.011994975619018078, + 0.009058032184839249, + 0.017823370173573494, + 0.011862028390169144, + 0.008964767679572105, + 0.01581578515470028, + 0.015482616610825062, + 0.016882263123989105, + 0.013842095620930195, + 0.01425598282366991, + 0.01757526583969593, + 0.006161247845739126, + 0.011790147982537746, + 0.014203504659235477, + 0.016876058652997017, + 0.03479126840829849, + 0.02047712169587612, + 0.030566396191716194, + 0.009327812120318413, + 0.006530902814120054, + 0.010723398067057133, + 0.011317546479403973, + 0.015157641842961311, + 0.017722798511385918, + 0.014986321330070496, + 0.016864372417330742, + 0.009238330647349358, + 0.009037061594426632, + 0.010248511098325253, + 0.014537638053297997, + 0.013879184611141682, + 0.007915331982076168, + 0.014956777915358543, + 0.01036281231790781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.1442396491765976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008681767620146275 + ], + "opposite_logprob_tokens": [ + -4.758681774139404 + ], + "logprob_reward": -0.008681767620146275, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008681767620146275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1620221734046936, + 0.09648296982049942, + 0.08894176781177521, + 0.0746244341135025, + 0.045470334589481354, + 0.05362647771835327, + 0.12265043705701828, + 0.08762140572071075, + 0.20292788743972778, + 0.06563209742307663 + ], + "credit_map": [ + 0.026517551392316818, + 0.03654297813773155, + 0.017314208671450615, + 0.02015865407884121, + 0.01750897243618965, + 0.009026512503623962, + 0.019885020330548286, + 0.015068300068378448, + 0.01520825270563364, + 0.023967541754245758, + 0.010813077911734581, + 0.015124538913369179, + 0.007873708382248878, + 0.004428245592862368, + 0.011073973961174488, + 0.007993638515472412, + 0.014921272173523903, + 0.022571362555027008, + 0.009138962253928185, + 0.012694907374680042, + 0.007444060407578945, + 0.0045776464976370335, + 0.009700586088001728, + 0.007892982102930546, + 0.01217836793512106, + 0.018755970522761345, + 0.0075451284646987915, + 0.010908761993050575, + 0.0066999574191868305, + 0.003408995224162936, + 0.008329167030751705, + 0.006798091810196638, + 0.007030019536614418, + 0.005148864816874266, + 0.0037964899092912674, + 0.009927707724273205, + 0.008010617457330227, + 0.0029109022580087185, + 0.00521023478358984, + 0.0034355013631284237, + 0.007342721335589886, + 0.010503565892577171, + 0.0057016341015696526, + 0.008137685246765614, + 0.0077745066955685616, + 0.002964518964290619, + 0.0069122593849897385, + 0.004289588890969753, + 0.017091086134314537, + 0.008070570416748524, + 0.00997073482722044, + 0.014026951044797897, + 0.01756821572780609, + 0.016018344089388847, + 0.02314896322786808, + 0.01675557903945446, + 0.009100240655243397, + 0.008189039304852486, + 0.009795223362743855, + 0.016875872388482094, + 0.010309874080121517, + 0.011359389871358871, + 0.012034726329147816, + 0.009957046248018742, + 0.017694702371954918, + 0.029829170554876328, + 0.017132751643657684, + 0.016675135120749474, + 0.026282301172614098, + 0.02772204391658306, + 0.047805771231651306, + 0.019786037504673004, + 0.009474516846239567, + 0.006932469550520182, + 0.007792026735842228, + 0.007305793929845095, + 0.009325875900685787, + 0.0055677094496786594, + 0.010802325792610645, + 0.008431385271251202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.008681767620146275 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1130073070526123 + ], + "opposite_logprob_tokens": [ + -2.2380073070526123 + ], + "logprob_reward": -0.1130073070526123, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1130073070526123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2325270026922226, + 0.07090707123279572, + 0.04207912087440491, + 0.027013925835490227, + 0.04821596294641495, + 0.061875082552433014, + 0.06875248998403549, + 0.10663830488920212, + 0.1613583117723465, + 0.18063265085220337 + ], + "credit_map": [ + 0.03199848532676697, + 0.034763310104608536, + 0.02673865109682083, + 0.040213990956544876, + 0.029809238389134407, + 0.011332041583955288, + 0.030807947739958763, + 0.026863358914852142, + 0.008666474372148514, + 0.01123699638992548, + 0.008769741281867027, + 0.011352194473147392, + 0.008894464932382107, + 0.004009070340543985, + 0.009953968226909637, + 0.008024170994758606, + 0.005565793719142675, + 0.008673468604683876, + 0.005552863702178001, + 0.006207823287695646, + 0.004315761383622885, + 0.002965715480968356, + 0.0046751839108765125, + 0.004122511018067598, + 0.0033555463887751102, + 0.004581016022711992, + 0.0028219071682542562, + 0.005005294922739267, + 0.00289778015576303, + 0.0015886491164565086, + 0.0035021298099309206, + 0.0032616043463349342, + 0.006432574708014727, + 0.0049263485707342625, + 0.007575778756290674, + 0.010171699337661266, + 0.005016673356294632, + 0.002963043749332428, + 0.005954498425126076, + 0.005175347439944744, + 0.007267641834914684, + 0.008660420775413513, + 0.007491381838917732, + 0.012498775497078896, + 0.005274707917124033, + 0.004751554224640131, + 0.009457257576286793, + 0.006473347079008818, + 0.014754265546798706, + 0.0077322437427937984, + 0.006713067181408405, + 0.0066175274550914764, + 0.010383524931967258, + 0.00651058042421937, + 0.008824153803288937, + 0.007217133417725563, + 0.01144791767001152, + 0.014521367847919464, + 0.00965405534952879, + 0.017938127741217613, + 0.010755660012364388, + 0.01412110310047865, + 0.015481342561542988, + 0.012718738988041878, + 0.019083114340901375, + 0.020434046164155006, + 0.013015980832278728, + 0.014492366462945938, + 0.022036710754036903, + 0.019991435110569, + 0.03934573382139206, + 0.012958932667970657, + 0.01740618608891964, + 0.01586826704442501, + 0.029390962794423103, + 0.03773462772369385, + 0.02196352742612362, + 0.016368653625249863, + 0.02312948741018772, + 0.018770961090922356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.1130073070526123 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036187181249260902 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036187181249260902, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036187181249260902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24661123752593994, + 0.11259768903255463, + 0.058756452053785324, + 0.0417519174516201, + 0.07278507947921753, + 0.045151904225349426, + 0.059040751308202744, + 0.07007583975791931, + 0.10663406550884247, + 0.18659506738185883 + ], + "credit_map": [ + 0.03465697914361954, + 0.03807449713349342, + 0.020896751433610916, + 0.04755231738090515, + 0.025276044383645058, + 0.02186509408056736, + 0.035737477242946625, + 0.022552073001861572, + 0.017187388613820076, + 0.020807724446058273, + 0.011818001978099346, + 0.021904945373535156, + 0.011559811420738697, + 0.007795916870236397, + 0.01181743573397398, + 0.00970646645873785, + 0.009675762616097927, + 0.012043679133057594, + 0.006709226872771978, + 0.010215301997959614, + 0.006494981236755848, + 0.0034210169687867165, + 0.005423463881015778, + 0.004773018881678581, + 0.006010808981955051, + 0.006209004670381546, + 0.005325146019458771, + 0.0066399965435266495, + 0.0038739657029509544, + 0.003082641400396824, + 0.005570033565163612, + 0.005040321033447981, + 0.010871282778680325, + 0.005676536355167627, + 0.007398602552711964, + 0.008741183206439018, + 0.01274615153670311, + 0.005091828294098377, + 0.013929830864071846, + 0.008329658769071102, + 0.0056398832239210606, + 0.008137202821671963, + 0.005052527412772179, + 0.00889388844370842, + 0.006189375184476376, + 0.002426235005259514, + 0.005245286971330643, + 0.0035675044637173414, + 0.01005111075937748, + 0.006787023972719908, + 0.004432706627994776, + 0.006999145727604628, + 0.00875320564955473, + 0.0070147402584552765, + 0.007826169021427631, + 0.007176643703132868, + 0.006338496692478657, + 0.0037230581510812044, + 0.006201290991157293, + 0.01094709150493145, + 0.007056954316794872, + 0.010895720683038235, + 0.011552869342267513, + 0.013360358774662018, + 0.011500109918415546, + 0.009041443467140198, + 0.00926840677857399, + 0.012471625581383705, + 0.01325939130038023, + 0.016530899330973625, + 0.02028377167880535, + 0.014278416521847248, + 0.018785880878567696, + 0.01656067743897438, + 0.031069783493876457, + 0.038575466722249985, + 0.025536732748150826, + 0.016202745959162712, + 0.02273055724799633, + 0.017133211717009544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0036187181249260902 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28134438395500183 + ], + "opposite_logprob_tokens": [ + -1.4063444137573242 + ], + "logprob_reward": -0.28134438395500183, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28134438395500183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09832736849784851, + 0.04836214333772659, + 0.04459404945373535, + 0.030234375968575478, + 0.08945027738809586, + 0.04022582620382309, + 0.043376147747039795, + 0.28067997097969055, + 0.2068834900856018, + 0.11786631494760513 + ], + "credit_map": [ + 0.012780288234353065, + 0.017547715455293655, + 0.008438145741820335, + 0.014374363236129284, + 0.012788519263267517, + 0.005525956396013498, + 0.015488699078559875, + 0.011383681558072567, + 0.005171594675630331, + 0.00793882180005312, + 0.005027634557336569, + 0.008411287330091, + 0.006153477821499109, + 0.0030245473608374596, + 0.007068069186061621, + 0.005566710140556097, + 0.004925490356981754, + 0.008488977327942848, + 0.0033788124565035105, + 0.006885457318276167, + 0.0062244730070233345, + 0.0030935073737055063, + 0.006748248357325792, + 0.004849080462008715, + 0.0032517886720597744, + 0.005195247940719128, + 0.0028295645024627447, + 0.005237277131527662, + 0.0040120938792824745, + 0.0022100270725786686, + 0.004123724531382322, + 0.003374652937054634, + 0.015264569781720638, + 0.022383926436305046, + 0.008896207436919212, + 0.012085202150046825, + 0.00962135847657919, + 0.005057492759078741, + 0.009525980800390244, + 0.006615540012717247, + 0.0039553348906338215, + 0.0032045471016317606, + 0.0036154938861727715, + 0.007370838429778814, + 0.004344266839325428, + 0.0037106687668710947, + 0.00800834409892559, + 0.006016330793499947, + 0.005509278737008572, + 0.004261595197021961, + 0.003520397702232003, + 0.005956710781902075, + 0.006325713358819485, + 0.006107302848249674, + 0.0060919104143977165, + 0.005603241268545389, + 0.026932591572403908, + 0.015535951592028141, + 0.023189585655927658, + 0.03026866354048252, + 0.025207020342350006, + 0.05592477694153786, + 0.0454121008515358, + 0.05820930376648903, + 0.018780263140797615, + 0.013992633670568466, + 0.01691249944269657, + 0.020175473764538765, + 0.026280583813786507, + 0.040370382368564606, + 0.03583117201924324, + 0.034540481865406036, + 0.011071843095123768, + 0.009740916080772877, + 0.01612929068505764, + 0.023303525522351265, + 0.016835104674100876, + 0.01090422086417675, + 0.016681864857673645, + 0.013199541717767715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.28134438395500183 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures an intimate and tense moment in a luxurious dressing room. A tall, statuesque woman with her sleek, raven-black hair pulled into a tight, high ponytail stands beside a shorter woman seated on a velvet stool. The tall woman, clad in an exquisite sleeveless black satin gown, exudes a calm authority in her posture. She is completely focused, one hand firmly and deliberately cupping the shorter woman's chin while the other carefully applies vibrant pale red lipstick to her lips. The shorter woman, with soft, wavy blonde hair and a simple cream silk blouse, looks up with a mixture of fear and resignation, her body tense. The background softly blurs, revealing a large vanity mirror framed by warm, glowing Hollywood-style bulbs, casting a soft golden light across the scene, creating delicate highlights on their skin and the glossy tube of lipstick. After completing the final stroke, the tall woman steps back, her expression shifting from concentration to calm scrutiny. She gave a curt, critical nod, then dismissed the smaller woman with a swift, decisive wave of her hand, a flick of her wrist that was both elegant and final, breaking eye contact as she shifted her focus, leaving behind a thick atmosphere of unspoken hierarchy and aloof control.", + "seed": 6958, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_2/index_5690_seed_6958_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08578865481540561, + "vlm_logprob": -0.08578865481540561, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.08578865481540561, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10162672400474548 + ], + "opposite_logprob_tokens": [ + -3.8516266345977783 + ], + "logprob_reward": -0.10162672400474548, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.10162672400474548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25978586077690125, + 0.1270848959684372, + 0.04721674695611, + 0.07377348095178604, + 0.0823727548122406, + 0.049818214029073715, + 0.052180152386426926, + 0.026790801435709, + 0.15085436403751373, + 0.13012275099754333 + ], + "credit_map": [ + 0.03359036147594452, + 0.025204414501786232, + 0.022010641172528267, + 0.03803592920303345, + 0.018344160169363022, + 0.02056242898106575, + 0.023925969377160072, + 0.07811195403337479, + 0.018780946731567383, + 0.01388948317617178, + 0.015023860149085522, + 0.014480776153504848, + 0.011809347197413445, + 0.03345431014895439, + 0.00592564744874835, + 0.013720527291297913, + 0.0073294732719659805, + 0.003958732355386019, + 0.006288464181125164, + 0.007040794938802719, + 0.002603864297270775, + 0.0033691676799207926, + 0.0037957828026264906, + 0.012830467894673347, + 0.014460297301411629, + 0.0045545934699475765, + 0.00682719936594367, + 0.00977414008229971, + 0.0035491830203682184, + 0.006705875974148512, + 0.005800457205623388, + 0.02210172824561596, + 0.012611539103090763, + 0.00804043747484684, + 0.009938941337168217, + 0.009038412012159824, + 0.00763919111341238, + 0.019879067316651344, + 0.004779439885169268, + 0.01044572051614523, + 0.007197516039013863, + 0.004699004348367453, + 0.005569584667682648, + 0.009580439887940884, + 0.0048815100453794, + 0.005206997040659189, + 0.004435756243765354, + 0.008247407153248787, + 0.007325286511331797, + 0.004776292946189642, + 0.0062411073595285416, + 0.007796429097652435, + 0.00433973828330636, + 0.014235110953450203, + 0.002985331928357482, + 0.004480855539441109, + 0.004153743386268616, + 0.002518378896638751, + 0.005289338994771242, + 0.005829040892422199, + 0.0026899841614067554, + 0.0021374565549194813, + 0.001976927975192666, + 0.002195928944274783, + 0.016348926350474358, + 0.019746888428926468, + 0.020986415445804596, + 0.023068061098456383, + 0.018204595893621445, + 0.012825480662286282, + 0.02667185291647911, + 0.013002142310142517, + 0.013220814056694508, + 0.01908162608742714, + 0.016161730512976646, + 0.014711765572428703, + 0.010123474523425102, + 0.016262637451291084, + 0.019607851281762123, + 0.0209528598934412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.10162672400474548 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005982468836009502 + ], + "opposite_logprob_tokens": [ + -5.130982398986816 + ], + "logprob_reward": -0.005982468836009502, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005982468836009502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36604639887809753, + 0.1349440962076187, + 0.08031431585550308, + 0.035747870802879333, + 0.017719844356179237, + 0.03560641407966614, + 0.04200747609138489, + 0.06974010169506073, + 0.11801560968160629, + 0.09985783696174622 + ], + "credit_map": [ + 0.052944328635931015, + 0.030069878324866295, + 0.042580172419548035, + 0.0578966960310936, + 0.015711026266217232, + 0.035110682249069214, + 0.025795280933380127, + 0.10593835264444351, + 0.018053801730275154, + 0.014566918835043907, + 0.020419372245669365, + 0.01982179470360279, + 0.009546998888254166, + 0.027294259518384933, + 0.005513587035238743, + 0.019727373495697975, + 0.014296799898147583, + 0.005454345140606165, + 0.010037022642791271, + 0.009785124100744724, + 0.0032453506719321012, + 0.009768580086529255, + 0.00534633407369256, + 0.02238076739013195, + 0.00582482386380434, + 0.005519304424524307, + 0.005437842104583979, + 0.005938876420259476, + 0.002115521812811494, + 0.0027154944837093353, + 0.002276040380820632, + 0.005919972434639931, + 0.002033424796536565, + 0.0023332503624260426, + 0.0021297757048159838, + 0.0027793939225375652, + 0.0014057343360036612, + 0.0036799455992877483, + 0.0014008787693455815, + 0.00195744214579463, + 0.0035245211329311132, + 0.004033133387565613, + 0.004108418710529804, + 0.006363857537508011, + 0.0028035109862685204, + 0.004975349176675081, + 0.0036502189468592405, + 0.006147404666990042, + 0.005502632353454828, + 0.0037753242067992687, + 0.004271484445780516, + 0.005265147425234318, + 0.003657235996797681, + 0.011280762031674385, + 0.004477627109736204, + 0.003777266712859273, + 0.011001487262547016, + 0.005877120885998011, + 0.017523329704999924, + 0.01501611340790987, + 0.005609235260635614, + 0.004503251519054174, + 0.004245666787028313, + 0.005963902920484543, + 0.015073686838150024, + 0.013708865270018578, + 0.01526745967566967, + 0.011524512432515621, + 0.014938367530703545, + 0.016696961596608162, + 0.014300496317446232, + 0.016505273059010506, + 0.008846113458275795, + 0.012974979355931282, + 0.00836797896772623, + 0.015289565548300743, + 0.01037090364843607, + 0.012400160543620586, + 0.018079448491334915, + 0.013528691604733467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.005982468836009502 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20165812969207764 + ], + "opposite_logprob_tokens": [ + -1.7016581296920776 + ], + "logprob_reward": -0.20165812969207764, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20165812969207764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12967681884765625, + 0.09527841210365295, + 0.08241279423236847, + 0.10259678214788437, + 0.10644376277923584, + 0.037766288965940475, + 0.04601925611495972, + 0.03414928913116455, + 0.22450895607471466, + 0.14114759862422943 + ], + "credit_map": [ + 0.01753709465265274, + 0.012456811033189297, + 0.02185673825442791, + 0.021035121753811836, + 0.011271493509411812, + 0.019297750666737556, + 0.009357611648738384, + 0.01686420477926731, + 0.00904075801372528, + 0.00829018373042345, + 0.010640128515660763, + 0.018137585371732712, + 0.007374523673206568, + 0.016852889209985733, + 0.00714477151632309, + 0.01779758557677269, + 0.010826745070517063, + 0.008049664087593555, + 0.01045315619558096, + 0.01146399974822998, + 0.006277460604906082, + 0.013580435886979103, + 0.006196063477545977, + 0.015565270557999611, + 0.018207404762506485, + 0.006309371907263994, + 0.009457787498831749, + 0.01338830403983593, + 0.0051751104183495045, + 0.009314440190792084, + 0.00828353688120842, + 0.032460834830999374, + 0.017536798492074013, + 0.00908841472119093, + 0.011383041739463806, + 0.013764250092208385, + 0.006336791906505823, + 0.02050034888088703, + 0.005161951296031475, + 0.022672176361083984, + 0.005004082340747118, + 0.004442091565579176, + 0.007009885739535093, + 0.007636389695107937, + 0.0035127021837979555, + 0.0026079078670591116, + 0.00286057498306036, + 0.004692660178989172, + 0.005676954984664917, + 0.004557973705232143, + 0.005156165920197964, + 0.007232369855046272, + 0.004095012322068214, + 0.010814273729920387, + 0.004105375614017248, + 0.004381136037409306, + 0.004277566447854042, + 0.0025471330154687166, + 0.008086728863418102, + 0.00724615715444088, + 0.0029441167134791613, + 0.00202389620244503, + 0.002577918814495206, + 0.004445776809006929, + 0.01998572051525116, + 0.03514797240495682, + 0.022817598655819893, + 0.026529546827077866, + 0.023978855460882187, + 0.019028129056096077, + 0.03979257494211197, + 0.03722858056426048, + 0.00993887335062027, + 0.017424926161766052, + 0.015171600505709648, + 0.012571527622640133, + 0.012531635351479053, + 0.027659986168146133, + 0.02272915467619896, + 0.023119904100894928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.20165812969207764 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735833827406168 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735833827406168, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735833827406168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2732311487197876, + 0.09749335050582886, + 0.196126326918602, + 0.035975925624370575, + 0.011476791463792324, + 0.02020011842250824, + 0.011730375699698925, + 0.01280492078512907, + 0.15533725917339325, + 0.18562377989292145 + ], + "credit_map": [ + 0.04087677597999573, + 0.020428869873285294, + 0.02371586672961712, + 0.04079017788171768, + 0.012118987739086151, + 0.034832991659641266, + 0.020557111129164696, + 0.07991036772727966, + 0.011834296397864819, + 0.011004932224750519, + 0.011989438906311989, + 0.015680471435189247, + 0.008758501149713993, + 0.021564031019806862, + 0.005581713747233152, + 0.011079970747232437, + 0.029070667922496796, + 0.016761837527155876, + 0.022979218512773514, + 0.025704912841320038, + 0.013079458847641945, + 0.026719681918621063, + 0.01571657322347164, + 0.046093977987766266, + 0.006053609307855368, + 0.003331852611154318, + 0.003621414303779602, + 0.004841947462409735, + 0.002320870291441679, + 0.004163502249866724, + 0.003022199496626854, + 0.008620532229542732, + 0.001515137031674385, + 0.0019208242883905768, + 0.0012860952410846949, + 0.002098825527355075, + 0.0008577699190936983, + 0.0012401051353663206, + 0.0008910303586162627, + 0.0016670043114572763, + 0.0028616557829082012, + 0.0029292262624949217, + 0.003207253059372306, + 0.004934451077133417, + 0.0010883439099416137, + 0.0015227744588628411, + 0.0016640688991174102, + 0.0019923450890928507, + 0.0016755611868575215, + 0.0011661323951557279, + 0.001513043069280684, + 0.002250467427074909, + 0.0011501468252390623, + 0.001626812038011849, + 0.0012603314826264977, + 0.0010878804605454206, + 0.002001075306907296, + 0.0012393802171573043, + 0.001486907247453928, + 0.0025867708027362823, + 0.001421321532689035, + 0.0012845159508287907, + 0.0012779835378751159, + 0.0015069660730659962, + 0.02318619005382061, + 0.0202453825622797, + 0.03012712672352791, + 0.02134833298623562, + 0.011479133740067482, + 0.01706722006201744, + 0.013445977121591568, + 0.018437905237078667, + 0.01738722249865532, + 0.021989941596984863, + 0.01763959787786007, + 0.02365661785006523, + 0.016960278153419495, + 0.02891600877046585, + 0.028585925698280334, + 0.03048819862306118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.006735833827406168 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294011771678925 + ], + "opposite_logprob_tokens": [ + -2.2379400730133057 + ], + "logprob_reward": -0.11294011771678925, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294011771678925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14468999207019806, + 0.06221437454223633, + 0.1025230884552002, + 0.052107878029346466, + 0.03149379789829254, + 0.023421941325068474, + 0.026641743257641792, + 0.03483985736966133, + 0.3587811589241028, + 0.16328617930412292 + ], + "credit_map": [ + 0.02329101786017418, + 0.016518227756023407, + 0.01845063641667366, + 0.019064845517277718, + 0.009766903705894947, + 0.02982984483242035, + 0.010288731195032597, + 0.01747978664934635, + 0.008150323294103146, + 0.008324935100972652, + 0.011125529184937477, + 0.011050869710743427, + 0.006935327313840389, + 0.006300217006355524, + 0.004444103222340345, + 0.0058830734342336655, + 0.016921984031796455, + 0.007379467599093914, + 0.012120435945689678, + 0.0128641277551651, + 0.0060316347517073154, + 0.011103122495114803, + 0.008291265927255154, + 0.027811048552393913, + 0.008662323467433453, + 0.006001900415867567, + 0.007448818068951368, + 0.007973700761795044, + 0.003584243357181549, + 0.006271912716329098, + 0.0032690197695046663, + 0.008895961567759514, + 0.0037302246782928705, + 0.0033173745032399893, + 0.004850299097597599, + 0.004510039463639259, + 0.002230763668194413, + 0.005635492969304323, + 0.0020468735601752996, + 0.0051727294921875, + 0.0043743145652115345, + 0.0017125960439443588, + 0.0025745253078639507, + 0.0031370746437460184, + 0.002491688821464777, + 0.002150119049474597, + 0.0025356891565024853, + 0.004445933271199465, + 0.003432769328355789, + 0.002771281637251377, + 0.0028242669068276882, + 0.0038130716420710087, + 0.0030523641034960747, + 0.005642039235681295, + 0.0026528669986873865, + 0.002453083172440529, + 0.005034780595451593, + 0.002946155145764351, + 0.005274490918964148, + 0.006439642980694771, + 0.003495304612442851, + 0.003998364321887493, + 0.002894340083003044, + 0.004756777081638575, + 0.04295491427183151, + 0.0418219156563282, + 0.03355446830391884, + 0.03799094259738922, + 0.04485265538096428, + 0.04045342281460762, + 0.057188451290130615, + 0.0599643848836422, + 0.015355541370809078, + 0.03710128739476204, + 0.018523871898651123, + 0.03657376393675804, + 0.012690181843936443, + 0.013727635145187378, + 0.01755541004240513, + 0.011758471839129925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.11294011771678925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures an intimate and tense moment in a luxurious dressing room. A tall, statuesque woman with her sleek, raven-black hair pulled into a tight, high ponytail stands beside a shorter woman seated on a velvet stool. The tall woman, clad in an exquisite sleeveless black satin gown, exudes a calm authority in her posture. She is completely focused, one hand firmly and deliberately cupping the shorter woman's chin while the other carefully applies vibrant pale red lipstick to her lips. The shorter woman, with soft, wavy blonde hair and a simple cream silk blouse, looks up with a mixture of fear and resignation, her body tense. The background softly blurs, revealing a large vanity mirror framed by warm, glowing Hollywood-style bulbs, casting a soft golden light across the scene, creating delicate highlights on their skin and the glossy tube of lipstick. After completing the final stroke, the tall woman steps back, her expression shifting from concentration to calm scrutiny. She gave a curt, critical nod, then dismissed the smaller woman with a swift, decisive wave of her hand, a flick of her wrist that was both elegant and final, breaking eye contact as she shifted her focus, leaving behind a thick atmosphere of unspoken hierarchy and aloof control.", + "seed": 6959, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_2/index_5690_seed_6959_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10293727638199926, + "vlm_logprob": -0.10293727638199926, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.10293727638199926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09897199273109436 + ], + "opposite_logprob_tokens": [ + -3.9739720821380615 + ], + "logprob_reward": -0.09897199273109436, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.09897199273109436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15483458340168, + 0.05484079569578171, + 0.053369440138339996, + 0.027089154347777367, + 0.03707149624824524, + 0.06086902692914009, + 0.12071140110492706, + 0.13399703800678253, + 0.10062892735004425, + 0.2565881609916687 + ], + "credit_map": [ + 0.01972813531756401, + 0.018141310662031174, + 0.024279886856675148, + 0.02489892952144146, + 0.025595320388674736, + 0.013706808909773827, + 0.011733141727745533, + 0.016751037910580635, + 0.006648613605648279, + 0.0062606800347566605, + 0.007941444404423237, + 0.01153124775737524, + 0.007468317169696093, + 0.003099574241787195, + 0.006888880860060453, + 0.005002032034099102, + 0.0068205478601157665, + 0.005696922540664673, + 0.006435107905417681, + 0.0075141871348023415, + 0.012963884510099888, + 0.0038257844280451536, + 0.00650439690798521, + 0.00360860419459641, + 0.004230922088027, + 0.0028491036500781775, + 0.0042566764168441296, + 0.004865565337240696, + 0.004197866190224886, + 0.001988731324672699, + 0.002332984237000346, + 0.002367303241044283, + 0.004320407286286354, + 0.004967641085386276, + 0.0050937519408762455, + 0.007074674591422081, + 0.005263115279376507, + 0.0024879048578441143, + 0.0027074653189629316, + 0.005156531929969788, + 0.007671467959880829, + 0.007700762245804071, + 0.006661379709839821, + 0.007808339782059193, + 0.015444540418684483, + 0.004796168766915798, + 0.005868688691407442, + 0.004917667247354984, + 0.011943900026381016, + 0.011662776581943035, + 0.015228104777634144, + 0.01944759674370289, + 0.01867777854204178, + 0.0049461135640740395, + 0.02937406487762928, + 0.009431058540940285, + 0.020957335829734802, + 0.015982914716005325, + 0.026477232575416565, + 0.020945308730006218, + 0.01590459607541561, + 0.010768303647637367, + 0.007695963606238365, + 0.015265370719134808, + 0.0164681114256382, + 0.011026537977159023, + 0.017559394240379333, + 0.01734587922692299, + 0.013401431031525135, + 0.008805396035313606, + 0.0063069891184568405, + 0.009715171530842781, + 0.027090994641184807, + 0.03431202471256256, + 0.039788927882909775, + 0.043707475066185, + 0.041218195110559464, + 0.022626584395766258, + 0.03151574730873108, + 0.01632821373641491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.09897199273109436 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016117870807647705 + ], + "opposite_logprob_tokens": [ + -4.141118049621582 + ], + "logprob_reward": -0.016117870807647705, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016117870807647705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1943659782409668, + 0.07407157868146896, + 0.040319979190826416, + 0.0348038487136364, + 0.022843796759843826, + 0.03203887119889259, + 0.04704248532652855, + 0.18142203986644745, + 0.08446896076202393, + 0.2886224687099457 + ], + "credit_map": [ + 0.02125926874577999, + 0.020787570625543594, + 0.03427784889936447, + 0.04023988917469978, + 0.019043993204832077, + 0.008843121118843555, + 0.03192037343978882, + 0.017993921414017677, + 0.008960600942373276, + 0.00738334096968174, + 0.012701427564024925, + 0.013709422200918198, + 0.012154129333794117, + 0.005438648629933596, + 0.006085688713937998, + 0.007638325449079275, + 0.005515703931450844, + 0.004296754486858845, + 0.0067345560528337955, + 0.0073645454831421375, + 0.004576361272484064, + 0.0029825849924236536, + 0.003682543756440282, + 0.005166930612176657, + 0.004477303940802813, + 0.0035078790970146656, + 0.0037292176857590675, + 0.004549878649413586, + 0.007103532552719116, + 0.00242166337557137, + 0.003778391284868121, + 0.00523598026484251, + 0.0028802624437958, + 0.002480751136317849, + 0.003759213024750352, + 0.0034104431979358196, + 0.0037328898906707764, + 0.0017921923426911235, + 0.0020659789443016052, + 0.0027220663614571095, + 0.0038335626013576984, + 0.003464493900537491, + 0.0038090753369033337, + 0.00475785369053483, + 0.007130519021302462, + 0.002760874107480049, + 0.003140411339700222, + 0.0031420846935361624, + 0.004089619033038616, + 0.005829041823744774, + 0.006713148672133684, + 0.0060657719150185585, + 0.011095186695456505, + 0.0037528472021222115, + 0.006182608660310507, + 0.003314261557534337, + 0.024601779878139496, + 0.021265046671032906, + 0.03152674436569214, + 0.037228457629680634, + 0.02394743263721466, + 0.01192136574536562, + 0.016082793474197388, + 0.014848423190414906, + 0.01257421262562275, + 0.01111630443483591, + 0.013211943209171295, + 0.018653608858585358, + 0.010334247723221779, + 0.006559385452419519, + 0.006749853026121855, + 0.005269404500722885, + 0.02927723526954651, + 0.03710862621665001, + 0.03212990239262581, + 0.039642687886953354, + 0.05503227189183235, + 0.032419171184301376, + 0.042083483189344406, + 0.020929105579853058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.016117870807647705 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1130046397447586 + ], + "opposite_logprob_tokens": [ + -2.238004684448242 + ], + "logprob_reward": -0.1130046397447586, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1130046397447586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17365553975105286, + 0.09534841030836105, + 0.07520276308059692, + 0.04046923294663429, + 0.04776538908481598, + 0.04724528267979622, + 0.1296049803495407, + 0.07495758682489395, + 0.11879058182239532, + 0.19696028530597687 + ], + "credit_map": [ + 0.0221466813236475, + 0.01784437894821167, + 0.036396708339452744, + 0.03636520728468895, + 0.014706130139529705, + 0.009820523671805859, + 0.016253381967544556, + 0.02012251317501068, + 0.012263917364180088, + 0.009833603166043758, + 0.012992659583687782, + 0.015958165749907494, + 0.019270550459623337, + 0.0082713533192873, + 0.00894737895578146, + 0.0078107817098498344, + 0.009484675712883472, + 0.008017194457352161, + 0.00905437208712101, + 0.008140108548104763, + 0.020447349175810814, + 0.006953836884349585, + 0.007786381058394909, + 0.005318835377693176, + 0.005108283367007971, + 0.004265917930752039, + 0.005391173996031284, + 0.005574141163378954, + 0.006216195411980152, + 0.0033277475740760565, + 0.004654761403799057, + 0.0059310090728104115, + 0.005655845161527395, + 0.005363565869629383, + 0.007953159511089325, + 0.009079835377633572, + 0.008178254589438438, + 0.0034481778275221586, + 0.0034809759818017483, + 0.004605570342391729, + 0.006329637952148914, + 0.005105708260089159, + 0.0076601700857281685, + 0.009384474717080593, + 0.007338270545005798, + 0.0031201420351862907, + 0.004141391720622778, + 0.004165485501289368, + 0.01345266867429018, + 0.016749920323491096, + 0.013307303190231323, + 0.012988032773137093, + 0.03321162983775139, + 0.011994038708508015, + 0.01870141737163067, + 0.009199964813888073, + 0.0096467025578022, + 0.008582750335335732, + 0.011261695995926857, + 0.01077658124268055, + 0.013595101423561573, + 0.0069722519256174564, + 0.006195566151291132, + 0.007926935330033302, + 0.014003957621753216, + 0.014133173041045666, + 0.019487500190734863, + 0.0244622603058815, + 0.014576845802366734, + 0.012206475250422955, + 0.008969461545348167, + 0.010950900614261627, + 0.01894979737699032, + 0.026059705764055252, + 0.020410116761922836, + 0.02635248564183712, + 0.04061964154243469, + 0.020527420565485954, + 0.029829461127519608, + 0.01421163510531187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.1130046397447586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256047938019037 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256047938019037, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256047938019037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19831977784633636, + 0.12222305685281754, + 0.06425229460000992, + 0.07010740041732788, + 0.09027262032032013, + 0.036755070090293884, + 0.13171522319316864, + 0.08432512730360031, + 0.13047699630260468, + 0.07155243307352066 + ], + "credit_map": [ + 0.025454072281718254, + 0.018812911584973335, + 0.03127680718898773, + 0.04011613130569458, + 0.03541213274002075, + 0.015153622254729271, + 0.01574859954416752, + 0.016345487907528877, + 0.017378859221935272, + 0.013672549277544022, + 0.016520364210009575, + 0.018900161609053612, + 0.02948419377207756, + 0.009365670382976532, + 0.010644610971212387, + 0.006256652530282736, + 0.00876695103943348, + 0.006754241418093443, + 0.010793227702379227, + 0.010363822802901268, + 0.013196796178817749, + 0.004220747388899326, + 0.005699490662664175, + 0.004457012750208378, + 0.007603705395013094, + 0.008267106488347054, + 0.006913332734256983, + 0.0056915548630058765, + 0.019969170913100243, + 0.006699417717754841, + 0.007937970571219921, + 0.007025144062936306, + 0.011780967935919762, + 0.010595628060400486, + 0.011373358778655529, + 0.010362882167100906, + 0.024950657039880753, + 0.007463684771209955, + 0.008556819520890713, + 0.005188623443245888, + 0.004495907574892044, + 0.0040693809278309345, + 0.005566863808780909, + 0.005761449690908194, + 0.008183476515114307, + 0.0026564868167042732, + 0.0031506901141256094, + 0.0028708158060908318, + 0.014047907665371895, + 0.015308544971048832, + 0.01005180086940527, + 0.017422392964363098, + 0.022117657586932182, + 0.0063535054214298725, + 0.0354757159948349, + 0.01093770656734705, + 0.010351425036787987, + 0.012100816704332829, + 0.013666054233908653, + 0.012434199452400208, + 0.012615167535841465, + 0.006792367435991764, + 0.006075367797166109, + 0.010289729572832584, + 0.020811472088098526, + 0.015477391891181469, + 0.02150730974972248, + 0.0243382528424263, + 0.01847122237086296, + 0.007951238192617893, + 0.012833451852202415, + 0.009086662903428078, + 0.008063570596277714, + 0.009969033300876617, + 0.011927323415875435, + 0.01269206777215004, + 0.009923246689140797, + 0.008806174620985985, + 0.00578220235183835, + 0.004388812463730574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.005256047938019037 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28133583068847656 + ], + "opposite_logprob_tokens": [ + -1.4063358306884766 + ], + "logprob_reward": -0.28133583068847656, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28133583068847656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1106514260172844, + 0.06314046680927277, + 0.020054880529642105, + 0.01858721859753132, + 0.02612629346549511, + 0.022341087460517883, + 0.07964562624692917, + 0.051948562264442444, + 0.0594111792743206, + 0.5480932593345642 + ], + "credit_map": [ + 0.011050084605813026, + 0.01162892859429121, + 0.020189547911286354, + 0.022755960002541542, + 0.014557864516973495, + 0.00872484315186739, + 0.009643075987696648, + 0.012101122178137302, + 0.007757175713777542, + 0.007718721404671669, + 0.008224720135331154, + 0.0068150050938129425, + 0.01588144712150097, + 0.005251842550933361, + 0.005877253599464893, + 0.00561430212110281, + 0.0019909830298274755, + 0.001980474218726158, + 0.003866164479404688, + 0.003454594872891903, + 0.0019853441044688225, + 0.0017772023566067219, + 0.0019068936817348003, + 0.0030932240188121796, + 0.002681358950212598, + 0.0018954456318169832, + 0.0026023718528449535, + 0.002774581778794527, + 0.0023372075520455837, + 0.0016445850487798452, + 0.0019903306383639574, + 0.002661336911842227, + 0.0029620779678225517, + 0.003051960375159979, + 0.0034485608339309692, + 0.004688566084951162, + 0.005270711146295071, + 0.0021219798363745213, + 0.002319874009117484, + 0.0022625629790127277, + 0.002678417367860675, + 0.002418085467070341, + 0.003959150053560734, + 0.004676465876400471, + 0.0028058115858584642, + 0.0014065132709220052, + 0.0021755965426564217, + 0.0022210462484508753, + 0.008505533449351788, + 0.0073197148740291595, + 0.01130833849310875, + 0.015472776256501675, + 0.009576916694641113, + 0.004961871076375246, + 0.015813661739230156, + 0.006686815991997719, + 0.005450444296002388, + 0.005997851025313139, + 0.008386489003896713, + 0.010418671183288097, + 0.0069979168474674225, + 0.004723796155303717, + 0.004148093052208424, + 0.005825301632285118, + 0.00808334443718195, + 0.008598806336522102, + 0.011129309423267841, + 0.015100259333848953, + 0.004730398301035166, + 0.00371822458691895, + 0.0033841757103800774, + 0.004666660912334919, + 0.054755356162786484, + 0.07283288240432739, + 0.056741926819086075, + 0.07055599242448807, + 0.1103016585111618, + 0.057524725794792175, + 0.08374956995248795, + 0.04163114354014397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.28133583068847656 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures an intimate and tense moment in a luxurious dressing room. A tall, statuesque woman with her sleek, raven-black hair pulled into a tight, high ponytail stands beside a shorter woman seated on a velvet stool. The tall woman, clad in an exquisite sleeveless black satin gown, exudes a calm authority in her posture. She is completely focused, one hand firmly and deliberately cupping the shorter woman's chin while the other carefully applies vibrant pale red lipstick to her lips. The shorter woman, with soft, wavy blonde hair and a simple cream silk blouse, looks up with a mixture of fear and resignation, her body tense. The background softly blurs, revealing a large vanity mirror framed by warm, glowing Hollywood-style bulbs, casting a soft golden light across the scene, creating delicate highlights on their skin and the glossy tube of lipstick. After completing the final stroke, the tall woman steps back, her expression shifting from concentration to calm scrutiny. She gave a curt, critical nod, then dismissed the smaller woman with a swift, decisive wave of her hand, a flick of her wrist that was both elegant and final, breaking eye contact as she shifted her focus, leaving behind a thick atmosphere of unspoken hierarchy and aloof control.", + "seed": 6960, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_2/index_5690_seed_6960_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.139997481694445, + "vlm_logprob": -0.139997481694445, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.139997481694445, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12551985681056976 + ], + "opposite_logprob_tokens": [ + -3.7505197525024414 + ], + "logprob_reward": -0.12551985681056976, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.12551985681056976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02835611253976822, + 0.012936791405081749, + 0.0110478475689888, + 0.02268126979470253, + 0.05472598597407341, + 0.024703919887542725, + 0.34704965353012085, + 0.176841601729393, + 0.1335693597793579, + 0.18808741867542267 + ], + "credit_map": [ + 0.002382156439125538, + 0.0040939319878816605, + 0.0017169435741379857, + 0.002826047595590353, + 0.004418864380568266, + 0.004474442917853594, + 0.004216628614813089, + 0.004227096680551767, + 0.0011323353974148631, + 0.0015999653842300177, + 0.0008617430576123297, + 0.0012039844878017902, + 0.001857543713413179, + 0.0019872223492711782, + 0.0016498792683705688, + 0.0026441183872520924, + 0.0012448737397789955, + 0.001612666412256658, + 0.000830660224892199, + 0.0009667460108175874, + 0.001182554755359888, + 0.0017459391383454204, + 0.0016766681801527739, + 0.0017877393402159214, + 0.0016214114148169756, + 0.0024981247261166573, + 0.0012192499125376344, + 0.0016259201802313328, + 0.004215122200548649, + 0.0032675168476998806, + 0.0024928238708525896, + 0.005741101689636707, + 0.0038332371041178703, + 0.006234224420040846, + 0.003438054583966732, + 0.004580019507557154, + 0.012543577700853348, + 0.007540778256952763, + 0.007460277993232012, + 0.0090958122164011, + 0.001441707368940115, + 0.0022327688056975603, + 0.001081149559468031, + 0.0016986094415187836, + 0.004388055298477411, + 0.0025623696856200695, + 0.0028384802863001823, + 0.008460777811706066, + 0.014127682894468307, + 0.020031338557600975, + 0.010957928374409676, + 0.01811351254582405, + 0.06546220183372498, + 0.048272185027599335, + 0.023552432656288147, + 0.14653238654136658, + 0.006929059512913227, + 0.008082514628767967, + 0.003630890743806958, + 0.006788964848965406, + 0.06166110932826996, + 0.06345393508672714, + 0.010872376151382923, + 0.015422750264406204, + 0.005957372486591339, + 0.006821316666901112, + 0.0049476660788059235, + 0.009761804714798927, + 0.004960028454661369, + 0.020384522154927254, + 0.06362594664096832, + 0.01711069792509079, + 0.012563511729240417, + 0.06438751518726349, + 0.006843480281531811, + 0.014091960154473782, + 0.010908988304436207, + 0.0526374876499176, + 0.008483759127557278, + 0.01817071996629238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.12551985681056976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02641504630446434 + ], + "opposite_logprob_tokens": [ + -3.6514151096343994 + ], + "logprob_reward": -0.02641504630446434, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02641504630446434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052341196686029434, + 0.019829900935292244, + 0.023806802928447723, + 0.015299721620976925, + 0.062344297766685486, + 0.09295569360256195, + 0.15504391491413116, + 0.39451196789741516, + 0.04418887197971344, + 0.13967768847942352 + ], + "credit_map": [ + 0.004307738970965147, + 0.008877052925527096, + 0.003923534881323576, + 0.005620890762656927, + 0.003964994568377733, + 0.005928145721554756, + 0.008281711488962173, + 0.0114371282979846, + 0.0018626580713316798, + 0.002444045152515173, + 0.0022257817909121513, + 0.0023122881539165974, + 0.002087125787511468, + 0.0025593782775104046, + 0.0032642330043017864, + 0.0030743908137083054, + 0.0014731292612850666, + 0.0027281937655061483, + 0.0014288602396845818, + 0.0022568644490092993, + 0.0025973476003855467, + 0.0035799855832010508, + 0.0033365024719387293, + 0.006405921187251806, + 0.0013017919845879078, + 0.0021471877116709948, + 0.0012075239792466164, + 0.0015947551000863314, + 0.0022074533626437187, + 0.0022693064529448748, + 0.0025642861146479845, + 0.0020074164494872093, + 0.004399229306727648, + 0.00730896508321166, + 0.003976866137236357, + 0.005326554179191589, + 0.013667354360222816, + 0.008741241879761219, + 0.008561838418245316, + 0.010362247005105019, + 0.004786791745573282, + 0.007119202520698309, + 0.003013373352587223, + 0.0046792724169790745, + 0.02232775278389454, + 0.008532602339982986, + 0.008774195797741413, + 0.03372250124812126, + 0.005053137894719839, + 0.00949642900377512, + 0.0047516534104943275, + 0.007942561991512775, + 0.050154320895671844, + 0.028444388881325722, + 0.011424219235777855, + 0.03777719661593437, + 0.014461244456470013, + 0.01803913526237011, + 0.00842659454792738, + 0.014872487634420395, + 0.14461693167686462, + 0.13766689598560333, + 0.02584799751639366, + 0.03058064728975296, + 0.004120208323001862, + 0.00928737223148346, + 0.004498703870922327, + 0.0049921548925340176, + 0.002492849249392748, + 0.0068614287301898, + 0.007337216287851334, + 0.0045989383943378925, + 0.009396177716553211, + 0.051759298890829086, + 0.005674743093550205, + 0.009991602040827274, + 0.006345293950289488, + 0.02969856932759285, + 0.008974742144346237, + 0.01783725805580616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.02641504630446434 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008416950702667 + ], + "opposite_logprob_tokens": [ + -2.6950840950012207 + ], + "logprob_reward": -0.07008416950702667, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008416950702667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01752280630171299, + 0.007472584955394268, + 0.01342442911118269, + 0.008121587336063385, + 0.01787756010890007, + 0.013207138516008854, + 0.47403189539909363, + 0.20531970262527466, + 0.055662304162979126, + 0.1873600035905838 + ], + "credit_map": [ + 0.0013913834700360894, + 0.0022065541706979275, + 0.0014021960087120533, + 0.0015601353952661157, + 0.0019922186620533466, + 0.0029538318049162626, + 0.003369266167283058, + 0.002647220389917493, + 0.0006905849440954626, + 0.0009319609380327165, + 0.0005567874759435654, + 0.0008735595620237291, + 0.0006202042568475008, + 0.0012790110195055604, + 0.0014819029020145535, + 0.0010385735658928752, + 0.0008806195692159235, + 0.0011913625057786703, + 0.0005093404906801879, + 0.0008865318959578872, + 0.0016748766647651792, + 0.00263842917047441, + 0.0017945144791156054, + 0.003848754335194826, + 0.0006395679665729403, + 0.0009515788406133652, + 0.0005676173605024815, + 0.0007102388190105557, + 0.0008461708202958107, + 0.002007704693824053, + 0.0012391890631988645, + 0.0011595190735533834, + 0.0013925754465162754, + 0.003013811307027936, + 0.0013277768157422543, + 0.001884883502498269, + 0.002741775941103697, + 0.0025973054580390453, + 0.0022717518731951714, + 0.0026476799976080656, + 0.0008599346620030701, + 0.0014829119900241494, + 0.0009190773707814515, + 0.0014233705587685108, + 0.0014013366308063269, + 0.0012972576078027487, + 0.002375276992097497, + 0.0034479720052331686, + 0.020352713763713837, + 0.02457897923886776, + 0.014640074223279953, + 0.023315487429499626, + 0.0444830097258091, + 0.10936375707387924, + 0.031926654279232025, + 0.2053711861371994, + 0.007192561868578196, + 0.008773650974035263, + 0.005010189022868872, + 0.007338414900004864, + 0.0645807608962059, + 0.07659351080656052, + 0.012169646099209785, + 0.02366098016500473, + 0.002540944842621684, + 0.003411229234188795, + 0.0033862905111163855, + 0.0035067484714090824, + 0.002310883952304721, + 0.006588011980056763, + 0.02607123740017414, + 0.007846955209970474, + 0.011724403128027916, + 0.07747569680213928, + 0.00615283427760005, + 0.015912814065814018, + 0.007542728446424007, + 0.031289540231227875, + 0.010422445833683014, + 0.026839539408683777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.07008416950702667 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036188371013849974 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036188371013849974, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036188371013849974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06042785570025444, + 0.023801028728485107, + 0.017221443355083466, + 0.03529255837202072, + 0.05343509837985039, + 0.15313947200775146, + 0.35323670506477356, + 0.16955913603305817, + 0.09674317389726639, + 0.0371435210108757 + ], + "credit_map": [ + 0.005382217466831207, + 0.009303649887442589, + 0.0032419534400105476, + 0.005687570199370384, + 0.0073602525517344475, + 0.008852130733430386, + 0.006517124827951193, + 0.014082960784435272, + 0.0025880811735987663, + 0.002704185899347067, + 0.0019493793370202184, + 0.002496708184480667, + 0.003662049537524581, + 0.003010367276147008, + 0.004703621380031109, + 0.0026866341941058636, + 0.0017655857373028994, + 0.0019306103931739926, + 0.001529029686935246, + 0.0018718903884291649, + 0.0024001123383641243, + 0.0024612066335976124, + 0.0028660639654845, + 0.0023969439789652824, + 0.001922696945257485, + 0.003842479782178998, + 0.0018378954846411943, + 0.0026159286499023438, + 0.006434095557779074, + 0.005295042879879475, + 0.004423321224749088, + 0.00892109889537096, + 0.003783962456509471, + 0.008641847409307957, + 0.003360235597938299, + 0.0058397226966917515, + 0.0073218997567892075, + 0.00742088258266449, + 0.008137192577123642, + 0.008929353207349777, + 0.008278915658593178, + 0.012374470010399818, + 0.004694809205830097, + 0.00801575556397438, + 0.0336129292845726, + 0.0136290667578578, + 0.016765493899583817, + 0.05576803907752037, + 0.015315299853682518, + 0.018815919756889343, + 0.012188740074634552, + 0.01822623983025551, + 0.036907490342855453, + 0.08236159384250641, + 0.022127678617835045, + 0.14729374647140503, + 0.006226814351975918, + 0.007979011163115501, + 0.004497217480093241, + 0.0063133034855127335, + 0.059651702642440796, + 0.06029350683093071, + 0.01064380630850792, + 0.01395376492291689, + 0.0063598440028727055, + 0.014110691845417023, + 0.007918674498796463, + 0.0074344202876091, + 0.00446258345618844, + 0.011183320544660091, + 0.03260877728462219, + 0.0126648573204875, + 0.004579022526741028, + 0.0075780898332595825, + 0.0025370903313159943, + 0.003223131410777569, + 0.0037724971771240234, + 0.005337022244930267, + 0.005308240652084351, + 0.0048084259033203125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.0036188371013849974 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tall woman wearing a sleeveless black satin gown standing beside a shorter woman seated on a velvet stool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tall woman apply vibrant pale red lipstick to the shorter woman's lips while cupping her chin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a large vanity mirror framed by warm, glowing Hollywood-style bulbs casting golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tall woman step back, nod critically, and dismiss the shorter woman with a swift wave of her hand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743494987487793 + ], + "opposite_logprob_tokens": [ + -0.9743494987487793 + ], + "logprob_reward": -0.4743494987487793, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743494987487793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03338320925831795, + 0.011144332587718964, + 0.011495999991893768, + 0.017464714124798775, + 0.00959840603172779, + 0.022557081654667854, + 0.34391260147094727, + 0.07097235321998596, + 0.09668996185064316, + 0.38278138637542725 + ], + "credit_map": [ + 0.002402320969849825, + 0.0060964226722717285, + 0.002279105829074979, + 0.0032677622511982918, + 0.0029231880325824022, + 0.0036049862392246723, + 0.004202717915177345, + 0.008606701157987118, + 0.0008705119253136218, + 0.0019474758300930262, + 0.0009002884617075324, + 0.0013469202676787972, + 0.0008122276631183922, + 0.0012894202955067158, + 0.0014507606392726302, + 0.0025267265737056732, + 0.0007757113198749721, + 0.0015137569280341268, + 0.0007054021116346121, + 0.0011308238608762622, + 0.0011704994831234217, + 0.0018119059968739748, + 0.0013367680367082357, + 0.00305113079957664, + 0.0011857706122100353, + 0.0021775842178612947, + 0.0009528286173008382, + 0.001517262775450945, + 0.0019857410807162523, + 0.0024926457554101944, + 0.0020903192926198244, + 0.005062560550868511, + 0.0006894111284054816, + 0.0014905401039868593, + 0.0007599320379085839, + 0.0011332710273563862, + 0.000986789003945887, + 0.0016131928423419595, + 0.0013102591037750244, + 0.0016150107840076089, + 0.0013961688382551074, + 0.0027480691205710173, + 0.0012331230100244284, + 0.0019587294664233923, + 0.0022083374205976725, + 0.002597850514575839, + 0.003337522502988577, + 0.007077279035001993, + 0.014757853001356125, + 0.01825464889407158, + 0.011801768094301224, + 0.017557980492711067, + 0.04230198636651039, + 0.07385329902172089, + 0.021534699946641922, + 0.14385034143924713, + 0.0033115462865680456, + 0.004525769501924515, + 0.003302814206108451, + 0.0034353549126535654, + 0.024075064808130264, + 0.021446552127599716, + 0.004719456657767296, + 0.006155788898468018, + 0.0050103929825127125, + 0.010336211882531643, + 0.0040032160468399525, + 0.008385248482227325, + 0.004520879592746496, + 0.015933305025100708, + 0.04017573595046997, + 0.00832496490329504, + 0.024013495072722435, + 0.14909759163856506, + 0.012894785962998867, + 0.029988501220941544, + 0.01784953661262989, + 0.08455006033182144, + 0.019577238708734512, + 0.04481012746691704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.4743494987487793 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_3/index_tensor([9542, 9542])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_3/index_tensor([9542, 9542])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8821cc39ea82be9fed9a67fb479b8a5beeb7f71f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_3/index_tensor([9542, 9542])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot, taken at eye level, intimately depicts the face of an adult koala. Its thick, woolly, silver-gray fur appears soft and dense, with tufts of white peeking out from its large, rounded ears. The koala's prominent, leathery black nose gleams in the light, and its small, black eyes, heavy-lidded, convey a deep, sleepy contentment. Perched on a eucalyptus tree, its jaws move in slow, rhythmic, circular, grinding motions as it lazily chews a waxy sickle-shaped eucalyptus leaf. Another fresh leaf is delicately tucked between its gray-padded paws beneath its chin. Warm afternoon sunlight filters through the treetops, creating a dappled effect of bright highlights and soft shadows, streaking across the koala's face and the rough, flaking bark of the branches behind it. The shallow depth of field transforms the background into a soft, impressionistic blur of coastal colors—the vibrant blue of the ocean and the pale yellow of the sand—creating an atmosphere of tranquility, serenity, and absolute peace.", + "seed": 27284, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_3/index_9542_seed_27284_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.043672950752079484, + "vlm_logprob": -0.043672950752079484, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.043672950752079484, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05162430182099342 + ], + "opposite_logprob_tokens": [ + -3.176624298095703 + ], + "logprob_reward": -0.05162430182099342, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05162430182099342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0655636340379715, + 0.10773700475692749, + 0.04756096005439758, + 0.02579568885266781, + 0.020880483090877533, + 0.10908795893192291, + 0.09821967035531998, + 0.043427396565675735, + 0.409584641456604, + 0.07214252650737762 + ], + "credit_map": [ + 0.0062443711794912815, + 0.007906705141067505, + 0.00772383576259017, + 0.008741420693695545, + 0.009302383288741112, + 0.0059910304844379425, + 0.012324914336204529, + 0.00732897175475955, + 0.006867385469377041, + 0.009290293790400028, + 0.006695475894957781, + 0.03683742508292198, + 0.009254257194697857, + 0.01090357918292284, + 0.019217491149902344, + 0.008671090006828308, + 0.0048972624354064465, + 0.005693384446203709, + 0.005294281989336014, + 0.0059870462864637375, + 0.005663717165589333, + 0.008603069931268692, + 0.006537182722240686, + 0.004885018337517977, + 0.002007368253543973, + 0.002143023069947958, + 0.0019147528801113367, + 0.0024786083959043026, + 0.0030431468039751053, + 0.005923103075474501, + 0.006371483206748962, + 0.0019142014207318425, + 0.0013109591091051698, + 0.002155349589884281, + 0.0015197271713986993, + 0.005091852508485317, + 0.0023995230440050364, + 0.0031423941254615784, + 0.003493244294077158, + 0.001767433132044971, + 0.007698926609009504, + 0.0067436606623232365, + 0.007222164887934923, + 0.040991511195898056, + 0.010551034472882748, + 0.01876567304134369, + 0.009738270193338394, + 0.007376725785434246, + 0.006260155700147152, + 0.01249636895954609, + 0.008356845937669277, + 0.016653794795274734, + 0.018561387434601784, + 0.010616728104650974, + 0.01569357141852379, + 0.009580818936228752, + 0.0040416233241558075, + 0.006201544776558876, + 0.003931175917387009, + 0.008864694274961948, + 0.005717907566577196, + 0.006031150929629803, + 0.005051767453551292, + 0.0035875311587005854, + 0.026353975757956505, + 0.0637684315443039, + 0.06400928646326065, + 0.06934940814971924, + 0.05180288851261139, + 0.039263464510440826, + 0.054509237408638, + 0.040527913719415665, + 0.004783915355801582, + 0.00871316622942686, + 0.005983581300824881, + 0.01299870666116476, + 0.017040686681866646, + 0.006631063763052225, + 0.00888889841735363, + 0.007102509494870901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.05162430182099342 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02063637599349022 + ], + "opposite_logprob_tokens": [ + -3.8956363201141357 + ], + "logprob_reward": -0.02063637599349022, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02063637599349022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16419322788715363, + 0.23290053009986877, + 0.06679966300725937, + 0.06463274359703064, + 0.08226259052753448, + 0.056553952395915985, + 0.103828564286232, + 0.05570129677653313, + 0.07460755854845047, + 0.09851983934640884 + ], + "credit_map": [ + 0.011430908925831318, + 0.01593615859746933, + 0.017636703327298164, + 0.020796658471226692, + 0.018882479518651962, + 0.02670445293188095, + 0.03431066498160362, + 0.01849520020186901, + 0.02013288624584675, + 0.021616360172629356, + 0.017893439158797264, + 0.06614810228347778, + 0.030719902366399765, + 0.020981479436159134, + 0.0396745391190052, + 0.01573382131755352, + 0.005191735923290253, + 0.009146161377429962, + 0.007238570135086775, + 0.00919212494045496, + 0.010484006255865097, + 0.007378065958619118, + 0.011117898859083652, + 0.007051090244203806, + 0.0042707109823822975, + 0.005373818799853325, + 0.006906483788043261, + 0.007564407307654619, + 0.008624846115708351, + 0.011550256051123142, + 0.01570580154657364, + 0.004636414814740419, + 0.0067213489674031734, + 0.010727730579674244, + 0.007432408165186644, + 0.01326026301831007, + 0.012481787241995335, + 0.012678999453783035, + 0.010816924273967743, + 0.008143125101923943, + 0.005230193957686424, + 0.006981618236750364, + 0.010141179896891117, + 0.009173128753900528, + 0.008396170102059841, + 0.006407610140740871, + 0.006461197044700384, + 0.003762852866202593, + 0.006927948445081711, + 0.011060183867812157, + 0.009416128508746624, + 0.015716657042503357, + 0.020616846159100533, + 0.013490368612110615, + 0.01855996996164322, + 0.00804045982658863, + 0.004354037344455719, + 0.005095774307847023, + 0.005437417421489954, + 0.00563095835968852, + 0.007564176805317402, + 0.009443772956728935, + 0.012690991163253784, + 0.005484165158122778, + 0.004297435283660889, + 0.013184581883251667, + 0.006434116046875715, + 0.010808756574988365, + 0.014169090427458286, + 0.007834861055016518, + 0.012013571336865425, + 0.00586513988673687, + 0.006663159467279911, + 0.012003045529127121, + 0.009487771429121494, + 0.012587244622409344, + 0.025322208181023598, + 0.00839152466505766, + 0.013730550184845924, + 0.010334338061511517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.02063637599349022 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553836219012737 + ], + "opposite_logprob_tokens": [ + -4.387553691864014 + ], + "logprob_reward": -0.012553836219012737, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553836219012737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19579610228538513, + 0.2880306541919708, + 0.046889305114746094, + 0.03430997580289841, + 0.022992590442299843, + 0.13271892070770264, + 0.06844207644462585, + 0.050300806760787964, + 0.0717114731669426, + 0.08880802243947983 + ], + "credit_map": [ + 0.012458711862564087, + 0.01297773327678442, + 0.01449657790362835, + 0.022693146020174026, + 0.022157317027449608, + 0.0404205396771431, + 0.05200600251555443, + 0.018586093559861183, + 0.020901765674352646, + 0.022178275510668755, + 0.01911921799182892, + 0.10019435733556747, + 0.028616074472665787, + 0.0282413549721241, + 0.046438973397016525, + 0.02234065905213356, + 0.00362670817412436, + 0.005540493410080671, + 0.0038031500298529863, + 0.005334490444511175, + 0.005996200256049633, + 0.009602350182831287, + 0.007868532091379166, + 0.005117384716868401, + 0.0021137543953955173, + 0.0027378960512578487, + 0.00199141982011497, + 0.003227067878469825, + 0.0034794446546584368, + 0.00806017592549324, + 0.009953929111361504, + 0.0027462909929454327, + 0.0025829204823821783, + 0.002296061720699072, + 0.0029668330680578947, + 0.005380125716328621, + 0.002666485495865345, + 0.0023279921151697636, + 0.0029201533179730177, + 0.0018520200392231345, + 0.009119013324379921, + 0.01067937258630991, + 0.011197547428309917, + 0.04546605423092842, + 0.01195431873202324, + 0.023769395425915718, + 0.012100261636078358, + 0.008432971313595772, + 0.00310928700491786, + 0.0067633674480021, + 0.004986683838069439, + 0.011803955771028996, + 0.014778676442801952, + 0.005420684814453125, + 0.016521891579031944, + 0.0050575388595461845, + 0.003481882158666849, + 0.005177715327590704, + 0.004472336266189814, + 0.01179941464215517, + 0.004805561620742083, + 0.011029714718461037, + 0.0066273328848183155, + 0.0029068528674542904, + 0.005433745216578245, + 0.008695981465280056, + 0.01094964612275362, + 0.012067541480064392, + 0.010892927646636963, + 0.006534835323691368, + 0.010522996075451374, + 0.006613810546696186, + 0.005849612411111593, + 0.011105254292488098, + 0.007969467900693417, + 0.013379676267504692, + 0.02077816054224968, + 0.007846799679100513, + 0.012477164156734943, + 0.009401892311871052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.012553836219012737 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296045780181885 + ], + "opposite_logprob_tokens": [ + -2.2379603385925293 + ], + "logprob_reward": -0.11296045780181885, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296045780181885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16271556913852692, + 0.3334316909313202, + 0.028645174577832222, + 0.0713706761598587, + 0.041087713092565536, + 0.05809812620282173, + 0.18075698614120483, + 0.03531841188669205, + 0.024446658790111542, + 0.06412902474403381 + ], + "credit_map": [ + 0.012781897559762001, + 0.014341655187308788, + 0.019662514328956604, + 0.018816350027918816, + 0.022600000724196434, + 0.024063870310783386, + 0.027152417227625847, + 0.023296859115362167, + 0.023683922365307808, + 0.024220578372478485, + 0.021607734262943268, + 0.11823413521051407, + 0.03156517818570137, + 0.03496569022536278, + 0.05241025611758232, + 0.026744231581687927, + 0.0030720133800059557, + 0.003582234727218747, + 0.0038390953559428453, + 0.0034933581482619047, + 0.004354081582278013, + 0.003537153359502554, + 0.0038590740878134966, + 0.002908163471147418, + 0.0056387148797512054, + 0.006044616457074881, + 0.005424397066235542, + 0.005606044549494982, + 0.006882197223603725, + 0.01657210662961006, + 0.020503511652350426, + 0.004699087236076593, + 0.0024253723677247763, + 0.003578288946300745, + 0.004100047517567873, + 0.010768220759928226, + 0.004914997145533562, + 0.007768045179545879, + 0.004571270197629929, + 0.002961469581350684, + 0.0038939551450312138, + 0.005376768298447132, + 0.005755770020186901, + 0.011860097758471966, + 0.016074877232313156, + 0.006039065308868885, + 0.005516295786947012, + 0.0035812954884022474, + 0.010854181833565235, + 0.020497851073741913, + 0.015360470861196518, + 0.02848268859088421, + 0.036538586020469666, + 0.024610910564661026, + 0.031594421714544296, + 0.01281785499304533, + 0.0022018379531800747, + 0.004359062761068344, + 0.0034463743213564157, + 0.006583444774150848, + 0.004658053163439035, + 0.002460036426782608, + 0.00907268188893795, + 0.0025369208306074142, + 0.0021258669439703226, + 0.0027875108644366264, + 0.0029525801073759794, + 0.003548314329236746, + 0.004448233637958765, + 0.0020251013338565826, + 0.004228651523590088, + 0.0023303995840251446, + 0.004422793630510569, + 0.007040966302156448, + 0.005137086845934391, + 0.008375107310712337, + 0.011656505987048149, + 0.009013478644192219, + 0.012056825682520866, + 0.006426254753023386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.11296045780181885 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589781925082207 + ], + "opposite_logprob_tokens": [ + -3.895589828491211 + ], + "logprob_reward": -0.020589781925082207, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589781925082207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2510315477848053, + 0.1284671127796173, + 0.0858110710978508, + 0.02160903811454773, + 0.03418176248669624, + 0.06604226678609848, + 0.06868714094161987, + 0.030606994405388832, + 0.0906691774725914, + 0.2228938192129135 + ], + "credit_map": [ + 0.015842925757169724, + 0.016204508021473885, + 0.02114417962729931, + 0.020922748371958733, + 0.019429555162787437, + 0.06852513551712036, + 0.06954734027385712, + 0.01941518485546112, + 0.01130397617816925, + 0.010466610081493855, + 0.009988641366362572, + 0.04003516957163811, + 0.010443469509482384, + 0.016849970445036888, + 0.017695199698209763, + 0.011684096418321133, + 0.00816718116402626, + 0.007867044769227505, + 0.0092253927141428, + 0.010456039570271969, + 0.01106637716293335, + 0.02282722294330597, + 0.008189587853848934, + 0.008012240752577782, + 0.002163191558793187, + 0.0025334495585411787, + 0.0033759702928364277, + 0.002697879681363702, + 0.0034894824493676424, + 0.0014643460744991899, + 0.0038852666039019823, + 0.0019994545727968216, + 0.002155303256586194, + 0.004350152797996998, + 0.0029225186444818974, + 0.006703456398099661, + 0.006663802079856396, + 0.003389457706362009, + 0.005792212672531605, + 0.002204864053055644, + 0.005247576627880335, + 0.005023859441280365, + 0.004878629464656115, + 0.010036821477115154, + 0.016066616401076317, + 0.012174306437373161, + 0.0075705042108893394, + 0.00504395691677928, + 0.005625206511467695, + 0.010366165079176426, + 0.006857475731521845, + 0.011231285519897938, + 0.010719318874180317, + 0.00562265096232295, + 0.011696531437337399, + 0.00656851613894105, + 0.0028274881187826395, + 0.003855276620015502, + 0.0037869447842240334, + 0.004299644380807877, + 0.004500729497522116, + 0.004404768347740173, + 0.005282160360366106, + 0.001649986719712615, + 0.007745229639112949, + 0.0102008618414402, + 0.013026447035372257, + 0.015796199440956116, + 0.01293101068586111, + 0.00808342918753624, + 0.013461919501423836, + 0.009424101561307907, + 0.015006464906036854, + 0.029396018013358116, + 0.019918033853173256, + 0.0360579714179039, + 0.04876871779561043, + 0.0218261256814003, + 0.029107827693223953, + 0.02281268872320652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.020589781925082207 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, taken at eye level, intimately depicts the face of an adult koala. Its thick, woolly, silver-gray fur appears soft and dense, with tufts of white peeking out from its large, rounded ears. The koala's prominent, leathery black nose gleams in the light, and its small, black eyes, heavy-lidded, convey a deep, sleepy contentment. Perched on a eucalyptus tree, its jaws move in slow, rhythmic, circular, grinding motions as it lazily chews a waxy sickle-shaped eucalyptus leaf. Another fresh leaf is delicately tucked between its gray-padded paws beneath its chin. Warm afternoon sunlight filters through the treetops, creating a dappled effect of bright highlights and soft shadows, streaking across the koala's face and the rough, flaking bark of the branches behind it. The shallow depth of field transforms the background into a soft, impressionistic blur of coastal colors—the vibrant blue of the ocean and the pale yellow of the sand—creating an atmosphere of tranquility, serenity, and absolute peace.", + "seed": 27285, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_3/index_9542_seed_27285_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022906765807420015, + "vlm_logprob": -0.022906765807420015, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.022906765807420015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023425143212080002 + ], + "opposite_logprob_tokens": [ + -3.8984251022338867 + ], + "logprob_reward": -0.023425143212080002, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023425143212080002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6108536124229431, + 0.10666798800230026, + 0.05511435493826866, + 0.09716444462537766, + 0.025091247633099556, + 0.015911780297756195, + 0.028379244729876518, + 0.013503523543477058, + 0.03686026856303215, + 0.010453509166836739 + ], + "credit_map": [ + 0.020757324993610382, + 0.031117402017116547, + 0.015947556123137474, + 0.02567196823656559, + 0.23368027806282043, + 0.029435355216264725, + 0.22944432497024536, + 0.02479938603937626, + 0.01318675559014082, + 0.013532917946577072, + 0.012029853649437428, + 0.016938090324401855, + 0.01247705239802599, + 0.010202951729297638, + 0.02243921346962452, + 0.005861151032149792, + 0.005703596863895655, + 0.006315167061984539, + 0.006781574804335833, + 0.008127879351377487, + 0.007637497968971729, + 0.007188794668763876, + 0.009874314069747925, + 0.003485528752207756, + 0.011144055984914303, + 0.014407032169401646, + 0.014138362370431423, + 0.012034524232149124, + 0.013176223263144493, + 0.010415001772344112, + 0.0137301255017519, + 0.008119125850498676, + 0.0029887768905609846, + 0.002928908448666334, + 0.0037812290247529745, + 0.002114086877554655, + 0.002830791287124157, + 0.003828557441011071, + 0.004205837845802307, + 0.002413060748949647, + 0.0017959141405299306, + 0.002358199330046773, + 0.0019421072211116552, + 0.001563207246363163, + 0.0024172081612050533, + 0.0024347833823412657, + 0.002010929398238659, + 0.001389429671689868, + 0.003171272110193968, + 0.005475756246596575, + 0.0026670407969504595, + 0.00317053752951324, + 0.00392691558226943, + 0.003499385667964816, + 0.003916170448064804, + 0.002552166348323226, + 0.0012700771912932396, + 0.0017890955787152052, + 0.002069327514618635, + 0.0012040415313094854, + 0.0016543767414987087, + 0.0027066096663475037, + 0.0018251105211675167, + 0.0009848850313574076, + 0.0032144631259143353, + 0.005527942441403866, + 0.004117791540920734, + 0.005200802348554134, + 0.0038816225714981556, + 0.0036023797001689672, + 0.00813576765358448, + 0.00317949871532619, + 0.0012507658684626222, + 0.0009654752211645246, + 0.001222691498696804, + 0.000944063940551132, + 0.001368753262795508, + 0.00248833023943007, + 0.0012662670342251658, + 0.000947161577641964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.023425143212080002 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014234318397939205 + ], + "opposite_logprob_tokens": [ + -4.26423454284668 + ], + "logprob_reward": -0.014234318397939205, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014234318397939205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18317408859729767, + 0.2655103802680969, + 0.05317598208785057, + 0.1755833774805069, + 0.07232389599084854, + 0.0468212328851223, + 0.033598609268665314, + 0.05559171363711357, + 0.07781149446964264, + 0.036409299820661545 + ], + "credit_map": [ + 0.01048295572400093, + 0.012998796068131924, + 0.01325035560876131, + 0.01581752300262451, + 0.0495341494679451, + 0.01917967200279236, + 0.05082812160253525, + 0.011082526296377182, + 0.011872501112520695, + 0.022722331807017326, + 0.02104043774306774, + 0.020517956465482712, + 0.027874575927853584, + 0.05707081779837608, + 0.08912678062915802, + 0.015284971334040165, + 0.008026319555938244, + 0.006432100664824247, + 0.006983701605349779, + 0.00474524637684226, + 0.008186476305127144, + 0.005760280415415764, + 0.009423082694411278, + 0.0036187770310789347, + 0.020757701247930527, + 0.026265623047947884, + 0.023844558745622635, + 0.01774309016764164, + 0.021458474919199944, + 0.020130924880504608, + 0.031176265329122543, + 0.014206741005182266, + 0.007136616390198469, + 0.01000578049570322, + 0.008958940394222736, + 0.009359916672110558, + 0.007541654631495476, + 0.009222823195159435, + 0.014109622687101364, + 0.00598854711279273, + 0.004979620687663555, + 0.008321660570800304, + 0.004520252346992493, + 0.0057961950078606606, + 0.005649592261761427, + 0.006577012594789267, + 0.006126598455011845, + 0.004850303754210472, + 0.0029051098972558975, + 0.0048326714895665646, + 0.003353901207447052, + 0.003307613078504801, + 0.003471544710919261, + 0.008227258920669556, + 0.004101882688701153, + 0.003398628905415535, + 0.0027001986745744944, + 0.003485481021925807, + 0.007986040785908699, + 0.0031299216207116842, + 0.0059183440171182156, + 0.01089001726359129, + 0.017991121858358383, + 0.0034905890934169292, + 0.006892770994454622, + 0.00657694460824132, + 0.011602838523685932, + 0.009255766868591309, + 0.006562546826899052, + 0.009615626186132431, + 0.021373027935624123, + 0.005931962747126818, + 0.003433781675994396, + 0.00346709368750453, + 0.004379487130790949, + 0.002835502615198493, + 0.004379195626825094, + 0.007858150638639927, + 0.006655385717749596, + 0.003400703426450491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.014234318397939205 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634391542524099 + ], + "opposite_logprob_tokens": [ + -4.882634162902832 + ], + "logprob_reward": -0.007634391542524099, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634391542524099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13277177512645721, + 0.21899884939193726, + 0.05730710178613663, + 0.22894714772701263, + 0.055593978613615036, + 0.029146557673811913, + 0.03926059231162071, + 0.09686112403869629, + 0.08015290647745132, + 0.06095988675951958 + ], + "credit_map": [ + 0.014220107346773148, + 0.011745858006179333, + 0.012256475165486336, + 0.011940492317080498, + 0.023142794147133827, + 0.022317998111248016, + 0.02514268085360527, + 0.012005390599370003, + 0.019902054220438004, + 0.016232043504714966, + 0.021550767123699188, + 0.021389221772551537, + 0.022769097238779068, + 0.03548892214894295, + 0.06654170900583267, + 0.015125061385333538, + 0.0074150655418634415, + 0.0074332887306809425, + 0.006814668420702219, + 0.0069052851758897305, + 0.009301207959651947, + 0.008071785792708397, + 0.00662141852080822, + 0.00474438676610589, + 0.02148601971566677, + 0.02918562851846218, + 0.02857791818678379, + 0.02206474542617798, + 0.02734518051147461, + 0.029278527945280075, + 0.0518687404692173, + 0.019140414893627167, + 0.006873198319226503, + 0.008303034119307995, + 0.007111974060535431, + 0.00720322597771883, + 0.006013181526213884, + 0.0077313585206866264, + 0.007767618168145418, + 0.004590397700667381, + 0.00292301457375288, + 0.004033506382256746, + 0.0029473057948052883, + 0.0035463222302496433, + 0.004192163236439228, + 0.00475427508354187, + 0.004048763774335384, + 0.0027012103237211704, + 0.003298199502751231, + 0.005614704452455044, + 0.004059366881847382, + 0.0037732147611677647, + 0.004533440340310335, + 0.009132582694292068, + 0.004921054467558861, + 0.003928032703697681, + 0.004797718487679958, + 0.007600010372698307, + 0.012499627657234669, + 0.00626167468726635, + 0.010535862296819687, + 0.019704436883330345, + 0.028589054942131042, + 0.006872756406664848, + 0.007766823284327984, + 0.010063284076750278, + 0.009124784730374813, + 0.010114908218383789, + 0.0076602306216955185, + 0.015013986267149448, + 0.013382314704358578, + 0.007026593200862408, + 0.005037229973822832, + 0.0056942836381495, + 0.00750568276271224, + 0.004350300412625074, + 0.006442011334002018, + 0.012923896312713623, + 0.013421691954135895, + 0.005584795027971268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.007634391542524099 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865988716483116 + ], + "opposite_logprob_tokens": [ + -3.0486598014831543 + ], + "logprob_reward": -0.04865988716483116, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865988716483116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5983966588973999, + 0.15284499526023865, + 0.026328420266509056, + 0.04546041786670685, + 0.025259554386138916, + 0.012985892593860626, + 0.023134157061576843, + 0.07142645120620728, + 0.030598433688282967, + 0.0135650634765625 + ], + "credit_map": [ + 0.02010175958275795, + 0.022653738036751747, + 0.013963399454951286, + 0.022578423842787743, + 0.23115696012973785, + 0.03623320162296295, + 0.22652986645698547, + 0.02517927624285221, + 0.007497522979974747, + 0.010857182554900646, + 0.013004965148866177, + 0.008466680534183979, + 0.015745362266898155, + 0.03585472330451012, + 0.053066104650497437, + 0.008352444507181644, + 0.0026004596147686243, + 0.004204153548926115, + 0.0028317768592387438, + 0.003191424300894141, + 0.0028279779944568872, + 0.0053355819545686245, + 0.0028383478056639433, + 0.002498698653653264, + 0.004965274129062891, + 0.006044477224349976, + 0.005790391471236944, + 0.004756382200866938, + 0.005278195720165968, + 0.007154928054660559, + 0.007956381887197495, + 0.0035143911372870207, + 0.0024816389195621014, + 0.002955581294372678, + 0.003568782936781645, + 0.0018464965978637338, + 0.0025619312655180693, + 0.004539396148175001, + 0.005358709953725338, + 0.0019470162224024534, + 0.0013799110893160105, + 0.0015368297463282943, + 0.0011733322171494365, + 0.0011582850711420178, + 0.0015821086708456278, + 0.002917743055149913, + 0.0020865024998784065, + 0.0011511791963130236, + 0.002755129709839821, + 0.0021187474485486746, + 0.002353687770664692, + 0.0016392400721088052, + 0.0028142433147877455, + 0.006221148185431957, + 0.003232433693483472, + 0.001999526284635067, + 0.0039533511735498905, + 0.004373825155198574, + 0.010414131917059422, + 0.0036615964490920305, + 0.0052572693675756454, + 0.014634677208960056, + 0.024560756981372833, + 0.004570844117552042, + 0.0026624605525285006, + 0.0030621157493442297, + 0.002273700898513198, + 0.0036546094343066216, + 0.0033793170005083084, + 0.007389172911643982, + 0.004608938470482826, + 0.0035681198351085186, + 0.001218268065713346, + 0.0016889236867427826, + 0.0016505400417372584, + 0.0015051125083118677, + 0.001529025612398982, + 0.002310708863660693, + 0.0025333319790661335, + 0.0011291524861007929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.04865988716483116 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058008871972561 + ], + "opposite_logprob_tokens": [ + -3.8955800533294678 + ], + "logprob_reward": -0.02058008871972561, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058008871972561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24180831015110016, + 0.09568266570568085, + 0.07449761033058167, + 0.07922324538230896, + 0.0398319773375988, + 0.032266512513160706, + 0.04112832248210907, + 0.08946652710437775, + 0.26231956481933594, + 0.043775252997875214 + ], + "credit_map": [ + 0.01484677940607071, + 0.014159001410007477, + 0.013282891362905502, + 0.014191865921020508, + 0.07654722779989243, + 0.02336091361939907, + 0.07082383334636688, + 0.014595798216760159, + 0.005346699617803097, + 0.007299658842384815, + 0.009252064861357212, + 0.006709645502269268, + 0.010674959048628807, + 0.018305007368326187, + 0.03328488767147064, + 0.004809749312698841, + 0.008967207744717598, + 0.010218720883131027, + 0.009608893655240536, + 0.00885305367410183, + 0.009939978830516338, + 0.01149195246398449, + 0.009596893563866615, + 0.005820913705974817, + 0.010171421803534031, + 0.008961708284914494, + 0.011684509925544262, + 0.00987304002046585, + 0.007379773072898388, + 0.008145591244101524, + 0.01806694082915783, + 0.004940264392644167, + 0.003325955243781209, + 0.005596550647169352, + 0.004514084197580814, + 0.0036299361381679773, + 0.0046820505522191525, + 0.007459116168320179, + 0.006278952118009329, + 0.004345329478383064, + 0.0033574062399566174, + 0.0055299969390034676, + 0.0033378740772604942, + 0.004319092724472284, + 0.0032255288679152727, + 0.005325993988662958, + 0.00448185671120882, + 0.0026887638960033655, + 0.004750214051455259, + 0.006816258188337088, + 0.004315485246479511, + 0.006454297341406345, + 0.003997350577265024, + 0.006181086413562298, + 0.005033898167312145, + 0.0035797334276139736, + 0.005288995336741209, + 0.005656587891280651, + 0.013218268752098083, + 0.005671381484717131, + 0.009175299666821957, + 0.01606367528438568, + 0.029577506706118584, + 0.0048148054629564285, + 0.017543265596032143, + 0.024214019998908043, + 0.038800112903118134, + 0.02895278111100197, + 0.021129077300429344, + 0.02501062862575054, + 0.08846490830183029, + 0.018204767256975174, + 0.0046617561019957066, + 0.004028363153338432, + 0.0048141940496861935, + 0.0028717261739075184, + 0.0067741540260612965, + 0.01046846155077219, + 0.005691573489457369, + 0.004465021658688784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.02058008871972561 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, taken at eye level, intimately depicts the face of an adult koala. Its thick, woolly, silver-gray fur appears soft and dense, with tufts of white peeking out from its large, rounded ears. The koala's prominent, leathery black nose gleams in the light, and its small, black eyes, heavy-lidded, convey a deep, sleepy contentment. Perched on a eucalyptus tree, its jaws move in slow, rhythmic, circular, grinding motions as it lazily chews a waxy sickle-shaped eucalyptus leaf. Another fresh leaf is delicately tucked between its gray-padded paws beneath its chin. Warm afternoon sunlight filters through the treetops, creating a dappled effect of bright highlights and soft shadows, streaking across the koala's face and the rough, flaking bark of the branches behind it. The shallow depth of field transforms the background into a soft, impressionistic blur of coastal colors—the vibrant blue of the ocean and the pale yellow of the sand—creating an atmosphere of tranquility, serenity, and absolute peace.", + "seed": 27286, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_3/index_9542_seed_27286_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03411424998193979, + "vlm_logprob": -0.03411424998193979, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.03411424998193979, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08411788195371628 + ], + "opposite_logprob_tokens": [ + -2.584117889404297 + ], + "logprob_reward": -0.08411788195371628, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08411788195371628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.272543340921402, + 0.05500252917408943, + 0.06936010718345642, + 0.025497639551758766, + 0.0960092544555664, + 0.065617635846138, + 0.020187143236398697, + 0.1232622042298317, + 0.16502399742603302, + 0.10749616473913193 + ], + "credit_map": [ + 0.030576132237911224, + 0.06812679767608643, + 0.024631256237626076, + 0.036315254867076874, + 0.04236801341176033, + 0.021552518010139465, + 0.022442977875471115, + 0.02653036266565323, + 0.00814516469836235, + 0.007339567877352238, + 0.004623296670615673, + 0.01123955100774765, + 0.006600520107895136, + 0.005968790035694838, + 0.005454310681670904, + 0.0056313276290893555, + 0.006079616025090218, + 0.010989704169332981, + 0.008794410154223442, + 0.011030271649360657, + 0.008631828241050243, + 0.005522533319890499, + 0.006840153131633997, + 0.01147158071398735, + 0.0030346119310706854, + 0.00586902117356658, + 0.002074082149192691, + 0.003316868795081973, + 0.00313377333804965, + 0.002963984152302146, + 0.003091397462412715, + 0.0020139003172516823, + 0.009320159442722797, + 0.015176616609096527, + 0.011091231368482113, + 0.015744246542453766, + 0.012289822101593018, + 0.010323338210582733, + 0.011602521874010563, + 0.010461320169270039, + 0.008799569681286812, + 0.014566075056791306, + 0.006493754684925079, + 0.008976000361144543, + 0.007217645179480314, + 0.006784388329833746, + 0.005376862827688456, + 0.007403336465358734, + 0.003077178727835417, + 0.00436550984159112, + 0.0020734306890517473, + 0.002606259658932686, + 0.0016307843616232276, + 0.0018928935751318932, + 0.0020570054184645414, + 0.0024840794503688812, + 0.013184157200157642, + 0.01831699162721634, + 0.009698132053017616, + 0.016659175977110863, + 0.018362488597631454, + 0.01648748107254505, + 0.020685261115431786, + 0.009868506342172623, + 0.017810162156820297, + 0.02511993795633316, + 0.020093780010938644, + 0.020316800102591515, + 0.01975366473197937, + 0.021900944411754608, + 0.022356359288096428, + 0.017672337591648102, + 0.0125415725633502, + 0.021739879623055458, + 0.010909202508628368, + 0.01371167041361332, + 0.013998561538755894, + 0.010478365235030651, + 0.01655496470630169, + 0.007561950944364071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.08411788195371628 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012571141123771667 + ], + "opposite_logprob_tokens": [ + -4.387571334838867 + ], + "logprob_reward": -0.012571141123771667, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012571141123771667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21800269186496735, + 0.10210219025611877, + 0.07641245424747467, + 0.029393259435892105, + 0.2213168442249298, + 0.03032665327191353, + 0.027051983401179314, + 0.024241281673312187, + 0.1861885040998459, + 0.0849640741944313 + ], + "credit_map": [ + 0.029955968260765076, + 0.03061152994632721, + 0.02096780203282833, + 0.02999076619744301, + 0.021926676854491234, + 0.019106483086943626, + 0.033796295523643494, + 0.031647197902202606, + 0.013986887410283089, + 0.01333303190767765, + 0.011290608905255795, + 0.013353277929127216, + 0.010940263979136944, + 0.012544576078653336, + 0.010668762028217316, + 0.015984781086444855, + 0.00864750612527132, + 0.009886998683214188, + 0.007916142232716084, + 0.010960014536976814, + 0.0074613443575799465, + 0.010791742242872715, + 0.010739819146692753, + 0.010008897632360458, + 0.00400129659101367, + 0.006527196615934372, + 0.0017511274199932814, + 0.004087984096258879, + 0.0034640785306692123, + 0.003723257454112172, + 0.0029427940025925636, + 0.0028955256566405296, + 0.01816028542816639, + 0.03445589542388916, + 0.0165545716881752, + 0.02958628721535206, + 0.02070707269012928, + 0.02425030618906021, + 0.0538451261818409, + 0.023757314309477806, + 0.004108443856239319, + 0.005992318969219923, + 0.004418536089360714, + 0.004691512323915958, + 0.00266715488396585, + 0.0035059566143900156, + 0.0025133593007922173, + 0.002429371466860175, + 0.0031074993312358856, + 0.0049678185023367405, + 0.0029818452894687653, + 0.003507171757519245, + 0.003232235088944435, + 0.00346538913436234, + 0.0033039958216249943, + 0.0024860315024852753, + 0.002731457818299532, + 0.004683529958128929, + 0.0022800476290285587, + 0.002566862152889371, + 0.003685447620227933, + 0.002581319771707058, + 0.003359321504831314, + 0.0023532966151833534, + 0.018371382728219032, + 0.030056191608309746, + 0.018069015815854073, + 0.02420417033135891, + 0.02315238118171692, + 0.02321738377213478, + 0.02689485065639019, + 0.022223137319087982, + 0.011277593672275543, + 0.017164962366223335, + 0.008976553566753864, + 0.010135265067219734, + 0.010112344287335873, + 0.00807972438633442, + 0.010525408200919628, + 0.008692226372659206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.012571141123771667 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464289728552103 + ], + "opposite_logprob_tokens": [ + -5.379642963409424 + ], + "logprob_reward": -0.00464289728552103, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464289728552103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24938179552555084, + 0.08005967736244202, + 0.06544000655412674, + 0.02998027205467224, + 0.1803261637687683, + 0.03030262142419815, + 0.025041300803422928, + 0.01875399984419346, + 0.16135023534297943, + 0.15936395525932312 + ], + "credit_map": [ + 0.03387434035539627, + 0.04638177528977394, + 0.022742291912436485, + 0.03204380348324776, + 0.02863350696861744, + 0.021997613832354546, + 0.03351546451449394, + 0.030192989856004715, + 0.01002174150198698, + 0.013635526411235332, + 0.008589614182710648, + 0.009707342833280563, + 0.009089944884181023, + 0.009782974608242512, + 0.008897321298718452, + 0.010335207916796207, + 0.006118503864854574, + 0.007113887462764978, + 0.007825028151273727, + 0.010045275092124939, + 0.006514557637274265, + 0.010452620685100555, + 0.008103959262371063, + 0.009266171604394913, + 0.003304100129753351, + 0.00534688588231802, + 0.002674093935638666, + 0.005210033617913723, + 0.0033806730061769485, + 0.0028583037201315165, + 0.003817247925326228, + 0.0033889333717525005, + 0.01574813202023506, + 0.03370843827724457, + 0.01802464760839939, + 0.027880100533366203, + 0.021800901740789413, + 0.0201246477663517, + 0.025731418281793594, + 0.017307888716459274, + 0.004545783158391714, + 0.003666933858767152, + 0.00371044734492898, + 0.004277281928807497, + 0.0032390709966421127, + 0.004730354994535446, + 0.0034181138034909964, + 0.0027146334759891033, + 0.0039524962194263935, + 0.003259201068431139, + 0.0018417767714709044, + 0.00334059982560575, + 0.0017558427061885595, + 0.005412556231021881, + 0.002559817396104336, + 0.002919010119512677, + 0.002413276582956314, + 0.003996601793915033, + 0.001439185463823378, + 0.002544430550187826, + 0.001778529374860227, + 0.0026745847426354885, + 0.001864910707809031, + 0.0020424802787601948, + 0.018436215817928314, + 0.023869702592492104, + 0.016667833551764488, + 0.019755570217967033, + 0.01938912272453308, + 0.024017583578824997, + 0.02005625143647194, + 0.01915796287357807, + 0.018004687502980232, + 0.03783196583390236, + 0.016473717987537384, + 0.022776957601308823, + 0.018478931859135628, + 0.013621795922517776, + 0.02163272351026535, + 0.010543175041675568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.00464289728552103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865807294845581 + ], + "opposite_logprob_tokens": [ + -3.0486581325531006 + ], + "logprob_reward": -0.04865807294845581, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865807294845581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18539132177829742, + 0.08474620431661606, + 0.09238284826278687, + 0.04650918021798134, + 0.17676305770874023, + 0.037045180797576904, + 0.03310677036643028, + 0.06743582338094711, + 0.16422373056411743, + 0.11239584535360336 + ], + "credit_map": [ + 0.029950439929962158, + 0.0312962643802166, + 0.016748473048210144, + 0.027529919520020485, + 0.02043086849153042, + 0.01567220687866211, + 0.02522256411612034, + 0.01854058727622032, + 0.012381678447127342, + 0.011416670866310596, + 0.008159087039530277, + 0.009571007452905178, + 0.011825691908597946, + 0.011680561117827892, + 0.008468273095786572, + 0.011243243701756, + 0.007784922141581774, + 0.010191231966018677, + 0.009058238938450813, + 0.014480679295957088, + 0.007786075118929148, + 0.013366655446588993, + 0.015736419707536697, + 0.013978629373013973, + 0.0062665073201060295, + 0.010552379302680492, + 0.003541974350810051, + 0.006169226951897144, + 0.005609455052763224, + 0.0060037244111299515, + 0.004600651562213898, + 0.0037652584724128246, + 0.012988461181521416, + 0.027461526915431023, + 0.011723103001713753, + 0.03712352365255356, + 0.017239553853869438, + 0.010542137548327446, + 0.04358191043138504, + 0.016102848574519157, + 0.005084352567791939, + 0.01025328692048788, + 0.003198065096512437, + 0.0041293976828455925, + 0.0036772601306438446, + 0.004869180731475353, + 0.0027507042977958918, + 0.0030829336028546095, + 0.0039885868318378925, + 0.00874555017799139, + 0.0026277059223502874, + 0.004161871504038572, + 0.00282563385553658, + 0.004282404202967882, + 0.0032961429096758366, + 0.003178876359015703, + 0.008569065481424332, + 0.010390267707407475, + 0.004996849223971367, + 0.00968579389154911, + 0.007787936367094517, + 0.012032661586999893, + 0.007582997437566519, + 0.006390252150595188, + 0.015719804912805557, + 0.026720106601715088, + 0.014578139409422874, + 0.02094716578722, + 0.021193573251366615, + 0.02113749273121357, + 0.024898817762732506, + 0.019028620794415474, + 0.015778597444295883, + 0.022651202976703644, + 0.013913940638303757, + 0.016680605709552765, + 0.014915436506271362, + 0.00850458350032568, + 0.01403233502060175, + 0.005919142160564661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.04865807294845581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581256598234177 + ], + "opposite_logprob_tokens": [ + -3.8955812454223633 + ], + "logprob_reward": -0.020581256598234177, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581256598234177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15432901680469513, + 0.23887479305267334, + 0.05185694247484207, + 0.05589286983013153, + 0.3463349938392639, + 0.023713670670986176, + 0.02602650225162506, + 0.018694698810577393, + 0.046492841094732285, + 0.0377836748957634 + ], + "credit_map": [ + 0.018201284110546112, + 0.03211194649338722, + 0.011742300353944302, + 0.021561160683631897, + 0.021045053377747536, + 0.01452729944139719, + 0.018344635143876076, + 0.0167953222990036, + 0.04299706965684891, + 0.026100751012563705, + 0.025621164590120316, + 0.026364553719758987, + 0.021245716139674187, + 0.040403228253126144, + 0.01954229548573494, + 0.036600008606910706, + 0.007616219576448202, + 0.008102169260382652, + 0.003998589236289263, + 0.007077476009726524, + 0.007527395151555538, + 0.005830593407154083, + 0.005800435785204172, + 0.005904061254113913, + 0.007060979027301073, + 0.010834777727723122, + 0.0039190310053527355, + 0.008779958821833134, + 0.006675384473055601, + 0.008592776022851467, + 0.005243282299488783, + 0.004786673467606306, + 0.04962641000747681, + 0.02367383800446987, + 0.021276602521538734, + 0.03182254731655121, + 0.02461863122880459, + 0.061404623091220856, + 0.10146696120500565, + 0.03244536742568016, + 0.003338421694934368, + 0.004353121388703585, + 0.002971092937514186, + 0.004160463809967041, + 0.0020549718756228685, + 0.0030489834025502205, + 0.0017026683781296015, + 0.002083944156765938, + 0.003070383332669735, + 0.007023328449577093, + 0.0023506497964262962, + 0.0035369431134313345, + 0.0023410061839967966, + 0.002669191686436534, + 0.0019847843796014786, + 0.003050212049856782, + 0.0029218897689133883, + 0.0019280160777270794, + 0.0017537538660690188, + 0.002227697055786848, + 0.002107679145410657, + 0.003513084491714835, + 0.001906480290926993, + 0.0023360964842140675, + 0.005494760815054178, + 0.007981352508068085, + 0.005761774722486734, + 0.005719482898712158, + 0.005184678826481104, + 0.004147779196500778, + 0.007277956698089838, + 0.00492505170404911, + 0.005107379984110594, + 0.008018539287149906, + 0.0034813638776540756, + 0.004979019984602928, + 0.003992708399891853, + 0.0040906332433223724, + 0.004526110365986824, + 0.0035879153292626143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.020581256598234177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot, taken at eye level, intimately depicts the face of an adult koala. Its thick, woolly, silver-gray fur appears soft and dense, with tufts of white peeking out from its large, rounded ears. The koala's prominent, leathery black nose gleams in the light, and its small, black eyes, heavy-lidded, convey a deep, sleepy contentment. Perched on a eucalyptus tree, its jaws move in slow, rhythmic, circular, grinding motions as it lazily chews a waxy sickle-shaped eucalyptus leaf. Another fresh leaf is delicately tucked between its gray-padded paws beneath its chin. Warm afternoon sunlight filters through the treetops, creating a dappled effect of bright highlights and soft shadows, streaking across the koala's face and the rough, flaking bark of the branches behind it. The shallow depth of field transforms the background into a soft, impressionistic blur of coastal colors—the vibrant blue of the ocean and the pale yellow of the sand—creating an atmosphere of tranquility, serenity, and absolute peace.", + "seed": 27287, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_3/index_9542_seed_27287_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.017609773157164454, + "vlm_logprob": -0.017609773157164454, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.017609773157164454, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016337087377905846 + ], + "opposite_logprob_tokens": [ + -4.516336917877197 + ], + "logprob_reward": -0.016337087377905846, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.016337087377905846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09231795370578766, + 0.11672390252351761, + 0.5694366097450256, + 0.02448875457048416, + 0.04762638360261917, + 0.029965611174702644, + 0.017967812716960907, + 0.01192097831517458, + 0.02376486547291279, + 0.0657871663570404 + ], + "credit_map": [ + 0.010006628930568695, + 0.014768444001674652, + 0.019506322219967842, + 0.015311561524868011, + 0.008529763668775558, + 0.006958795245736837, + 0.007315197493880987, + 0.009921238757669926, + 0.034035541117191315, + 0.01740996353328228, + 0.010997158475220203, + 0.011513491161167622, + 0.0066244518384337425, + 0.0075826579704880714, + 0.004780931398272514, + 0.02377971261739731, + 0.052030049264431, + 0.11495332419872284, + 0.04304318130016327, + 0.026051517575979233, + 0.24645879864692688, + 0.040498681366443634, + 0.02447335794568062, + 0.021927693858742714, + 0.002777526620775461, + 0.006494042929261923, + 0.0025896038860082626, + 0.006947816349565983, + 0.001485674176365137, + 0.0012070114025846124, + 0.0016461501363664865, + 0.001340928371064365, + 0.008206920698285103, + 0.0074473596177995205, + 0.006980519741773605, + 0.007226995192468166, + 0.005067241378128529, + 0.0029060072265565395, + 0.006193187087774277, + 0.003598154056817293, + 0.00479217991232872, + 0.006915385834872723, + 0.003535263007506728, + 0.004304073750972748, + 0.003947112709283829, + 0.0017151888459920883, + 0.0025824429467320442, + 0.0021739641670137644, + 0.0016517357435077429, + 0.0040762838907539845, + 0.0026284309569746256, + 0.0032674104440957308, + 0.001586283789947629, + 0.0013139151269569993, + 0.0016317764529958367, + 0.0018119753804057837, + 0.0021149273961782455, + 0.001634284621104598, + 0.0011838028440251946, + 0.002038891427218914, + 0.0013474137522280216, + 0.0009951484389603138, + 0.0010782856261357665, + 0.0015282246749848127, + 0.004798708483576775, + 0.0028221276588737965, + 0.0027003444265574217, + 0.0033545049373060465, + 0.002591058611869812, + 0.0020887600257992744, + 0.002852334175258875, + 0.002557026920840144, + 0.007879135198891163, + 0.01529165729880333, + 0.0074440124444663525, + 0.009525978937745094, + 0.005722295958548784, + 0.006276217754930258, + 0.007006283849477768, + 0.006641583517193794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.016337087377905846 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007649652194231749 + ], + "opposite_logprob_tokens": [ + -4.8826494216918945 + ], + "logprob_reward": -0.007649652194231749, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007649652194231749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17160601913928986, + 0.10037688165903091, + 0.3076254725456238, + 0.1563294380903244, + 0.050065044313669205, + 0.07797685265541077, + 0.026844656094908714, + 0.04151054844260216, + 0.028552018105983734, + 0.039113037288188934 + ], + "credit_map": [ + 0.016820579767227173, + 0.02581537701189518, + 0.052478671073913574, + 0.021290268748998642, + 0.01633475534617901, + 0.01252345833927393, + 0.007862133905291557, + 0.01848076470196247, + 0.020302604883909225, + 0.011843769811093807, + 0.010958959348499775, + 0.00747930770739913, + 0.005685379728674889, + 0.01261881273239851, + 0.004800199531018734, + 0.026687851175665855, + 0.029286155477166176, + 0.058810051530599594, + 0.026327893137931824, + 0.015430976636707783, + 0.13182252645492554, + 0.021586207672953606, + 0.012498997151851654, + 0.01186267752200365, + 0.010168706066906452, + 0.06520286202430725, + 0.012950887903571129, + 0.04382015019655228, + 0.011124867014586926, + 0.002678708638995886, + 0.004838923458009958, + 0.005544339306652546, + 0.006910084746778011, + 0.011754538863897324, + 0.00550346914678812, + 0.007659338414669037, + 0.006149821914732456, + 0.0033374319318681955, + 0.004683875944465399, + 0.004066480789333582, + 0.008876079693436623, + 0.020894868299365044, + 0.009519916027784348, + 0.01003674790263176, + 0.010975303128361702, + 0.004034408833831549, + 0.007536165416240692, + 0.006103367544710636, + 0.0031319500412791967, + 0.006802092306315899, + 0.004989486187696457, + 0.0033203884959220886, + 0.002064717700704932, + 0.002094210358336568, + 0.001650204067118466, + 0.0027916058897972107, + 0.009617358446121216, + 0.005988840013742447, + 0.004284414928406477, + 0.00642583891749382, + 0.0051983315497636795, + 0.0032748791854828596, + 0.0033416668884456158, + 0.0033792145550251007, + 0.00691057788208127, + 0.0038983579725027084, + 0.0028587630949914455, + 0.0038782162591814995, + 0.002974394243210554, + 0.0026328193489462137, + 0.0025096770841628313, + 0.0028892147820442915, + 0.004150849301367998, + 0.007494290359318256, + 0.005017523653805256, + 0.0063690668903291225, + 0.003819110104814172, + 0.0032286408822983503, + 0.004912806674838066, + 0.004120750818401575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.007649652194231749 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028208012226969004 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028208012226969004, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028208012226969004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14215432107448578, + 0.06838668882846832, + 0.4730096161365509, + 0.04277792572975159, + 0.09757325798273087, + 0.03199142590165138, + 0.018869202584028244, + 0.031925033777952194, + 0.026025889441370964, + 0.06728669255971909 + ], + "credit_map": [ + 0.014418531209230423, + 0.017054369673132896, + 0.017749831080436707, + 0.021694296970963478, + 0.018166132271289825, + 0.02590866945683956, + 0.009513230063021183, + 0.017649253830313683, + 0.006209724582731724, + 0.01320281159132719, + 0.019461100921034813, + 0.006661910563707352, + 0.0032903901301324368, + 0.004482380114495754, + 0.0033989690709859133, + 0.011679394170641899, + 0.04198385775089264, + 0.08835267275571823, + 0.041766539216041565, + 0.022508323192596436, + 0.20421232283115387, + 0.0369127057492733, + 0.02039075642824173, + 0.01688244193792343, + 0.003965241834521294, + 0.008697259239852428, + 0.0036826387513428926, + 0.01574532873928547, + 0.0029425122775137424, + 0.0020975566003471613, + 0.00234212470240891, + 0.003305258695036173, + 0.015280244871973991, + 0.01962054893374443, + 0.012316245585680008, + 0.015680598095059395, + 0.010165108367800713, + 0.006606693845242262, + 0.01096352469176054, + 0.006940280552953482, + 0.005467316135764122, + 0.005099332891404629, + 0.004016720689833164, + 0.0046595120802521706, + 0.004559046123176813, + 0.002193886786699295, + 0.0034950184635818005, + 0.0025005924981087446, + 0.0017287475056946278, + 0.00475212000310421, + 0.0030888915061950684, + 0.0029254809487611055, + 0.0015987608348950744, + 0.001668288721702993, + 0.0013487235410138965, + 0.0017581874271854758, + 0.007438588421791792, + 0.005798245780169964, + 0.003043617121875286, + 0.004635578487068415, + 0.0035904196556657553, + 0.0020914506167173386, + 0.0024673128500580788, + 0.002859820146113634, + 0.006286118645220995, + 0.0030776532366871834, + 0.002800743794068694, + 0.003852918278425932, + 0.0026575063820928335, + 0.0029218574054539204, + 0.0022473924327641726, + 0.0021816976368427277, + 0.008598782122135162, + 0.012504964135587215, + 0.007428594399243593, + 0.008429124020040035, + 0.008429361507296562, + 0.0066946991719305515, + 0.008198394440114498, + 0.007002769038081169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0028208012226969004 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305940493941307 + ], + "opposite_logprob_tokens": [ + -3.1680593490600586 + ], + "logprob_reward": -0.04305940493941307, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305940493941307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17152170836925507, + 0.11394891887903214, + 0.32007038593292236, + 0.05845388397574425, + 0.08325869590044022, + 0.08413707464933395, + 0.02413998916745186, + 0.03675410896539688, + 0.06236167252063751, + 0.04535358026623726 + ], + "credit_map": [ + 0.02143237367272377, + 0.026787998154759407, + 0.02426249161362648, + 0.03857593238353729, + 0.0179002583026886, + 0.019290629774332047, + 0.010344404727220535, + 0.012927631847560406, + 0.010506709106266499, + 0.019824111834168434, + 0.03497469052672386, + 0.011808822862803936, + 0.006455251481384039, + 0.007967346347868443, + 0.004468386061489582, + 0.017943592742085457, + 0.025883829221129417, + 0.06339187920093536, + 0.02798445336520672, + 0.01588822528719902, + 0.13756714761257172, + 0.024281037971377373, + 0.013263716362416744, + 0.011810100637376308, + 0.010810129344463348, + 0.014129996299743652, + 0.004510084167122841, + 0.012768791057169437, + 0.003537791082635522, + 0.00327280443161726, + 0.002950780326500535, + 0.0064735072664916515, + 0.011917121708393097, + 0.019601212814450264, + 0.011175143532454967, + 0.010991565883159637, + 0.009409526363015175, + 0.004956837743520737, + 0.010745668783783913, + 0.004461618606001139, + 0.010296356864273548, + 0.01793028600513935, + 0.008589057251811028, + 0.012440198101103306, + 0.012733490206301212, + 0.005768523085862398, + 0.010561970993876457, + 0.005817190743982792, + 0.0022457472514361143, + 0.007795044686645269, + 0.004326103255152702, + 0.0034587092231959105, + 0.0016333761159330606, + 0.0010031424462795258, + 0.0013171563623473048, + 0.0023607087787240744, + 0.008462665602564812, + 0.005307063460350037, + 0.0038556132931262255, + 0.0057454234920442104, + 0.0044225165620446205, + 0.0030535641126334667, + 0.0030243995133787394, + 0.002882862463593483, + 0.015481291338801384, + 0.01158920954912901, + 0.006225254852324724, + 0.008448602631688118, + 0.0060650999657809734, + 0.0064054373651742935, + 0.0034724404104053974, + 0.004674339666962624, + 0.0063507528975605965, + 0.0064790821634233, + 0.003820078680291772, + 0.006287305615842342, + 0.003443030407652259, + 0.006078398320823908, + 0.005377659574151039, + 0.00751727307215333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.04305940493941307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of an adult koala's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the koala's fur thick, woolly, and silver-gray with white tufts in its ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the koala's nose prominent, leathery, black, and gleaming in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the koala chewing a waxy sickle-shaped eucalyptus leaf with slow, rhythmic, circular jaw motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, impressionistic blur of coastal colors (blue ocean and pale yellow sand) due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181920051574707 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.018181920051574707, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181920051574707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16010785102844238, + 0.07526009529829025, + 0.5830032229423523, + 0.023829039186239243, + 0.01446212362498045, + 0.03632083162665367, + 0.023281924426555634, + 0.013933354057371616, + 0.011635346338152885, + 0.058166272938251495 + ], + "credit_map": [ + 0.015506179071962833, + 0.02272498793900013, + 0.048742543905973434, + 0.012336577288806438, + 0.013775737956166267, + 0.022612975910305977, + 0.00672772666439414, + 0.017681123688817024, + 0.014556696638464928, + 0.011399480514228344, + 0.02174130082130432, + 0.0071885846555233, + 0.0041520423255860806, + 0.006726809311658144, + 0.002451876178383827, + 0.007043302990496159, + 0.048883143812417984, + 0.10281717777252197, + 0.06243079900741577, + 0.02732456475496292, + 0.25164690613746643, + 0.04481843113899231, + 0.024706896394491196, + 0.020375283434987068, + 0.0017704747151583433, + 0.009250318631529808, + 0.0018068035133183002, + 0.006281246431171894, + 0.001845320570282638, + 0.0008019944070838392, + 0.0010825508506968617, + 0.0009903311729431152, + 0.0023920913226902485, + 0.0025099485646933317, + 0.0016370947705581784, + 0.001926919911056757, + 0.001579949981532991, + 0.0011525360168889165, + 0.0019201291725039482, + 0.001343454234302044, + 0.0037253431510180235, + 0.005312629975378513, + 0.003431417979300022, + 0.00585183035582304, + 0.006569505203515291, + 0.0028221672400832176, + 0.004239997826516628, + 0.004367939196527004, + 0.002459166571497917, + 0.0040646931156516075, + 0.0028184461407363415, + 0.0032638865523040295, + 0.0026578460820019245, + 0.0029100796673446894, + 0.0020968555472791195, + 0.0030109514482319355, + 0.0016326664481312037, + 0.0016802357276901603, + 0.0015874907840043306, + 0.002141457749530673, + 0.001979105407372117, + 0.0014186135958880186, + 0.0014006738783791661, + 0.0020931113976985216, + 0.0018146221991628408, + 0.0018926418852061033, + 0.001382028334774077, + 0.0014202238526195288, + 0.0010393403936177492, + 0.0014543889556080103, + 0.0013717543333768845, + 0.001260346150957048, + 0.006328842137008905, + 0.007135951891541481, + 0.005723929964005947, + 0.008317241445183754, + 0.005028370767831802, + 0.00798390619456768, + 0.005062130279839039, + 0.012585901655256748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.018181920051574707 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_4/index_tensor([2044, 2044])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_4/index_tensor([2044, 2044])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..443fc4ff536d4e78cd9b766009c336a4be99f807 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_4/index_tensor([2044, 2044])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An ultra-wide-angle lens, starting from a slightly elevated perspective, slowly and steadily sweeps from left to right, revealing the courtyard of an ancient Taoist temple shrouded in morning mist. Gray eaves and angled corners appear faintly through the milky white haze, while the damp bluestone floor reflects soft, diffuse sunlight. A few golden ginkgo leaves are scattered on the ground. In the center of the courtyard, an elderly man with white hair and beard, a gaunt face, clad in a loose, flowing, pure white silk Tai Chi suit, his graying hair pulled into a neat bun, is absorbed in practicing Tai Chi. As the camera pans, he steadily shifts his weight, his arms spreading out like flowing water, forming a classic \"white crane spreads its wings\" pose. His wide sleeves flutter gently, and his expression, with closed eyes and a focused expression, reveals a tranquility that blends with nature. The entire picture has a cool tone, dominated by the gray of the stone, the dark color of the wood and the white of the characters, creating an atmosphere of extreme tranquility, concentration and full of oriental philosophical wisdom. The air seems to be filled with moist soil and a faint scent of sandalwood.", + "seed": 5069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_4/index_2044_seed_5069_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0491354470141232, + "vlm_logprob": -0.0491354470141232, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.0491354470141232, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1704421490430832 + ], + "opposite_logprob_tokens": [ + -1.9204421043395996 + ], + "logprob_reward": -0.1704421490430832, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1704421490430832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35075417160987854, + 0.061743855476379395, + 0.02297940105199814, + 0.15785479545593262, + 0.03766616806387901, + 0.048638392239809036, + 0.05282127857208252, + 0.05628441274166107, + 0.14478987455368042, + 0.06646771728992462 + ], + "credit_map": [ + 0.025139344856142998, + 0.024830980226397514, + 0.08352301269769669, + 0.0319865345954895, + 0.040572620928287506, + 0.0182192325592041, + 0.0326593853533268, + 0.09382304549217224, + 0.006390166003257036, + 0.010934915393590927, + 0.00739397294819355, + 0.008971963077783585, + 0.007596439681947231, + 0.00428653322160244, + 0.008039744570851326, + 0.008130121044814587, + 0.0020528212189674377, + 0.0038446870166808367, + 0.002980586374178529, + 0.0035485520493239164, + 0.0013800904853269458, + 0.0018712605815380812, + 0.003604755736887455, + 0.0036966463085263968, + 0.014022807590663433, + 0.0167679563164711, + 0.028644468635320663, + 0.02306019328534603, + 0.012322735972702503, + 0.012142211198806763, + 0.03687671944499016, + 0.014017703011631966, + 0.0026732280384749174, + 0.004046933725476265, + 0.006193314678966999, + 0.004261423368006945, + 0.003953349776566029, + 0.004936869256198406, + 0.005114279221743345, + 0.006486769765615463, + 0.0034037153236567974, + 0.0042943828739225864, + 0.006808497942984104, + 0.006373813841491938, + 0.00599499000236392, + 0.005107088480144739, + 0.010560860857367516, + 0.006095041986554861, + 0.005389476660639048, + 0.007231986615806818, + 0.005996118299663067, + 0.006015693303197622, + 0.005978621542453766, + 0.0053719510324299335, + 0.0072966329753398895, + 0.009540795348584652, + 0.0033987804781645536, + 0.006762687582522631, + 0.0075443643145263195, + 0.012449042871594429, + 0.008696929551661015, + 0.0036883195862174034, + 0.007018779404461384, + 0.006725512910634279, + 0.010178455151617527, + 0.029759852215647697, + 0.01730867102742195, + 0.020424168556928635, + 0.015974584966897964, + 0.01337371300905943, + 0.017848452553153038, + 0.01992196962237358, + 0.006821407936513424, + 0.010436523705720901, + 0.007181881926953793, + 0.008673533797264099, + 0.00857809092849493, + 0.007312044966965914, + 0.007864169776439667, + 0.009600060060620308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.1704421490430832 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025064502842724323 + ], + "opposite_logprob_tokens": [ + -6.002506256103516 + ], + "logprob_reward": -0.0025064502842724323, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025064502842724323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.310710608959198, + 0.06325127184391022, + 0.04853254184126854, + 0.2551848888397217, + 0.0461181104183197, + 0.052688874304294586, + 0.040780190378427505, + 0.07669985294342041, + 0.050822481513023376, + 0.05521122366189957 + ], + "credit_map": [ + 0.02273014932870865, + 0.02251437120139599, + 0.07710304856300354, + 0.026762671768665314, + 0.035286761820316315, + 0.0163867250084877, + 0.027383793145418167, + 0.08254308253526688, + 0.006390136666595936, + 0.009454209357500076, + 0.0073735895566642284, + 0.012178972363471985, + 0.006151957903057337, + 0.004726353567093611, + 0.008387119509279728, + 0.008588920347392559, + 0.0039601558819413185, + 0.008926388807594776, + 0.006683975923806429, + 0.007017267867922783, + 0.003824556013569236, + 0.0037378170527517796, + 0.007998242042958736, + 0.006384130567312241, + 0.023713139817118645, + 0.02820567786693573, + 0.04108477756381035, + 0.02970392070710659, + 0.021509293466806412, + 0.01866075024008751, + 0.0670497789978981, + 0.02525753155350685, + 0.0041166190057992935, + 0.005994170904159546, + 0.00675416411831975, + 0.00455496646463871, + 0.006279313936829567, + 0.005478124134242535, + 0.005828302353620529, + 0.007112446706742048, + 0.003494475269690156, + 0.004916935693472624, + 0.006943405140191317, + 0.006686740554869175, + 0.007902349345386028, + 0.007090355735272169, + 0.008383753709495068, + 0.007270855363458395, + 0.0032908557914197445, + 0.004461407661437988, + 0.005454160273075104, + 0.010285407304763794, + 0.0036924390587955713, + 0.0026349001564085484, + 0.006620686501264572, + 0.0043403273448348045, + 0.005262298975139856, + 0.009729905053973198, + 0.00795348547399044, + 0.013853231444954872, + 0.010026922449469566, + 0.006544362287968397, + 0.009113901294767857, + 0.014215746894478798, + 0.0038888815324753523, + 0.009815267287194729, + 0.006082954816520214, + 0.007509833667427301, + 0.006208283361047506, + 0.004213673993945122, + 0.006562167312949896, + 0.0065414151176810265, + 0.005484934896230698, + 0.006618314888328314, + 0.005531507544219494, + 0.00836669746786356, + 0.00869266502559185, + 0.004513170570135117, + 0.007944147102534771, + 0.008059783838689327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.0025064502842724323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253676325082779 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.005253676325082779, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253676325082779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17747750878334045, + 0.060110267251729965, + 0.043093644082546234, + 0.4341166615486145, + 0.05206988751888275, + 0.037697017192840576, + 0.04886573925614357, + 0.08898178488016129, + 0.018683113157749176, + 0.038904380053281784 + ], + "credit_map": [ + 0.012307094410061836, + 0.013810936361551285, + 0.041529055684804916, + 0.016269125044345856, + 0.02397950179874897, + 0.01081068441271782, + 0.01540306769311428, + 0.04336803779006004, + 0.005313050467520952, + 0.009845135733485222, + 0.007430169731378555, + 0.007577894255518913, + 0.008351538330316544, + 0.005694591905921698, + 0.007384218741208315, + 0.008513666689395905, + 0.004156634211540222, + 0.006643610540777445, + 0.005176861770451069, + 0.005098981782793999, + 0.00383122474886477, + 0.004809487611055374, + 0.006955379620194435, + 0.006421458441764116, + 0.03276591747999191, + 0.04576198756694794, + 0.0688369870185852, + 0.08323000371456146, + 0.03739692270755768, + 0.06027897447347641, + 0.06205769255757332, + 0.04378817230463028, + 0.004589118994772434, + 0.006336333230137825, + 0.00709001999348402, + 0.006215887609869242, + 0.007547623477876186, + 0.006773252971470356, + 0.008486254140734673, + 0.005031396169215441, + 0.002921739360317588, + 0.004204648081213236, + 0.004259094595909119, + 0.004349334165453911, + 0.00665872311219573, + 0.005179228261113167, + 0.005375633481889963, + 0.004748616833239794, + 0.004255414940416813, + 0.003002702724188566, + 0.0062169828452169895, + 0.012881397269666195, + 0.0053908186964690685, + 0.00265084533020854, + 0.010062034241855145, + 0.004405542742460966, + 0.005766468122601509, + 0.012831629253923893, + 0.010080443695187569, + 0.017935432493686676, + 0.01346661802381277, + 0.006097717676311731, + 0.011263214983046055, + 0.01154025923460722, + 0.0017323150532320142, + 0.0024902315344661474, + 0.0028270184993743896, + 0.003023696830496192, + 0.002089593792334199, + 0.0014941084664314985, + 0.0025538636837154627, + 0.0024722826201468706, + 0.004434006288647652, + 0.004378292243927717, + 0.005441871006041765, + 0.006087642628699541, + 0.004515267442911863, + 0.0023847108241170645, + 0.004916029050946236, + 0.0067465598694980145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.005253676325082779 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493466183543205 + ], + "opposite_logprob_tokens": [ + -2.9299347400665283 + ], + "logprob_reward": -0.05493466183543205, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493466183543205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2952251434326172, + 0.13957370817661285, + 0.04055782034993172, + 0.08754787594079971, + 0.043453242629766464, + 0.18764014542102814, + 0.06624335050582886, + 0.05391308292746544, + 0.042640138417482376, + 0.04320554435253143 + ], + "credit_map": [ + 0.020847227424383163, + 0.02775484137237072, + 0.06286587566137314, + 0.028103433549404144, + 0.04384500905871391, + 0.01953776553273201, + 0.03374355658888817, + 0.05852741003036499, + 0.016743313521146774, + 0.024927346035838127, + 0.015889985486865044, + 0.02212933637201786, + 0.014986139722168446, + 0.010063275694847107, + 0.018787605687975883, + 0.016046691685914993, + 0.004007593262940645, + 0.0056824516505002975, + 0.005507874768227339, + 0.006077027879655361, + 0.0029358472675085068, + 0.0040154363960027695, + 0.005837216507643461, + 0.006494371686130762, + 0.007481295615434647, + 0.010033036582171917, + 0.01340700313448906, + 0.010805359110236168, + 0.006518403068184853, + 0.0080966642126441, + 0.021143821999430656, + 0.010062281973659992, + 0.003624072065576911, + 0.004900592844933271, + 0.00935263279825449, + 0.005671198479831219, + 0.0050541190430521965, + 0.002698878524824977, + 0.007353846449404955, + 0.004797903820872307, + 0.012707711197435856, + 0.01572452299296856, + 0.024622656404972076, + 0.02718346007168293, + 0.03238891810178757, + 0.015058638527989388, + 0.04004589468240738, + 0.01990833505988121, + 0.005694147665053606, + 0.007296775933355093, + 0.010667964816093445, + 0.007974636740982533, + 0.007704333867877722, + 0.007032182067632675, + 0.008871609345078468, + 0.01100169587880373, + 0.0028771567158401012, + 0.005003315396606922, + 0.0057450467720627785, + 0.00805220752954483, + 0.006876146886497736, + 0.004219429567456245, + 0.007158104330301285, + 0.013981672003865242, + 0.003302454249933362, + 0.008793475106358528, + 0.0061260005459189415, + 0.007678098976612091, + 0.003369937650859356, + 0.002828494645655155, + 0.005999372806400061, + 0.00454229861497879, + 0.0044309538789093494, + 0.00461607426404953, + 0.00541574414819479, + 0.005084417760372162, + 0.005084235221147537, + 0.005199187435209751, + 0.007958248257637024, + 0.005416682921350002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.05493466183543205 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540297582745552 + ], + "opposite_logprob_tokens": [ + -4.387540340423584 + ], + "logprob_reward": -0.012540297582745552, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540297582745552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27524900436401367, + 0.06704447418451309, + 0.03431873023509979, + 0.10307976603507996, + 0.047592464834451675, + 0.0881427526473999, + 0.11776182055473328, + 0.056074127554893494, + 0.10207495093345642, + 0.10866191238164902 + ], + "credit_map": [ + 0.02421337552368641, + 0.01935630664229393, + 0.0611422099173069, + 0.02485247515141964, + 0.034908175468444824, + 0.015642566606402397, + 0.026061007753014565, + 0.06907287985086441, + 0.008611100725829601, + 0.013269124552607536, + 0.007086003199219704, + 0.009142118506133556, + 0.00735582085326314, + 0.004127079620957375, + 0.00864393450319767, + 0.008809290826320648, + 0.002511420054361224, + 0.0056049637496471405, + 0.004762747325003147, + 0.004757990129292011, + 0.0024100681766867638, + 0.002833052072674036, + 0.005491821561008692, + 0.005946665070950985, + 0.007407213095575571, + 0.010546157136559486, + 0.013985357247292995, + 0.015432031825184822, + 0.010613350197672844, + 0.01344740018248558, + 0.020358504727482796, + 0.011289754882454872, + 0.003458324819803238, + 0.005542587023228407, + 0.006076303776353598, + 0.004083586856722832, + 0.007527487352490425, + 0.008014780469238758, + 0.006617569364607334, + 0.006271826568990946, + 0.006715313531458378, + 0.007511925883591175, + 0.012738323770463467, + 0.013518337160348892, + 0.010870063677430153, + 0.006063968408852816, + 0.02111542969942093, + 0.009609398432075977, + 0.010705907829105854, + 0.015560672618448734, + 0.018317271023988724, + 0.020358292385935783, + 0.009876609779894352, + 0.00801630038768053, + 0.015193375758826733, + 0.019733401015400887, + 0.002747263992205262, + 0.0051440345123410225, + 0.005794732831418514, + 0.011630255728960037, + 0.007470494601875544, + 0.0038258216809481382, + 0.006717267446219921, + 0.012744251638650894, + 0.007428036071360111, + 0.022219493985176086, + 0.012346486561000347, + 0.01783710904419422, + 0.010307722724974155, + 0.010124203749001026, + 0.0127952191978693, + 0.009016674943268299, + 0.01159349549561739, + 0.015379001386463642, + 0.013807877898216248, + 0.017354417592287064, + 0.014481066726148129, + 0.006977851502597332, + 0.012944650836288929, + 0.01612355187535286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.012540297582745552 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An ultra-wide-angle lens, starting from a slightly elevated perspective, slowly and steadily sweeps from left to right, revealing the courtyard of an ancient Taoist temple shrouded in morning mist. Gray eaves and angled corners appear faintly through the milky white haze, while the damp bluestone floor reflects soft, diffuse sunlight. A few golden ginkgo leaves are scattered on the ground. In the center of the courtyard, an elderly man with white hair and beard, a gaunt face, clad in a loose, flowing, pure white silk Tai Chi suit, his graying hair pulled into a neat bun, is absorbed in practicing Tai Chi. As the camera pans, he steadily shifts his weight, his arms spreading out like flowing water, forming a classic \"white crane spreads its wings\" pose. His wide sleeves flutter gently, and his expression, with closed eyes and a focused expression, reveals a tranquility that blends with nature. The entire picture has a cool tone, dominated by the gray of the stone, the dark color of the wood and the white of the characters, creating an atmosphere of extreme tranquility, concentration and full of oriental philosophical wisdom. The air seems to be filled with moist soil and a faint scent of sandalwood.", + "seed": 5070, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_4/index_2044_seed_5070_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09350732588209212, + "vlm_logprob": -0.09350732588209212, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.09350732588209212, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3637690544128418 + ], + "opposite_logprob_tokens": [ + -1.2387690544128418 + ], + "logprob_reward": -0.3637690544128418, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3637690544128418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5009222030639648, + 0.05653392896056175, + 0.044772353023290634, + 0.04624604806303978, + 0.042749471962451935, + 0.060974765568971634, + 0.1258014440536499, + 0.032075945287942886, + 0.04881072789430618, + 0.04111308604478836 + ], + "credit_map": [ + 0.0405249148607254, + 0.05980338528752327, + 0.03157828748226166, + 0.11577428877353668, + 0.08840309828519821, + 0.038117825984954834, + 0.03282749280333519, + 0.0938928872346878, + 0.004575980361551046, + 0.01131140161305666, + 0.00460744509473443, + 0.007366965990513563, + 0.011027878150343895, + 0.004064936190843582, + 0.008976220153272152, + 0.00460310373455286, + 0.0073721762746572495, + 0.007861743681132793, + 0.005576424766331911, + 0.010323495604097843, + 0.004129247739911079, + 0.0019784667529165745, + 0.0037200814113020897, + 0.0038107188884168863, + 0.006923850625753403, + 0.005667988210916519, + 0.004730040207505226, + 0.012539667077362537, + 0.005014168564230204, + 0.0027316026389598846, + 0.004943132866173983, + 0.003695598104968667, + 0.0047774906270205975, + 0.005716023501008749, + 0.004981634207069874, + 0.0052618831396102905, + 0.00778629444539547, + 0.002952100243419409, + 0.006012694910168648, + 0.005261349491775036, + 0.00780201330780983, + 0.00657678721472621, + 0.009545338340103626, + 0.010301200672984123, + 0.007455596700310707, + 0.006771485786885023, + 0.007674314547330141, + 0.004848025273531675, + 0.03625784441828728, + 0.008869840763509274, + 0.01474524661898613, + 0.01194105576723814, + 0.007395484484732151, + 0.007614534813910723, + 0.01865174062550068, + 0.02032569982111454, + 0.003476472571492195, + 0.0035037368070334196, + 0.0028980516362935305, + 0.005239102989435196, + 0.005807987879961729, + 0.002890676725655794, + 0.004678231664001942, + 0.0035816850140690804, + 0.010934652760624886, + 0.005970137659460306, + 0.0063829608261585236, + 0.004305612295866013, + 0.005511367227882147, + 0.005118933506309986, + 0.005232761148363352, + 0.005354306194931269, + 0.008203505538403988, + 0.004754148423671722, + 0.004967282991856337, + 0.004006354138255119, + 0.003750080009922385, + 0.004202835727483034, + 0.004624412395060062, + 0.006604467052966356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.3637690544128418 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025159630458801985 + ], + "opposite_logprob_tokens": [ + -6.00251579284668 + ], + "logprob_reward": -0.0025159630458801985, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025159630458801985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2905775010585785, + 0.11676173657178879, + 0.0964871272444725, + 0.06468401104211807, + 0.05697469040751457, + 0.04967591539025307, + 0.18082886934280396, + 0.07230565696954727, + 0.025292638689279556, + 0.046411894261837006 + ], + "credit_map": [ + 0.019691800698637962, + 0.03743838891386986, + 0.019585484638810158, + 0.06894207000732422, + 0.06057574227452278, + 0.021964749321341515, + 0.02121112309396267, + 0.04116812348365784, + 0.007802398409694433, + 0.02427726425230503, + 0.010799753479659557, + 0.017462514340877533, + 0.019350003451108932, + 0.007425693329423666, + 0.017562463879585266, + 0.012081636115908623, + 0.009989430196583271, + 0.023655755445361137, + 0.014274606481194496, + 0.015937520191073418, + 0.012461762875318527, + 0.004635636694729328, + 0.008251156657934189, + 0.007281250786036253, + 0.0066850814037024975, + 0.007056666072458029, + 0.011885303072631359, + 0.011206473223865032, + 0.009905025362968445, + 0.004136167466640472, + 0.007430345751345158, + 0.006378944963216782, + 0.006683725863695145, + 0.00796355027705431, + 0.0055540199391543865, + 0.0072386846877634525, + 0.010754775255918503, + 0.003466513939201832, + 0.008906712755560875, + 0.006406704895198345, + 0.0064812046475708485, + 0.0053892009891569614, + 0.007952372543513775, + 0.008764859288930893, + 0.005389803554862738, + 0.005075132939964533, + 0.006615750025957823, + 0.004007592331618071, + 0.05869156867265701, + 0.013964728452265263, + 0.010796893388032913, + 0.0148066570982337, + 0.009198455139994621, + 0.006387357600033283, + 0.05093501880764961, + 0.016048172488808632, + 0.007953852415084839, + 0.00983746349811554, + 0.007780025247484446, + 0.01101989857852459, + 0.011958178132772446, + 0.00581393064931035, + 0.010616845451295376, + 0.007325457409024239, + 0.005919523071497679, + 0.0028247085865586996, + 0.003624618286266923, + 0.002064724685624242, + 0.002592759905382991, + 0.0027299148496240377, + 0.0026309120003134012, + 0.0029054763726890087, + 0.009199784137308598, + 0.004482409451156855, + 0.00583608215674758, + 0.004484310746192932, + 0.005184118635952473, + 0.004270229954272509, + 0.006022133864462376, + 0.006932820193469524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0025159630458801985 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097400698810816 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097400698810816, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097400698810816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3387928307056427, + 0.05384720489382744, + 0.08152669668197632, + 0.10867641121149063, + 0.08861737698316574, + 0.03968190774321556, + 0.17183825373649597, + 0.05456363037228584, + 0.025268884375691414, + 0.037186842411756516 + ], + "credit_map": [ + 0.015765441581606865, + 0.02552918903529644, + 0.015226504765450954, + 0.1043456420302391, + 0.0550217479467392, + 0.024682391434907913, + 0.022675352171063423, + 0.07554653286933899, + 0.003776284633204341, + 0.008896330371499062, + 0.005655328743159771, + 0.0067350356839597225, + 0.010182695463299751, + 0.004080041777342558, + 0.007777821272611618, + 0.006743669975548983, + 0.011468835175037384, + 0.010904023423790932, + 0.010806990787386894, + 0.014959531836211681, + 0.007988553494215012, + 0.003710636403411627, + 0.013946768827736378, + 0.007741360925137997, + 0.011403973214328289, + 0.011102769523859024, + 0.016937071457505226, + 0.019513998180627823, + 0.015757277607917786, + 0.007657195907086134, + 0.012461873702704906, + 0.013842253014445305, + 0.010656408965587616, + 0.008809174410998821, + 0.015046480111777782, + 0.013577837496995926, + 0.010615854524075985, + 0.0047296383418142796, + 0.012204973958432674, + 0.01297700684517622, + 0.006164579652249813, + 0.004673215094953775, + 0.007027862127870321, + 0.005535265896469355, + 0.0040926458314061165, + 0.0027302727103233337, + 0.004690750502049923, + 0.004767316859215498, + 0.05329873785376549, + 0.013122654519975185, + 0.011116951704025269, + 0.014474118128418922, + 0.009722528047859669, + 0.005560177844017744, + 0.04905140399932861, + 0.015491674654185772, + 0.005868035368621349, + 0.005467014387249947, + 0.004668375942856073, + 0.012338832020759583, + 0.008719666860997677, + 0.0033196001313626766, + 0.007698706816881895, + 0.006483395583927631, + 0.0034855499397963285, + 0.004070383962243795, + 0.002663909224793315, + 0.0028820657171308994, + 0.0038501673843711615, + 0.002154975663870573, + 0.0029946223367005587, + 0.003167211776599288, + 0.005581756588071585, + 0.0042968797497451305, + 0.0044764322228729725, + 0.0032768994569778442, + 0.004795849323272705, + 0.00442513357847929, + 0.004275602754205465, + 0.006058288738131523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.004097400698810816 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896234095096588 + ], + "opposite_logprob_tokens": [ + -2.5789623260498047 + ], + "logprob_reward": -0.07896234095096588, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896234095096588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3538527190685272, + 0.08214126527309418, + 0.08203473687171936, + 0.0416472889482975, + 0.024204183369874954, + 0.0370652936398983, + 0.0828545019030571, + 0.1480448842048645, + 0.075600765645504, + 0.07255431264638901 + ], + "credit_map": [ + 0.028737571090459824, + 0.030053040012717247, + 0.024738091975450516, + 0.08250024169683456, + 0.03814235329627991, + 0.03334317356348038, + 0.02415645495057106, + 0.09218178689479828, + 0.006156853400170803, + 0.014138326048851013, + 0.009315789677202702, + 0.010715312324464321, + 0.013203120790421963, + 0.008807571604847908, + 0.008456116542220116, + 0.011348173022270203, + 0.012862375006079674, + 0.008988333866000175, + 0.01316223293542862, + 0.016102075576782227, + 0.010135679505765438, + 0.006522825453430414, + 0.007961530238389969, + 0.006299684755504131, + 0.005838148295879364, + 0.005210317205637693, + 0.006516302470117807, + 0.005026842933148146, + 0.005771559197455645, + 0.004620704334229231, + 0.0039653233252465725, + 0.004698093514889479, + 0.005291296169161797, + 0.0022999895736575127, + 0.003551447531208396, + 0.002418584655970335, + 0.0025080672930926085, + 0.002019132487475872, + 0.0030235385056585073, + 0.003092127852141857, + 0.00548932608217001, + 0.004756735637784004, + 0.0035768223460763693, + 0.004761465359479189, + 0.006527905818074942, + 0.0022993318270891905, + 0.006130062974989414, + 0.003523642895743251, + 0.022421857342123985, + 0.00803941860795021, + 0.007650299929082394, + 0.00840796995908022, + 0.006366113666445017, + 0.004894724581390619, + 0.01861843839287758, + 0.006455680821090937, + 0.02220982126891613, + 0.027961574494838715, + 0.019820213317871094, + 0.02643638849258423, + 0.012306032702326775, + 0.007528313901275396, + 0.01656949892640114, + 0.015213035978376865, + 0.01572389155626297, + 0.01100285816937685, + 0.01049965713173151, + 0.008012289181351662, + 0.006423846818506718, + 0.006067116279155016, + 0.008781279437243938, + 0.009089826606214046, + 0.011123009957373142, + 0.007135705091059208, + 0.00985010527074337, + 0.00724708242341876, + 0.009505837224423885, + 0.009553213603794575, + 0.008794564753770828, + 0.009344788268208504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.07896234095096588 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181918703019619 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.0181918703019619, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181918703019619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.312578946352005, + 0.1375671625137329, + 0.035626303404569626, + 0.06400921940803528, + 0.03385278210043907, + 0.04712139070034027, + 0.2414746880531311, + 0.022961171343922615, + 0.06041451543569565, + 0.0443938709795475 + ], + "credit_map": [ + 0.02573362924158573, + 0.04372522607445717, + 0.020823504775762558, + 0.055750347673892975, + 0.07402484118938446, + 0.030921345576643944, + 0.023110952228307724, + 0.038489099591970444, + 0.00913870707154274, + 0.019255351275205612, + 0.016467010602355003, + 0.015952812507748604, + 0.021431731060147285, + 0.012041502632200718, + 0.01947741210460663, + 0.023802638053894043, + 0.003305418649688363, + 0.006793079897761345, + 0.004776024259626865, + 0.004558888729661703, + 0.004263877868652344, + 0.0019933911971747875, + 0.004404781386256218, + 0.005530840717256069, + 0.008835614658892155, + 0.007433145772665739, + 0.007654963526874781, + 0.015432296320796013, + 0.006120635662227869, + 0.003837586846202612, + 0.0072130910120904446, + 0.007481891196221113, + 0.003961633890867233, + 0.0042910086922347546, + 0.004146690480411053, + 0.003746533300727606, + 0.006913456600159407, + 0.0021716717164963484, + 0.004953885450959206, + 0.003667902434244752, + 0.005972892977297306, + 0.005118806380778551, + 0.007290266919881105, + 0.008567813783884048, + 0.006114447955042124, + 0.004743030294775963, + 0.005523037165403366, + 0.003791089402511716, + 0.07162749767303467, + 0.021841930225491524, + 0.016391253098845482, + 0.017387252300977707, + 0.012415970675647259, + 0.01212528720498085, + 0.0719129666686058, + 0.01777251809835434, + 0.002097000367939472, + 0.002343884203583002, + 0.0017426933627575636, + 0.006082657724618912, + 0.003346913494169712, + 0.0014589362544938922, + 0.003125207731500268, + 0.002763879019767046, + 0.013761650770902634, + 0.01037503406405449, + 0.007676530629396439, + 0.0051718358881771564, + 0.006536470726132393, + 0.005592129193246365, + 0.005500506609678268, + 0.005800359882414341, + 0.009333987720310688, + 0.005469590425491333, + 0.00564905721694231, + 0.0046644737012684345, + 0.0037404231261461973, + 0.004644562024623156, + 0.004404544830322266, + 0.006487228907644749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0181918703019619 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An ultra-wide-angle lens, starting from a slightly elevated perspective, slowly and steadily sweeps from left to right, revealing the courtyard of an ancient Taoist temple shrouded in morning mist. Gray eaves and angled corners appear faintly through the milky white haze, while the damp bluestone floor reflects soft, diffuse sunlight. A few golden ginkgo leaves are scattered on the ground. In the center of the courtyard, an elderly man with white hair and beard, a gaunt face, clad in a loose, flowing, pure white silk Tai Chi suit, his graying hair pulled into a neat bun, is absorbed in practicing Tai Chi. As the camera pans, he steadily shifts his weight, his arms spreading out like flowing water, forming a classic \"white crane spreads its wings\" pose. His wide sleeves flutter gently, and his expression, with closed eyes and a focused expression, reveals a tranquility that blends with nature. The entire picture has a cool tone, dominated by the gray of the stone, the dark color of the wood and the white of the characters, creating an atmosphere of extreme tranquility, concentration and full of oriental philosophical wisdom. The air seems to be filled with moist soil and a faint scent of sandalwood.", + "seed": 5071, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_4/index_2044_seed_5071_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.043493061792105436, + "vlm_logprob": -0.043493061792105436, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.043493061792105436, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13625112175941467 + ], + "opposite_logprob_tokens": [ + -2.136251211166382 + ], + "logprob_reward": -0.13625112175941467, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13625112175941467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.270303875207901, + 0.05449146032333374, + 0.04127965494990349, + 0.04607878625392914, + 0.22093495726585388, + 0.19500070810317993, + 0.04789164662361145, + 0.02961309254169464, + 0.053241848945617676, + 0.041164010763168335 + ], + "credit_map": [ + 0.02352765202522278, + 0.028379039838910103, + 0.03680824860930443, + 0.03272926062345505, + 0.036642950028181076, + 0.03280654922127724, + 0.03164542838931084, + 0.04776473715901375, + 0.0056692808866500854, + 0.0057078758254647255, + 0.006389007903635502, + 0.008001849055290222, + 0.004943800624459982, + 0.006084918975830078, + 0.01181872934103012, + 0.005875998176634312, + 0.00386089482344687, + 0.004130050539970398, + 0.007417052518576384, + 0.006509582046419382, + 0.004571991506963968, + 0.005634657107293606, + 0.004155384376645088, + 0.005000040400773287, + 0.002992029068991542, + 0.004450123757123947, + 0.008023222908377647, + 0.008206618018448353, + 0.005222593434154987, + 0.0060897329822182655, + 0.005454124417155981, + 0.005640340968966484, + 0.010183914564549923, + 0.005797530524432659, + 0.018602998927235603, + 0.010979888029396534, + 0.01058224868029356, + 0.01777125522494316, + 0.08467997610569, + 0.06233714520931244, + 0.01242057979106903, + 0.020151276141405106, + 0.016476470977067947, + 0.0279085710644722, + 0.02075284719467163, + 0.03335583955049515, + 0.03603425621986389, + 0.02790086530148983, + 0.002800652990117669, + 0.003022267948836088, + 0.004155353177338839, + 0.005166302900761366, + 0.006313389167189598, + 0.00603076908737421, + 0.013812323100864887, + 0.006590591743588448, + 0.0025821735616773367, + 0.004299665801227093, + 0.003865624312311411, + 0.0031716255471110344, + 0.002901364117860794, + 0.005150428507477045, + 0.003920775838196278, + 0.0037214329931885004, + 0.0064152986742556095, + 0.008525981567800045, + 0.005553358234465122, + 0.007000377401709557, + 0.00648901704698801, + 0.007489123847335577, + 0.005908672697842121, + 0.0058600204065442085, + 0.006648601032793522, + 0.005227757152169943, + 0.004074769094586372, + 0.006070286501199007, + 0.004680134356021881, + 0.004990439862012863, + 0.005149677861481905, + 0.004322342574596405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.13625112175941467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004119126591831446 + ], + "opposite_logprob_tokens": [ + -5.504118919372559 + ], + "logprob_reward": -0.004119126591831446, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004119126591831446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06424460560083389, + 0.04017588123679161, + 0.047857195138931274, + 0.02551407739520073, + 0.26426059007644653, + 0.37872663140296936, + 0.03193957731127739, + 0.03891618549823761, + 0.07339467108249664, + 0.034970540553331375 + ], + "credit_map": [ + 0.005328761413693428, + 0.00623314967378974, + 0.006341929081827402, + 0.010048381052911282, + 0.010604574345052242, + 0.00873190350830555, + 0.009124968200922012, + 0.007830937393009663, + 0.004230852704495192, + 0.006773340981453657, + 0.006447440013289452, + 0.005628760904073715, + 0.004255874082446098, + 0.0037025969941169024, + 0.004365886561572552, + 0.004771132953464985, + 0.0060616228729486465, + 0.0057196770794689655, + 0.008182683028280735, + 0.005828255787491798, + 0.0032624672167003155, + 0.008369014598429203, + 0.005498165730386972, + 0.004935310687869787, + 0.0016855569556355476, + 0.003496368182823062, + 0.003349351231008768, + 0.003279872005805373, + 0.002704733284190297, + 0.0030069544445723295, + 0.003280971897765994, + 0.004710270091891289, + 0.014740293845534325, + 0.005089010577648878, + 0.045067377388477325, + 0.011435814201831818, + 0.009491082280874252, + 0.007909427396953106, + 0.10779263824224472, + 0.0627349391579628, + 0.021314578130841255, + 0.025219716131687164, + 0.018939416855573654, + 0.0377584844827652, + 0.033858273178339005, + 0.044795360416173935, + 0.10781881213188171, + 0.08902197331190109, + 0.0027464169543236494, + 0.0034588249400258064, + 0.002508479403331876, + 0.003967933356761932, + 0.003704311326146126, + 0.0024827145971357822, + 0.006819508038461208, + 0.006251389626413584, + 0.003125503659248352, + 0.006378392688930035, + 0.004375500604510307, + 0.006020596716552973, + 0.004275931511074305, + 0.0037658405490219593, + 0.004381727892905474, + 0.0065926904790103436, + 0.006011958699673414, + 0.009000785648822784, + 0.006393468007445335, + 0.011721841059625149, + 0.017207946628332138, + 0.005007797386497259, + 0.01151940505951643, + 0.006531475577503443, + 0.0038202661089599133, + 0.003556201234459877, + 0.00315451854839921, + 0.0051947953179478645, + 0.003985458053648472, + 0.004921702202409506, + 0.004329807590693235, + 0.006007794756442308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.004119126591831446 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081226170063019 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011081226170063019, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081226170063019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03430516645312309, + 0.03338925540447235, + 0.023853816092014313, + 0.09406335651874542, + 0.5999062061309814, + 0.030076563358306885, + 0.08737828582525253, + 0.056564800441265106, + 0.01546479668468237, + 0.02499770186841488 + ], + "credit_map": [ + 0.0031913602724671364, + 0.0028323326259851456, + 0.006009920500218868, + 0.0060451701283454895, + 0.0036818350199609995, + 0.0025929678231477737, + 0.004632347729057074, + 0.005319234915077686, + 0.0020905141718685627, + 0.0028743459843099117, + 0.0040962290950119495, + 0.006183895748108625, + 0.0021614553406834602, + 0.0027491902001202106, + 0.010196286253631115, + 0.0030373409390449524, + 0.0023678976576775312, + 0.0037323711439967155, + 0.0036561216693371534, + 0.0034497736487537622, + 0.0026156571693718433, + 0.0024270787835121155, + 0.003043654141947627, + 0.0025612639728933573, + 0.0063169170171022415, + 0.014761284925043583, + 0.011632278561592102, + 0.017733817920088768, + 0.01028405036777258, + 0.009572912938892841, + 0.015947114676237106, + 0.007814987562596798, + 0.02794814482331276, + 0.009241476655006409, + 0.06641437113285065, + 0.023926354944705963, + 0.02013435959815979, + 0.01921028085052967, + 0.2551141083240509, + 0.17791715264320374, + 0.0018923014868050814, + 0.002992964582517743, + 0.0017170790815725923, + 0.0031688841991126537, + 0.0027945355977863073, + 0.003354214131832123, + 0.007482367102056742, + 0.006674219388514757, + 0.00889135804027319, + 0.013702057301998138, + 0.008177649229764938, + 0.014821690507233143, + 0.011526549234986305, + 0.008107299916446209, + 0.01322789303958416, + 0.00892380066215992, + 0.004655064083635807, + 0.008723495528101921, + 0.007049231324344873, + 0.007583872880786657, + 0.00712523702532053, + 0.006355721037834883, + 0.006427247542887926, + 0.00864493940025568, + 0.001987611874938011, + 0.00325511721894145, + 0.0012088201474398375, + 0.002296335995197296, + 0.002116111107170582, + 0.0011663446202874184, + 0.0024037130642682314, + 0.0010307435877621174, + 0.0021290997974574566, + 0.0028776624239981174, + 0.003430354641750455, + 0.0035602422431111336, + 0.002616386627778411, + 0.00448518106713891, + 0.0025326544418931007, + 0.0033661234192550182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.011081226170063019 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493319779634476 + ], + "opposite_logprob_tokens": [ + -2.9299333095550537 + ], + "logprob_reward": -0.05493319779634476, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493319779634476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20697398483753204, + 0.1227949783205986, + 0.04938637837767601, + 0.09280204772949219, + 0.18889755010604858, + 0.0936744436621666, + 0.08262892067432404, + 0.06507346034049988, + 0.04290536046028137, + 0.054862894117832184 + ], + "credit_map": [ + 0.026012668386101723, + 0.02172810211777687, + 0.022046441212296486, + 0.0255266260355711, + 0.029486054554581642, + 0.028552837669849396, + 0.026742076501250267, + 0.026879170909523964, + 0.010828673839569092, + 0.01404508389532566, + 0.02297508902847767, + 0.013515681028366089, + 0.011690287850797176, + 0.010655030608177185, + 0.027533775195479393, + 0.01155135128647089, + 0.005346063058823347, + 0.007654855959117413, + 0.011510168202221394, + 0.005642398726195097, + 0.005096542648971081, + 0.005075598601251841, + 0.003977905493229628, + 0.005082839634269476, + 0.004416930954903364, + 0.005431607831269503, + 0.01108535099774599, + 0.009786638431251049, + 0.006700115278363228, + 0.01238175667822361, + 0.022822603583335876, + 0.02017703279852867, + 0.008603179827332497, + 0.0034282607957720757, + 0.0172630213201046, + 0.008158081211149693, + 0.006548100151121616, + 0.0062559135258197784, + 0.08051210641860962, + 0.05812885984778404, + 0.006262735463678837, + 0.007191091310232878, + 0.00693069351837039, + 0.012123375199735165, + 0.011398270726203918, + 0.014605298638343811, + 0.018319016322493553, + 0.01684396341443062, + 0.005122852977365255, + 0.004963692743331194, + 0.008263269439339638, + 0.010691075585782528, + 0.01049668900668621, + 0.010061605833470821, + 0.02227954752743244, + 0.010750186629593372, + 0.0045605916529893875, + 0.007247853092849255, + 0.005938942544162273, + 0.008380064740777016, + 0.012135052122175694, + 0.006667728070169687, + 0.007939128205180168, + 0.012204097583889961, + 0.005203422158956528, + 0.007204855792224407, + 0.004358302801847458, + 0.00663525378331542, + 0.005763834808021784, + 0.0036914607044309378, + 0.005996013525873423, + 0.0040522171184420586, + 0.004675589501857758, + 0.004607470240443945, + 0.0050491164438426495, + 0.005242127925157547, + 0.0055387080647051334, + 0.008871386758983135, + 0.007027771789580584, + 0.013850720599293709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.05493319779634476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080636642873287 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080636642873287, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080636642873287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1389027088880539, + 0.06431416422128677, + 0.03761934116482735, + 0.03983248025178909, + 0.48640766739845276, + 0.09608334302902222, + 0.03480716049671173, + 0.053065426647663116, + 0.034448422491550446, + 0.014519237913191319 + ], + "credit_map": [ + 0.01382032223045826, + 0.019003894180059433, + 0.013151435181498528, + 0.020630573853850365, + 0.020760739222168922, + 0.01894976757466793, + 0.015597363002598286, + 0.01698862388730049, + 0.005324792582541704, + 0.00394844776019454, + 0.011354814283549786, + 0.008962067775428295, + 0.00462113693356514, + 0.003896119073033333, + 0.020964622497558594, + 0.0052421726286411285, + 0.004164519254118204, + 0.0038801126647740602, + 0.007606782950460911, + 0.004684059880673885, + 0.004336318001151085, + 0.004571586847305298, + 0.003548566484823823, + 0.004827399738132954, + 0.00236911466345191, + 0.0036567782517522573, + 0.004092406015843153, + 0.00712954206392169, + 0.0029621701687574387, + 0.004586867522448301, + 0.008305624127388, + 0.00672998046502471, + 0.021695179864764214, + 0.006849560420960188, + 0.04537001624703407, + 0.019663674756884575, + 0.016368327662348747, + 0.016983244568109512, + 0.20877718925476074, + 0.15070053935050964, + 0.00677265040576458, + 0.00828910619020462, + 0.007229783106595278, + 0.008414536714553833, + 0.009075216948986053, + 0.014864853583276272, + 0.021100124344229698, + 0.020337074995040894, + 0.0023041218519210815, + 0.002966269850730896, + 0.00306776468642056, + 0.005824059247970581, + 0.00557350181043148, + 0.003838077886030078, + 0.006693160627037287, + 0.0045402091927826405, + 0.0038313381373882294, + 0.005853225477039814, + 0.005405341740697622, + 0.007148542907088995, + 0.012716324999928474, + 0.004768819082528353, + 0.007863947190344334, + 0.005477889906615019, + 0.0037342680152505636, + 0.0058739944361150265, + 0.003262655809521675, + 0.004982868209481239, + 0.005823494400829077, + 0.003423641435801983, + 0.0038846044335514307, + 0.0034629006404429674, + 0.0021130028180778027, + 0.002202404662966728, + 0.001343044568784535, + 0.0023555653169751167, + 0.0015279606450349092, + 0.001357594970613718, + 0.0018087259959429502, + 0.0018109404481947422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.011080636642873287 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An ultra-wide-angle lens, starting from a slightly elevated perspective, slowly and steadily sweeps from left to right, revealing the courtyard of an ancient Taoist temple shrouded in morning mist. Gray eaves and angled corners appear faintly through the milky white haze, while the damp bluestone floor reflects soft, diffuse sunlight. A few golden ginkgo leaves are scattered on the ground. In the center of the courtyard, an elderly man with white hair and beard, a gaunt face, clad in a loose, flowing, pure white silk Tai Chi suit, his graying hair pulled into a neat bun, is absorbed in practicing Tai Chi. As the camera pans, he steadily shifts his weight, his arms spreading out like flowing water, forming a classic \"white crane spreads its wings\" pose. His wide sleeves flutter gently, and his expression, with closed eyes and a focused expression, reveals a tranquility that blends with nature. The entire picture has a cool tone, dominated by the gray of the stone, the dark color of the wood and the white of the characters, creating an atmosphere of extreme tranquility, concentration and full of oriental philosophical wisdom. The air seems to be filled with moist soil and a faint scent of sandalwood.", + "seed": 5072, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_4/index_2044_seed_5072_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.024957272922620177, + "vlm_logprob": -0.024957272922620177, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.024957272922620177, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06606660038232803 + ], + "opposite_logprob_tokens": [ + -2.8160665035247803 + ], + "logprob_reward": -0.06606660038232803, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06606660038232803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17675501108169556, + 0.039586033672094345, + 0.15945762395858765, + 0.08818400651216507, + 0.12674957513809204, + 0.08651269972324371, + 0.14592373371124268, + 0.05130719393491745, + 0.04745636507868767, + 0.07806769758462906 + ], + "credit_map": [ + 0.020513324066996574, + 0.022793350741267204, + 0.022482696920633316, + 0.021602407097816467, + 0.029885534197092056, + 0.013890421018004417, + 0.024154648184776306, + 0.021432632580399513, + 0.0036212559789419174, + 0.006593383848667145, + 0.005732948891818523, + 0.006555297877639532, + 0.004692124202847481, + 0.0031578908674418926, + 0.005226050969213247, + 0.00400708569213748, + 0.014817814342677593, + 0.017761720344424248, + 0.020223325118422508, + 0.02375352755188942, + 0.025903815403580666, + 0.01698428764939308, + 0.0221993550658226, + 0.017813799902796745, + 0.007922641932964325, + 0.012317754328250885, + 0.012620413675904274, + 0.011679697781801224, + 0.014281876385211945, + 0.006367343477904797, + 0.013309652917087078, + 0.009684637188911438, + 0.010885500349104404, + 0.011827170848846436, + 0.01402370911091566, + 0.015100494027137756, + 0.019426902756094933, + 0.014136288315057755, + 0.015401300974190235, + 0.02594822458922863, + 0.015211653895676136, + 0.011483771726489067, + 0.01135213952511549, + 0.011597023345530033, + 0.009083734825253487, + 0.008903459645807743, + 0.008722182363271713, + 0.010158742778003216, + 0.013343281112611294, + 0.022458117455244064, + 0.028507083654403687, + 0.02590477466583252, + 0.015768317505717278, + 0.010543068870902061, + 0.017067821696400642, + 0.012331273406744003, + 0.004203867167234421, + 0.005336328875273466, + 0.01044515985995531, + 0.01474766619503498, + 0.0031090439297258854, + 0.0035152093041688204, + 0.007437090389430523, + 0.0025128331035375595, + 0.006806327495723963, + 0.0065752011723816395, + 0.009915863163769245, + 0.005338291637599468, + 0.006205067038536072, + 0.0046409429050982, + 0.004692151676863432, + 0.0032825227826833725, + 0.010046512819826603, + 0.014292187988758087, + 0.012297037057578564, + 0.013826058246195316, + 0.006148854270577431, + 0.006532437168061733, + 0.005288529209792614, + 0.009636086411774158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.06606660038232803 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003208491252735257 + ], + "opposite_logprob_tokens": [ + -5.753208637237549 + ], + "logprob_reward": -0.003208491252735257, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003208491252735257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17930887639522552, + 0.06223967671394348, + 0.2560390830039978, + 0.10442133247852325, + 0.07847901433706284, + 0.06453133374452591, + 0.06508554518222809, + 0.07171807438135147, + 0.035580843687057495, + 0.08259619772434235 + ], + "credit_map": [ + 0.013764436356723309, + 0.02239098958671093, + 0.021088965237140656, + 0.02463657222688198, + 0.033497605472803116, + 0.017002258449792862, + 0.021208470687270164, + 0.02571958117187023, + 0.006002176087349653, + 0.007918065413832664, + 0.011462753638625145, + 0.006246975157409906, + 0.00865507498383522, + 0.0056232307106256485, + 0.008442491292953491, + 0.007888915948569775, + 0.023585813120007515, + 0.027932770550251007, + 0.03585471957921982, + 0.039505310356616974, + 0.03868186101317406, + 0.028292711824178696, + 0.033556293696165085, + 0.028629634529352188, + 0.008826269768178463, + 0.01478527020663023, + 0.017969155684113503, + 0.013922342099249363, + 0.015597536228597164, + 0.008371300995349884, + 0.014511970803141594, + 0.010437498800456524, + 0.007519997656345367, + 0.008304250426590443, + 0.008787542581558228, + 0.010604334063827991, + 0.013461566530168056, + 0.007852911949157715, + 0.010839235037565231, + 0.011109184473752975, + 0.00668759411200881, + 0.006998550612479448, + 0.011770649813115597, + 0.008753394708037376, + 0.012347856536507607, + 0.005075755994766951, + 0.008288980461657047, + 0.004608559887856245, + 0.006101755425333977, + 0.009067676030099392, + 0.009785723872482777, + 0.01261421199887991, + 0.007125935051590204, + 0.005996311083436012, + 0.008061490021646023, + 0.006332443095743656, + 0.006724291015416384, + 0.00794986542314291, + 0.015255671925842762, + 0.018108488991856575, + 0.005723894573748112, + 0.003745164256542921, + 0.009984026663005352, + 0.00422667944803834, + 0.0054285526275634766, + 0.0039266543462872505, + 0.006030823104083538, + 0.003745413850992918, + 0.006228066049516201, + 0.0028933752328157425, + 0.00400102324783802, + 0.003326940583065152, + 0.008298088796436787, + 0.012976132333278656, + 0.01835761032998562, + 0.013507203198969364, + 0.007861197926104069, + 0.007511145900934935, + 0.006736591923981905, + 0.007348233833909035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.003208491252735257 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630960550159216 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630960550159216, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630960550159216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09750847518444061, + 0.0343003012239933, + 0.09120962023735046, + 0.049001991748809814, + 0.05767272040247917, + 0.056728821247816086, + 0.10124260187149048, + 0.37294137477874756, + 0.051792211830616, + 0.08760188519954681 + ], + "credit_map": [ + 0.007041573524475098, + 0.011524076573550701, + 0.010441029444336891, + 0.010759202763438225, + 0.020827580243349075, + 0.008888806216418743, + 0.013677140697836876, + 0.014349064789712429, + 0.003174138255417347, + 0.004151481203734875, + 0.004613173194229603, + 0.006388736888766289, + 0.004275578539818525, + 0.0031921016052365303, + 0.004762266296893358, + 0.0037428238429129124, + 0.008335416205227375, + 0.009904413484036922, + 0.015021403320133686, + 0.014479742385447025, + 0.014652516692876816, + 0.010713411495089531, + 0.010619605891406536, + 0.007483114022761583, + 0.003990813624113798, + 0.00675149355083704, + 0.005307365674525499, + 0.0073715937323868275, + 0.008386598899960518, + 0.00303464662283659, + 0.009186153300106525, + 0.0049733263440430164, + 0.005618068855255842, + 0.005884827580302954, + 0.009246177971363068, + 0.009697838686406612, + 0.008172689005732536, + 0.004070613998919725, + 0.01032187044620514, + 0.004660631064325571, + 0.006074914708733559, + 0.00816552247852087, + 0.007962812669575214, + 0.00828312337398529, + 0.004998886492103338, + 0.005739862099289894, + 0.006369392853230238, + 0.009134308435022831, + 0.009936498478055, + 0.010853279381990433, + 0.01583675667643547, + 0.02549133263528347, + 0.00787421315908432, + 0.005996129475533962, + 0.017300521954894066, + 0.007953870110213757, + 0.02895553596317768, + 0.03786604106426239, + 0.07622947543859482, + 0.10268250852823257, + 0.030667852610349655, + 0.014952081255614758, + 0.056394148617982864, + 0.025193726643919945, + 0.00641342019662261, + 0.004391450434923172, + 0.011184437200427055, + 0.007546934299170971, + 0.006410561036318541, + 0.0033420072868466377, + 0.006483351346105337, + 0.006020049564540386, + 0.006703602150082588, + 0.010608485899865627, + 0.019659383222460747, + 0.005698519758880138, + 0.013029171153903008, + 0.0145562794059515, + 0.010119927115738392, + 0.0072265141643583775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.007630960550159216 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033682309091091156 + ], + "opposite_logprob_tokens": [ + -3.408682346343994 + ], + "logprob_reward": -0.033682309091091156, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033682309091091156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17520193755626678, + 0.12238302081823349, + 0.09137799590826035, + 0.216929093003273, + 0.07693767547607422, + 0.09379919618368149, + 0.12091187387704849, + 0.05111943557858467, + 0.0271547120064497, + 0.02418512850999832 + ], + "credit_map": [ + 0.019817417487502098, + 0.022180454805493355, + 0.023810626938939095, + 0.0245834868401289, + 0.025463607162237167, + 0.017076663672924042, + 0.02325647696852684, + 0.019013186916708946, + 0.012745353393256664, + 0.022709820419549942, + 0.0239018015563488, + 0.015773026272654533, + 0.013907112181186676, + 0.009242425672709942, + 0.014378412626683712, + 0.009725058451294899, + 0.00659357151016593, + 0.009137053973972797, + 0.012171369977295399, + 0.013587171211838722, + 0.011651408858597279, + 0.010203489102423191, + 0.014282843098044395, + 0.013751083984971046, + 0.024311166256666183, + 0.03188357874751091, + 0.036039799451828, + 0.026685252785682678, + 0.029008373618125916, + 0.020631039515137672, + 0.024045830592513084, + 0.02432403527200222, + 0.006346156354993582, + 0.008205139078199863, + 0.010722490958869457, + 0.01343535166233778, + 0.010370573960244656, + 0.007057570852339268, + 0.011037487536668777, + 0.009762902744114399, + 0.016181666404008865, + 0.015386455692350864, + 0.01647213287651539, + 0.010557970963418484, + 0.013151977211236954, + 0.00682074436917901, + 0.008748809807002544, + 0.006479435600340366, + 0.01196193415671587, + 0.01320680882781744, + 0.016214611008763313, + 0.022059647366404533, + 0.012720846571028233, + 0.010690752416849136, + 0.02029898576438427, + 0.013758279383182526, + 0.005223792977631092, + 0.006847409997135401, + 0.007245221640914679, + 0.009913992136716843, + 0.005686082411557436, + 0.0036493942607194185, + 0.0072149066254496574, + 0.005338632967323065, + 0.0031532940920442343, + 0.002645625500008464, + 0.004824345000088215, + 0.004475332330912352, + 0.003939757123589516, + 0.0019069383852183819, + 0.003262640442699194, + 0.0029467775020748377, + 0.0024838612880557775, + 0.003910311497747898, + 0.002511453814804554, + 0.0034384564496576786, + 0.00421827333047986, + 0.0035804554354399443, + 0.0020849162247031927, + 0.001957397907972336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.033682309091091156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the courtyard of an ancient Taoist temple shrouded in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with white hair and beard wearing a loose, flowing, pure white silk Tai Chi suit visible in the center of the courtyard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing the 'white crane spreads its wings' Tai Chi pose with his arms spreading out and wide sleeves fluttering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a slightly elevated perspective and slowly sweep from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are gray eaves, angled corners, and a damp bluestone floor reflecting soft sunlight visible through the mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198003336787224 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014198003336787224, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198003336787224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1282288134098053, + 0.08826428651809692, + 0.12979549169540405, + 0.0944768637418747, + 0.13402220606803894, + 0.08722864091396332, + 0.09901146590709686, + 0.07222463190555573, + 0.07804923504590988, + 0.08869834244251251 + ], + "credit_map": [ + 0.010588666424155235, + 0.01599060371518135, + 0.014252147637307644, + 0.015463539399206638, + 0.027196211740374565, + 0.01344231702387333, + 0.01393247302621603, + 0.01736285164952278, + 0.007084098644554615, + 0.008590317331254482, + 0.017120255157351494, + 0.01275835931301117, + 0.012557019479572773, + 0.01115728635340929, + 0.009771226905286312, + 0.009225725196301937, + 0.011835645884275436, + 0.014811582863330841, + 0.017700890079140663, + 0.01980636641383171, + 0.021615739911794662, + 0.015016580931842327, + 0.01629328913986683, + 0.012715386226773262, + 0.00889625959098339, + 0.008815394714474678, + 0.012608744204044342, + 0.011236436665058136, + 0.016794683411717415, + 0.005795310717076063, + 0.016992054879665375, + 0.01333798561245203, + 0.010100152343511581, + 0.011846957728266716, + 0.019868405535817146, + 0.01564759388566017, + 0.015619922429323196, + 0.014477595686912537, + 0.014463488012552261, + 0.03199809417128563, + 0.01558074913918972, + 0.015009772032499313, + 0.011728594079613686, + 0.009750002063810825, + 0.008942686952650547, + 0.011621551588177681, + 0.0075337826274335384, + 0.007061503361910582, + 0.009203922934830189, + 0.012331690639257431, + 0.011810838244855404, + 0.026062093675136566, + 0.008743597194552422, + 0.005976980086416006, + 0.01834697835147381, + 0.006535361520946026, + 0.007638041861355305, + 0.005692115519195795, + 0.016097292304039, + 0.014644747599959373, + 0.005386214703321457, + 0.007234762888401747, + 0.009574521332979202, + 0.0059569356963038445, + 0.010912491008639336, + 0.005952534731477499, + 0.02362293004989624, + 0.00753528717905283, + 0.006482559721916914, + 0.008442986756563187, + 0.00963718630373478, + 0.005463257431983948, + 0.012422717176377773, + 0.02077806368470192, + 0.010901366360485554, + 0.016669753938913345, + 0.007163955830037594, + 0.008663994260132313, + 0.005649195052683353, + 0.006449297070503235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.014198003336787224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_5/index_tensor([3576, 3576])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_5/index_tensor([3576, 3576])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..caabfc9d25a26aa5c2c1175dfea15cdc7ae99a44 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_5/index_tensor([3576, 3576])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a chaotic handheld camera shot imbued with the iconic aesthetic of '80s anime, the frame shakes violently with the pounding wind, struggling to focus on a lone Sophora japonica tree. It's the epicenter of a ferocious storm, the sky churning into a maelstrom of deep purple and ominous magenta, while rough, high-contrast brushstrokes fill the background. The ancient Sophora japonica tree waged a desperate struggle, its stout trunk bent stubbornly to the brink of breaking under the relentless force of the hurricane, its twisted branches lashing wildly at the air like struggling arms. Torrential rain streaked across the frame like dense, white slashes, lashing the tree's gnarled, deeply cracked bark with a wet gleam in the dim light. Every shake of the camera captured the vivid details of the water coursing along the veins. Dark green leaves were ruthlessly torn off and danced wildly in the chaotic air currents. A dazzling flash of lightning instantly tore through the sky, illuminating this doomsday scene full of dynamic tension with a brief pale light, creating a primitive, intense and tragic visual atmosphere.", + "seed": 2075, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_5/index_3576_seed_2075_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.6272812757641077, + "vlm_logprob": -0.6272812757641077, + "vlm_yes_no_margin": 1.575, + "vlm_token_credit": -0.6272812757641077, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.6983649730682373 + ], + "opposite_logprob_tokens": [ + -0.0733649879693985 + ], + "logprob_reward": -2.6983649730682373, + "yes_no_margin_reward": -2.625, + "token_credit_reward": -2.6983649730682373, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.47170141339302063, + 0.08020380139350891, + 0.023178840056061745, + 0.024763839319348335, + 0.0290476493537426, + 0.014472892507910728, + 0.05238773673772812, + 0.02406277507543564, + 0.1321965903043747, + 0.14798448979854584 + ], + "credit_map": [ + 0.04203970730304718, + 0.045968178659677505, + 0.13361266255378723, + 0.032072413712739944, + 0.09724898636341095, + 0.03946245461702347, + 0.021578796207904816, + 0.05971820279955864, + 0.014264373108744621, + 0.008753110654652119, + 0.011346925050020218, + 0.01589340716600418, + 0.00679343705996871, + 0.00796415749937296, + 0.006700032856315374, + 0.008488362655043602, + 0.003004444297403097, + 0.002575518563389778, + 0.003580802585929632, + 0.0038110220339149237, + 0.002369508147239685, + 0.00290083815343678, + 0.0025285156443715096, + 0.0024081901647150517, + 0.0035647910553961992, + 0.0032946234568953514, + 0.0034946314990520477, + 0.004022006411105394, + 0.0024971081875264645, + 0.0033727327827364206, + 0.002172674983739853, + 0.0023452723398804665, + 0.0030654289294034243, + 0.0036153439432382584, + 0.00509983254596591, + 0.0038332792464643717, + 0.005202324595302343, + 0.0039455764926970005, + 0.0020467054564505816, + 0.0022391576785594225, + 0.002147211693227291, + 0.0011566956527531147, + 0.0025341298896819353, + 0.0018670677673071623, + 0.0012085363268852234, + 0.0017479141242802143, + 0.0018410549964755774, + 0.001970282755792141, + 0.006523385178297758, + 0.005219972692430019, + 0.00939621590077877, + 0.0074364491738379, + 0.004353106953203678, + 0.005161984357982874, + 0.008499875664710999, + 0.005796747747808695, + 0.004094260279089212, + 0.0025738365948200226, + 0.004032236058264971, + 0.004608118906617165, + 0.0017907485598698258, + 0.0027840863913297653, + 0.001616601599380374, + 0.0025628875009715557, + 0.011197582818567753, + 0.011842547915875912, + 0.034041717648506165, + 0.013262390159070492, + 0.027507781982421875, + 0.011573642492294312, + 0.008832448162138462, + 0.013938479125499725, + 0.033228643238544464, + 0.021468818187713623, + 0.019753841683268547, + 0.014182302169501781, + 0.01347337756305933, + 0.012442112900316715, + 0.013047534972429276, + 0.02038785070180893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -2.6983649730682373 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03377151861786842 + ], + "opposite_logprob_tokens": [ + -3.408771514892578 + ], + "logprob_reward": -0.03377151861786842, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03377151861786842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5818661451339722, + 0.05946352332830429, + 0.029985353350639343, + 0.020824873819947243, + 0.03740135207772255, + 0.02765108272433281, + 0.02284826897084713, + 0.04364778473973274, + 0.07452776283025742, + 0.1017838642001152 + ], + "credit_map": [ + 0.05629643797874451, + 0.04861985892057419, + 0.17472504079341888, + 0.03857000172138214, + 0.09997531026601791, + 0.047903526574373245, + 0.03270299360156059, + 0.08307298272848129, + 0.00675942562520504, + 0.005316578783094883, + 0.013359101489186287, + 0.006399850361049175, + 0.004676864482462406, + 0.010313545353710651, + 0.004122142679989338, + 0.008516013622283936, + 0.0034700785763561726, + 0.0034647465217858553, + 0.004859200678765774, + 0.004527384415268898, + 0.0029413001611828804, + 0.003639103379100561, + 0.003285916754975915, + 0.0037976240273565054, + 0.003280313452705741, + 0.0024667552206665277, + 0.0035238435957580805, + 0.002353122690692544, + 0.001934928703121841, + 0.002420505741611123, + 0.0021974085830152035, + 0.0026479975786060095, + 0.0033588199876248837, + 0.003978328313678503, + 0.009189698845148087, + 0.0023074178025126457, + 0.008886227384209633, + 0.004291142337024212, + 0.0018326034769415855, + 0.003557112766429782, + 0.0038423631340265274, + 0.003058577887713909, + 0.005902566947042942, + 0.0032585037406533957, + 0.001862361910752952, + 0.003200514940544963, + 0.0028423380572348833, + 0.0036838543601334095, + 0.002222767099738121, + 0.00349818030372262, + 0.002837304724380374, + 0.0037482099141925573, + 0.002379829529672861, + 0.0035089654847979546, + 0.002030397532507777, + 0.0026226146146655083, + 0.005908517632633448, + 0.003865613602101803, + 0.010137239471077919, + 0.005732941906899214, + 0.0025733995717018843, + 0.004244276788085699, + 0.00430689612403512, + 0.0068789031356573105, + 0.006379620637744665, + 0.006521185394376516, + 0.02010120265185833, + 0.005966729950159788, + 0.01700447127223015, + 0.007201214320957661, + 0.0039970590732991695, + 0.007356278598308563, + 0.02017456851899624, + 0.01564989611506462, + 0.01597633771598339, + 0.012755115516483784, + 0.012524086982011795, + 0.009486692026257515, + 0.006515109445899725, + 0.008702063001692295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.03377151861786842 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201468512415886 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201468512415886, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201468512415886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6014775633811951, + 0.09315444529056549, + 0.03955100104212761, + 0.022968987002968788, + 0.02036971226334572, + 0.013928833417594433, + 0.017709553241729736, + 0.021420270204544067, + 0.030394183471798897, + 0.1390254646539688 + ], + "credit_map": [ + 0.0549292266368866, + 0.05816083028912544, + 0.18495489656925201, + 0.030739983543753624, + 0.1305873692035675, + 0.045205216854810715, + 0.01880268193781376, + 0.07809733599424362, + 0.011847405694425106, + 0.009376228787004948, + 0.019621344283223152, + 0.009627566672861576, + 0.007680185604840517, + 0.016096243634819984, + 0.007680072449147701, + 0.011225398629903793, + 0.0044087800197303295, + 0.003915421664714813, + 0.007535077165812254, + 0.004887581802904606, + 0.003129808232188225, + 0.004838518798351288, + 0.00375413428992033, + 0.007081679999828339, + 0.0023266086354851723, + 0.003281372832134366, + 0.003950781188905239, + 0.0023837436456233263, + 0.0021772338077425957, + 0.0028992064762860537, + 0.0023257818538695574, + 0.003624256933107972, + 0.002757337177172303, + 0.0022893669083714485, + 0.004252617713063955, + 0.0020565595477819443, + 0.0027516414411365986, + 0.0021100800950080156, + 0.0023577306419610977, + 0.0017943778075277805, + 0.0019848381634801626, + 0.0012645877432078123, + 0.0023604820016771555, + 0.0019597006030380726, + 0.0011520993430167437, + 0.0016724059823900461, + 0.0017707280348986387, + 0.0017639921279624104, + 0.002178949536755681, + 0.0028953312430530787, + 0.0028895926661789417, + 0.0017373127629980445, + 0.0019218422239646316, + 0.002358105732128024, + 0.0015861898427829146, + 0.002142228651791811, + 0.0024825141299515963, + 0.0021544902119785547, + 0.005999880377203226, + 0.0018452961230650544, + 0.001476799137890339, + 0.0023616377729922533, + 0.001557421637699008, + 0.0035422316286712885, + 0.0036545763723552227, + 0.00211904663592577, + 0.0076071713119745255, + 0.003440611995756626, + 0.0047908928245306015, + 0.0027526025660336018, + 0.002428183564916253, + 0.0036010979674756527, + 0.03378821536898613, + 0.02157745324075222, + 0.01603371649980545, + 0.019116109237074852, + 0.01743488200008869, + 0.012057967483997345, + 0.007495493162423372, + 0.011521629057824612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.018201468512415886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16043877601623535 + ], + "opposite_logprob_tokens": [ + -1.9104387760162354 + ], + "logprob_reward": -0.16043877601623535, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16043877601623535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24987782537937164, + 0.08099063485860825, + 0.028590580448508263, + 0.024393009021878242, + 0.10161486268043518, + 0.07201491296291351, + 0.03967532142996788, + 0.0326995849609375, + 0.04956410825252533, + 0.3205791711807251 + ], + "credit_map": [ + 0.019409259781241417, + 0.02869756519794464, + 0.04691635072231293, + 0.023350728675723076, + 0.05249384418129921, + 0.025796374306082726, + 0.029279878363013268, + 0.023933835327625275, + 0.013016321696341038, + 0.008968944661319256, + 0.014273811131715775, + 0.010663415305316448, + 0.007323765195906162, + 0.012564430944621563, + 0.005867855157703161, + 0.008312082849442959, + 0.0035889942664653063, + 0.003403106238692999, + 0.003980386070907116, + 0.00444050133228302, + 0.0025263759307563305, + 0.0032721790485084057, + 0.0035338320303708315, + 0.0038452078588306904, + 0.003077314468100667, + 0.003539394587278366, + 0.003877246519550681, + 0.0030657537281513214, + 0.0025620346423238516, + 0.004062224179506302, + 0.0019082075450569391, + 0.002300834748893976, + 0.010794652625918388, + 0.009729058481752872, + 0.029348941519856453, + 0.008691971190273762, + 0.014558781869709492, + 0.009154361672699451, + 0.007380913011729717, + 0.01195617113262415, + 0.00990966334939003, + 0.007671330589801073, + 0.011244770139455795, + 0.00985290203243494, + 0.005800572689622641, + 0.008672300726175308, + 0.009975624270737171, + 0.00888774823397398, + 0.0044270833022892475, + 0.005193337332457304, + 0.0067323860712349415, + 0.004509091842919588, + 0.003686064388602972, + 0.004010117147117853, + 0.006476757116615772, + 0.0046404823660850525, + 0.00353311188519001, + 0.004472291097044945, + 0.0058414386585354805, + 0.0037042023614048958, + 0.0029486443381756544, + 0.004556187894195318, + 0.0031609053257852793, + 0.004482802469283342, + 0.004587022587656975, + 0.005330739542841911, + 0.012320682406425476, + 0.004316383972764015, + 0.010817913338541985, + 0.0051520518027246, + 0.0025727541651576757, + 0.004466562997549772, + 0.05633733049035072, + 0.03717771917581558, + 0.0546501949429512, + 0.027286911383271217, + 0.03564675152301788, + 0.03680671751499176, + 0.030482610687613487, + 0.04219095781445503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.16043877601623535 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562964260578156 + ], + "opposite_logprob_tokens": [ + -1.6006296873092651 + ], + "logprob_reward": -0.22562964260578156, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562964260578156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8169805407524109, + 0.023499686270952225, + 0.007944769226014614, + 0.013488671742379665, + 0.008893042802810669, + 0.006378881633281708, + 0.014633544720709324, + 0.01625797711312771, + 0.012218164280056953, + 0.07970473170280457 + ], + "credit_map": [ + 0.07572665810585022, + 0.07828273624181747, + 0.26001057028770447, + 0.035914815962314606, + 0.17837519943714142, + 0.05592355132102966, + 0.022182542830705643, + 0.1105644553899765, + 0.004374480340629816, + 0.0026449472643435, + 0.0039246780797839165, + 0.0025981978978961706, + 0.0020132679492235184, + 0.0023932093754410744, + 0.0023867960553616285, + 0.003164108842611313, + 0.001172001357190311, + 0.0007866595406085253, + 0.00110315578058362, + 0.0012748283334076405, + 0.0008214404224418104, + 0.0009218541090376675, + 0.0010654921643435955, + 0.000799337518401444, + 0.0020776595920324326, + 0.0017451583407819271, + 0.0018224039813503623, + 0.001729160430841148, + 0.001278541749343276, + 0.0018707399722188711, + 0.0014890979509800673, + 0.0014759096084162593, + 0.0009665166726335883, + 0.0010793794644996524, + 0.0017761652125045657, + 0.0010936958715319633, + 0.0012694122269749641, + 0.0010595874628052115, + 0.0006332345074042678, + 0.0010150515008717775, + 0.0007519128848798573, + 0.0005571527290157974, + 0.0013753378298133612, + 0.0010799415176734328, + 0.00041397844324819744, + 0.0006574439466930926, + 0.000519977358635515, + 0.0010231371270492673, + 0.00158635841216892, + 0.0018909808713942766, + 0.0019708324689418077, + 0.0020598256960511208, + 0.001446134876459837, + 0.0016369873192161322, + 0.002433598041534424, + 0.001608828199096024, + 0.001664472627453506, + 0.0017635205294936895, + 0.002660051453858614, + 0.0021723699755966663, + 0.0013749458594247699, + 0.002202011877670884, + 0.001980036264285445, + 0.0024405664298683405, + 0.0011469711316749454, + 0.001467381021939218, + 0.002689588814973831, + 0.0014448873698711395, + 0.0021798769012093544, + 0.0012590192491188645, + 0.0008206464117392898, + 0.0012097933795303106, + 0.013402316719293594, + 0.012385750189423561, + 0.008613073267042637, + 0.011569339781999588, + 0.012284708209335804, + 0.008876300416886806, + 0.00601312518119812, + 0.006560115143656731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.22562964260578156 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a chaotic handheld camera shot imbued with the iconic aesthetic of '80s anime, the frame shakes violently with the pounding wind, struggling to focus on a lone Sophora japonica tree. It's the epicenter of a ferocious storm, the sky churning into a maelstrom of deep purple and ominous magenta, while rough, high-contrast brushstrokes fill the background. The ancient Sophora japonica tree waged a desperate struggle, its stout trunk bent stubbornly to the brink of breaking under the relentless force of the hurricane, its twisted branches lashing wildly at the air like struggling arms. Torrential rain streaked across the frame like dense, white slashes, lashing the tree's gnarled, deeply cracked bark with a wet gleam in the dim light. Every shake of the camera captured the vivid details of the water coursing along the veins. Dark green leaves were ruthlessly torn off and danced wildly in the chaotic air currents. A dazzling flash of lightning instantly tore through the sky, illuminating this doomsday scene full of dynamic tension with a brief pale light, creating a primitive, intense and tragic visual atmosphere.", + "seed": 2076, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_5/index_3576_seed_2076_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.6666951881721616, + "vlm_logprob": -0.6666951881721616, + "vlm_yes_no_margin": 2.025, + "vlm_token_credit": -0.6666951881721616, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -3.1723310947418213 + ], + "opposite_logprob_tokens": [ + -0.0473310686647892 + ], + "logprob_reward": -3.1723310947418213, + "yes_no_margin_reward": -3.125, + "token_credit_reward": -3.1723310947418213, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1356678009033203, + 0.03390870243310928, + 0.04197398200631142, + 0.019711168482899666, + 0.009357072412967682, + 0.010968336835503578, + 0.01800728775560856, + 0.01366584375500679, + 0.5650986433029175, + 0.15164121985435486 + ], + "credit_map": [ + 0.025949284434318542, + 0.013887675479054451, + 0.023897681385278702, + 0.019555605947971344, + 0.008712091483175755, + 0.013354580849409103, + 0.01647469773888588, + 0.013836178928613663, + 0.00541213434189558, + 0.0038727656938135624, + 0.004820161033421755, + 0.0035345060750842094, + 0.003701481269672513, + 0.004127363208681345, + 0.004159198608249426, + 0.004281089641153812, + 0.005436156876385212, + 0.00453416770324111, + 0.005650434177368879, + 0.004293649457395077, + 0.0035246992483735085, + 0.005711183417588472, + 0.006279261317104101, + 0.006544427014887333, + 0.003793667769059539, + 0.0017195094842463732, + 0.004262383095920086, + 0.0027577332220971584, + 0.0012179698096588254, + 0.0017742142081260681, + 0.002156112575903535, + 0.0020295754075050354, + 0.0010574074694886804, + 0.0010131722083315253, + 0.0014361668145284057, + 0.0012454577954486012, + 0.0007337850402109325, + 0.0008905649301595986, + 0.0014504624996334314, + 0.0015300544910132885, + 0.0016095517203211784, + 0.0010965017136186361, + 0.0015747661236673594, + 0.001248902641236782, + 0.0009957043221220374, + 0.0011940523982048035, + 0.001654616091400385, + 0.001594240078702569, + 0.0024339244700968266, + 0.0018334289779886603, + 0.002980997320264578, + 0.00237558176741004, + 0.0015316687058657408, + 0.0023169706109911203, + 0.0022871557157486677, + 0.0022475579753518105, + 0.0019860712345689535, + 0.0013175916392356157, + 0.002017499879002571, + 0.0015203198418021202, + 0.0012722313404083252, + 0.0010653469944372773, + 0.001987709663808346, + 0.002499071881175041, + 0.2123628407716751, + 0.022178782150149345, + 0.2389630675315857, + 0.03377891331911087, + 0.009578035213053226, + 0.012800132855772972, + 0.02033773623406887, + 0.01509903185069561, + 0.0306747704744339, + 0.014254326932132244, + 0.03135626018047333, + 0.020486067980527878, + 0.010887506417930126, + 0.014820598065853119, + 0.014016742818057537, + 0.015144934877753258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -3.1723310947418213 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04321148619055748 + ], + "opposite_logprob_tokens": [ + -3.1682114601135254 + ], + "logprob_reward": -0.04321148619055748, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04321148619055748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11306598037481308, + 0.07232574373483658, + 0.04807822406291962, + 0.02788858488202095, + 0.03269273415207863, + 0.027121160179376602, + 0.05570223182439804, + 0.05745454877614975, + 0.3308624029159546, + 0.23480838537216187 + ], + "credit_map": [ + 0.019838325679302216, + 0.009508916176855564, + 0.01913057081401348, + 0.01936856284737587, + 0.007607933133840561, + 0.01209079660475254, + 0.013537206687033176, + 0.011983661912381649, + 0.010310487821698189, + 0.00694880960509181, + 0.012089276686310768, + 0.007539353333413601, + 0.007072817534208298, + 0.007390408776700497, + 0.01016046293079853, + 0.010814128443598747, + 0.008564221672713757, + 0.004105747677385807, + 0.009494053199887276, + 0.004968846216797829, + 0.0035102202091366053, + 0.004104346036911011, + 0.005989535711705685, + 0.007341253571212292, + 0.005332916975021362, + 0.0029554166831076145, + 0.0066106063313782215, + 0.0030741465743631124, + 0.0023167552426457405, + 0.0019920573104172945, + 0.003084643045440316, + 0.0025220427196472883, + 0.0036978351417928934, + 0.003623037366196513, + 0.003780822269618511, + 0.004502760246396065, + 0.0037158976774662733, + 0.004029170610010624, + 0.004788269754499197, + 0.004554941784590483, + 0.003343562362715602, + 0.003001418663188815, + 0.003952362574636936, + 0.0034814791288226843, + 0.002844205591827631, + 0.002122039906680584, + 0.004235677421092987, + 0.004140413366258144, + 0.011143441312015057, + 0.004242925439029932, + 0.012649835087358952, + 0.004755537025630474, + 0.0036693811416625977, + 0.004252218641340733, + 0.007118665147572756, + 0.007870226167142391, + 0.009523056447505951, + 0.005209249444305897, + 0.010775825008749962, + 0.005461565218865871, + 0.004557711537927389, + 0.004199758172035217, + 0.008150629699230194, + 0.009576753713190556, + 0.11125525087118149, + 0.02068473957479, + 0.12950089573860168, + 0.021619148552417755, + 0.00889525841921568, + 0.010777122341096401, + 0.014279894530773163, + 0.013850078918039799, + 0.04232777655124664, + 0.023245420306921005, + 0.042426299303770065, + 0.034484073519706726, + 0.01531131099909544, + 0.025569738820195198, + 0.023271674290299416, + 0.028172075748443604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.04321148619055748 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01422197837382555 + ], + "opposite_logprob_tokens": [ + -4.264222145080566 + ], + "logprob_reward": -0.01422197837382555, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01422197837382555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10964734107255936, + 0.1268901824951172, + 0.05672706663608551, + 0.07745938003063202, + 0.026162385940551758, + 0.06961353868246078, + 0.03452443331480026, + 0.05140133202075958, + 0.30009159445762634, + 0.147482767701149 + ], + "credit_map": [ + 0.021520614624023438, + 0.00933147594332695, + 0.026025990024209023, + 0.010875137522816658, + 0.007003332953900099, + 0.011883466504514217, + 0.01131424866616726, + 0.011693074367940426, + 0.02718961052596569, + 0.011180179193615913, + 0.02768869698047638, + 0.013721275143325329, + 0.008864115923643112, + 0.01704913005232811, + 0.009586257860064507, + 0.0116109075024724, + 0.009652506560087204, + 0.005659563932567835, + 0.009221925400197506, + 0.006300811190158129, + 0.004423534497618675, + 0.007006779313087463, + 0.008054345846176147, + 0.006407597102224827, + 0.016773266717791557, + 0.006200372241437435, + 0.01914132945239544, + 0.007340498734265566, + 0.004975536372512579, + 0.006055100820958614, + 0.0072739385068416595, + 0.009699327871203423, + 0.005923716351389885, + 0.0035116851795464754, + 0.004173301160335541, + 0.0036148875951766968, + 0.00150838284753263, + 0.001835919450968504, + 0.0027227529790252447, + 0.0028717389795929193, + 0.011554612778127193, + 0.00868218019604683, + 0.012385714799165726, + 0.0070162867195904255, + 0.007240437436848879, + 0.0054771192371845245, + 0.009129896759986877, + 0.00812728051096201, + 0.00644829124212265, + 0.0032289205119013786, + 0.006974071264266968, + 0.004102768376469612, + 0.0025934225413948298, + 0.0027489266358315945, + 0.0033102333545684814, + 0.005117796361446381, + 0.009838653728365898, + 0.0048597101122140884, + 0.0106227807700634, + 0.003999016713351011, + 0.00413164496421814, + 0.003968322183936834, + 0.006286913529038429, + 0.007694287225604057, + 0.10775785893201828, + 0.015045383013784885, + 0.1154952272772789, + 0.01939644291996956, + 0.008540893904864788, + 0.01042527612298727, + 0.01324523612856865, + 0.010185259394347668, + 0.02239833027124405, + 0.01315983571112156, + 0.03400920704007149, + 0.018011746928095818, + 0.01596316695213318, + 0.012594705447554588, + 0.015181018970906734, + 0.016164740547537804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.01422197837382555 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055007778108119965 + ], + "opposite_logprob_tokens": [ + -2.9300076961517334 + ], + "logprob_reward": -0.055007778108119965, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055007778108119965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07430513948202133, + 0.034117307513952255, + 0.04775545001029968, + 0.021006111055612564, + 0.02899719402194023, + 0.01929938979446888, + 0.04734504967927933, + 0.02650674432516098, + 0.5525844097137451, + 0.14808325469493866 + ], + "credit_map": [ + 0.011327244341373444, + 0.008259803056716919, + 0.010868974030017853, + 0.010027267038822174, + 0.008283114060759544, + 0.008960581384599209, + 0.008727816864848137, + 0.007850337773561478, + 0.003557569347321987, + 0.0035756598226726055, + 0.005155754741281271, + 0.004788240883499384, + 0.0036757763009518385, + 0.005011189728975296, + 0.004115904215723276, + 0.004237215965986252, + 0.006206694059073925, + 0.0045328387059271336, + 0.007318203337490559, + 0.0044188471511006355, + 0.003419180866330862, + 0.005495357792824507, + 0.00837106816470623, + 0.007993259467184544, + 0.004624645691365004, + 0.0014148998307064176, + 0.005162864923477173, + 0.0021912301890552044, + 0.0014339712215587497, + 0.0015445512253791094, + 0.0024124085903167725, + 0.002221539616584778, + 0.005205944180488586, + 0.0019890153780579567, + 0.005322850309312344, + 0.002667888766154647, + 0.0021903151646256447, + 0.0022423104383051395, + 0.003681400790810585, + 0.005697471555322409, + 0.0026459144428372383, + 0.0020524715073406696, + 0.002031819662079215, + 0.002199947601184249, + 0.0015073333634063601, + 0.0025743155274540186, + 0.0028724228031933308, + 0.0034151640720665455, + 0.013644429855048656, + 0.002739747054874897, + 0.017058296129107475, + 0.0038228975608944893, + 0.0015283508691936731, + 0.0026515063364058733, + 0.0033198255114257336, + 0.0025799928698688745, + 0.005295857787132263, + 0.002943924628198147, + 0.0057709841057658195, + 0.0027671647258102894, + 0.0019875820726156235, + 0.002302444539964199, + 0.003029445419088006, + 0.00240934151224792, + 0.2053210288286209, + 0.023625951260328293, + 0.22842076420783997, + 0.031246419996023178, + 0.012196877039968967, + 0.013148018158972263, + 0.020528113469481468, + 0.018097221851348877, + 0.027935201302170753, + 0.015249041840434074, + 0.023184722289443016, + 0.021305471658706665, + 0.010553557425737381, + 0.018211811780929565, + 0.01628630794584751, + 0.015357146970927715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.055007778108119965 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870360344648361 + ], + "opposite_logprob_tokens": [ + -3.048703670501709 + ], + "logprob_reward": -0.04870360344648361, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870360344648361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057624876499176025, + 0.022073671221733093, + 0.01727028749883175, + 0.01475170161575079, + 0.008600073866546154, + 0.00820977333933115, + 0.022793244570493698, + 0.01911468245089054, + 0.4297862648963928, + 0.3997753858566284 + ], + "credit_map": [ + 0.010008256882429123, + 0.0044282651506364346, + 0.011547564528882504, + 0.00654916325584054, + 0.005195523612201214, + 0.007751004304736853, + 0.006909139920026064, + 0.005235962104052305, + 0.0046349563635885715, + 0.0018516475101932883, + 0.004926147870719433, + 0.002323441905900836, + 0.0016310347709804773, + 0.002144316677004099, + 0.0025187749415636063, + 0.002043350599706173, + 0.004627358168363571, + 0.000930583628360182, + 0.005423205904662609, + 0.0010647379094734788, + 0.0010162561666220427, + 0.0014721559127792716, + 0.001505781547166407, + 0.0012302083196118474, + 0.002586792455986142, + 0.0012703872052952647, + 0.003142704488709569, + 0.0014186766929924488, + 0.0011493625352159142, + 0.0017309235408902168, + 0.001379115041345358, + 0.0020737405866384506, + 0.0013400118332356215, + 0.0008130321511998773, + 0.0015878489939495921, + 0.0010034649167209864, + 0.0006575770676136017, + 0.0009972157422453165, + 0.001240151934325695, + 0.0009607715765014291, + 0.0012993330601602793, + 0.0007595241186209023, + 0.0014377807965502143, + 0.0009709647856652737, + 0.000668030115775764, + 0.000989490537904203, + 0.0008262307965196669, + 0.0012584197102114558, + 0.005535681266337633, + 0.001604935503564775, + 0.006568076554685831, + 0.0023218048736453056, + 0.0010556615889072418, + 0.0017678668955340981, + 0.0017595374956727028, + 0.0021796829532831907, + 0.004201729781925678, + 0.00211787736043334, + 0.004059114959090948, + 0.0019375726114958525, + 0.0013451235136017203, + 0.0016018247697502375, + 0.0019487603567540646, + 0.0019026818918064237, + 0.16417178511619568, + 0.014648359268903732, + 0.18223069608211517, + 0.02541111409664154, + 0.007935861125588417, + 0.009319133125245571, + 0.01584223099052906, + 0.010227113030850887, + 0.07308115065097809, + 0.03559223189949989, + 0.08327481150627136, + 0.05310146510601044, + 0.031236547976732254, + 0.038599707186222076, + 0.03908155858516693, + 0.04580793157219887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.04870360344648361 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a chaotic handheld camera shot imbued with the iconic aesthetic of '80s anime, the frame shakes violently with the pounding wind, struggling to focus on a lone Sophora japonica tree. It's the epicenter of a ferocious storm, the sky churning into a maelstrom of deep purple and ominous magenta, while rough, high-contrast brushstrokes fill the background. The ancient Sophora japonica tree waged a desperate struggle, its stout trunk bent stubbornly to the brink of breaking under the relentless force of the hurricane, its twisted branches lashing wildly at the air like struggling arms. Torrential rain streaked across the frame like dense, white slashes, lashing the tree's gnarled, deeply cracked bark with a wet gleam in the dim light. Every shake of the camera captured the vivid details of the water coursing along the veins. Dark green leaves were ruthlessly torn off and danced wildly in the chaotic air currents. A dazzling flash of lightning instantly tore through the sky, illuminating this doomsday scene full of dynamic tension with a brief pale light, creating a primitive, intense and tragic visual atmosphere.", + "seed": 2077, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_5/index_3576_seed_2077_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.619628370180726, + "vlm_logprob": -0.619628370180726, + "vlm_yes_no_margin": 1.375, + "vlm_token_credit": -0.619628370180726, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.580561399459839 + ], + "opposite_logprob_tokens": [ + -0.0805613324046135 + ], + "logprob_reward": -2.580561399459839, + "yes_no_margin_reward": -2.5, + "token_credit_reward": -2.580561399459839, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09702442586421967, + 0.09500081837177277, + 0.08935794979333878, + 0.0735875740647316, + 0.0822356715798378, + 0.3650991916656494, + 0.1449444442987442, + 0.03189175948500633, + 0.012318032793700695, + 0.008540105074644089 + ], + "credit_map": [ + 0.008721544407308102, + 0.015109511092305183, + 0.007635797839611769, + 0.013292854651808739, + 0.012533657252788544, + 0.014994516968727112, + 0.011959588155150414, + 0.012776955962181091, + 0.008884239941835403, + 0.011280909180641174, + 0.007485578767955303, + 0.03363632410764694, + 0.009704573079943657, + 0.009840195067226887, + 0.007272682152688503, + 0.0068963151425123215, + 0.008808269165456295, + 0.009858673438429832, + 0.009364600293338299, + 0.015910174697637558, + 0.009284223429858685, + 0.010518156923353672, + 0.012313570827245712, + 0.013300281018018723, + 0.0086457384750247, + 0.011678499169647694, + 0.0053869313560426235, + 0.013517824001610279, + 0.009450736455619335, + 0.009292714297771454, + 0.007975381799042225, + 0.00763975502923131, + 0.008081878535449505, + 0.010916189290583134, + 0.00843305978924036, + 0.01246109139174223, + 0.011694427579641342, + 0.010228886269032955, + 0.010560343973338604, + 0.009859799407422543, + 0.03210848569869995, + 0.04555324837565422, + 0.08721154928207397, + 0.030292661860585213, + 0.032766737043857574, + 0.07706412672996521, + 0.03472164273262024, + 0.025380773469805717, + 0.0139311533421278, + 0.014373195357620716, + 0.013996963389217854, + 0.04062698036432266, + 0.018149225041270256, + 0.015684932470321655, + 0.01310751773416996, + 0.015074480324983597, + 0.002766543533653021, + 0.003341004718095064, + 0.0026127498131245375, + 0.003936300054192543, + 0.004017931409180164, + 0.0058158086612820625, + 0.006867885589599609, + 0.002533537335693836, + 0.0016410297248512506, + 0.0012347191805019975, + 0.0011343940859660506, + 0.0013642407720908523, + 0.0014423838583752513, + 0.0015408797189593315, + 0.0012914182152599096, + 0.002668967703357339, + 0.0008270558319054544, + 0.0015889769420027733, + 0.0006867742049507797, + 0.0008205134654417634, + 0.0015915981493890285, + 0.0011552763171494007, + 0.0007545459666289389, + 0.0011153637897223234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -2.580561399459839 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06214362382888794 + ], + "opposite_logprob_tokens": [ + -2.812143564224243 + ], + "logprob_reward": -0.06214362382888794, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06214362382888794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06673792004585266, + 0.06669002771377563, + 0.1419651210308075, + 0.08141676336526871, + 0.08329476416110992, + 0.16544954478740692, + 0.06937326490879059, + 0.25805404782295227, + 0.04640428349375725, + 0.020614271983504295 + ], + "credit_map": [ + 0.0067372629418969154, + 0.008312098681926727, + 0.008022134192287922, + 0.005584200378507376, + 0.0076510407961905, + 0.015122439712285995, + 0.0066546364687383175, + 0.008654101751744747, + 0.007431930396705866, + 0.009263682179152966, + 0.005212605930864811, + 0.013634973205626011, + 0.010496113449335098, + 0.007929358631372452, + 0.006902809254825115, + 0.005818553734570742, + 0.015456116758286953, + 0.01589600369334221, + 0.011646471917629242, + 0.016905944794416428, + 0.018750449642539024, + 0.015964945778250694, + 0.02723989449441433, + 0.020105283707380295, + 0.0110425790771842, + 0.012451021000742912, + 0.009205498732626438, + 0.009823123924434185, + 0.010002527385950089, + 0.009594835340976715, + 0.00936182215809822, + 0.00993536040186882, + 0.009301271289587021, + 0.009273853152990341, + 0.008547987788915634, + 0.010203850455582142, + 0.013085883110761642, + 0.008132467046380043, + 0.011104118078947067, + 0.013645334169268608, + 0.011593803763389587, + 0.02075682021677494, + 0.039176095277071, + 0.012472063302993774, + 0.014647270552814007, + 0.050529103726148605, + 0.008425291627645493, + 0.007849092595279217, + 0.008939475752413273, + 0.008908270858228207, + 0.005836819764226675, + 0.009069652296602726, + 0.010308955796062946, + 0.006871074438095093, + 0.010634738951921463, + 0.008804275654256344, + 0.027010733261704445, + 0.02948652021586895, + 0.019678860902786255, + 0.02336827665567398, + 0.03134367614984512, + 0.04465475305914879, + 0.06034279614686966, + 0.02216840535402298, + 0.004353748168796301, + 0.0063757821917533875, + 0.0040257154032588005, + 0.004945652559399605, + 0.006757454946637154, + 0.005369347985833883, + 0.005915638990700245, + 0.008660943247377872, + 0.0025476047303527594, + 0.003264108207076788, + 0.0020977214444428682, + 0.002329309703782201, + 0.0027199103496968746, + 0.0025302572175860405, + 0.0017811300931498408, + 0.0033442298881709576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.06214362382888794 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330962009727955 + ], + "opposite_logprob_tokens": [ + -3.7733097076416016 + ], + "logprob_reward": -0.02330962009727955, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330962009727955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.076422318816185, + 0.08327388018369675, + 0.1491607278585434, + 0.05941719934344292, + 0.07148246467113495, + 0.3645443618297577, + 0.09968571364879608, + 0.06882443279027939, + 0.017428971827030182, + 0.00975991040468216 + ], + "credit_map": [ + 0.00758944870904088, + 0.01145810168236494, + 0.006830811500549316, + 0.008731120266020298, + 0.00998744834214449, + 0.012275546789169312, + 0.007682868745177984, + 0.011866972781717777, + 0.006978790275752544, + 0.01025095209479332, + 0.006697779055684805, + 0.025875983759760857, + 0.010105093009769917, + 0.008645666763186455, + 0.00773646030575037, + 0.00698316004127264, + 0.028071299195289612, + 0.015965988859534264, + 0.01979617029428482, + 0.01677439548075199, + 0.014845934696495533, + 0.01787164993584156, + 0.0207645483314991, + 0.015070738270878792, + 0.007580404635518789, + 0.006154519040137529, + 0.00772184319794178, + 0.01091317180544138, + 0.006673916708678007, + 0.008658341132104397, + 0.005843490362167358, + 0.005871513858437538, + 0.007830923423171043, + 0.009639160707592964, + 0.006195502355694771, + 0.009300063364207745, + 0.009155483916401863, + 0.008992315270006657, + 0.012500000186264515, + 0.007869016379117966, + 0.018068332225084305, + 0.027929795905947685, + 0.11917239427566528, + 0.02090691216289997, + 0.021498864516615868, + 0.134989395737648, + 0.011286580935120583, + 0.010692080482840538, + 0.010910761542618275, + 0.011753472499549389, + 0.008233489468693733, + 0.01649029739201069, + 0.016523918136954308, + 0.009975376538932323, + 0.015452663414180279, + 0.010345732793211937, + 0.008214322850108147, + 0.007352936081588268, + 0.006818101741373539, + 0.006903800182044506, + 0.007653139531612396, + 0.011683652177453041, + 0.014314540661871433, + 0.0058839404955506325, + 0.0022441288456320763, + 0.0028117839246988297, + 0.0015705301193520427, + 0.002597035840153694, + 0.0021812976337969303, + 0.0017552870558574796, + 0.0018230522982776165, + 0.002445854712277651, + 0.0010400224709883332, + 0.0014977857936173677, + 0.0009847718756645918, + 0.001135638332925737, + 0.001489178859628737, + 0.001038788235746324, + 0.0013127914862707257, + 0.0012609335826709867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.02330962009727955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17997747659683228 + ], + "opposite_logprob_tokens": [ + -1.8049774169921875 + ], + "logprob_reward": -0.17997747659683228, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17997747659683228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07178089022636414, + 0.15095491707324982, + 0.13381683826446533, + 0.07967568933963776, + 0.17603352665901184, + 0.12698888778686523, + 0.11382666230201721, + 0.10458800196647644, + 0.020859386771917343, + 0.021475203335285187 + ], + "credit_map": [ + 0.007233343552798033, + 0.008535316213965416, + 0.006365407723933458, + 0.006516395136713982, + 0.012685374356806278, + 0.01236187107861042, + 0.007639237213879824, + 0.010443950071930885, + 0.014437010511755943, + 0.021233487874269485, + 0.01254054345190525, + 0.02224172092974186, + 0.02334865741431713, + 0.01997782289981842, + 0.019990505650639534, + 0.017185160890221596, + 0.019156966358423233, + 0.01839316636323929, + 0.015424758195877075, + 0.016725242137908936, + 0.014921711757779121, + 0.016816100105643272, + 0.016133734956383705, + 0.0162451583892107, + 0.006491527426987886, + 0.009185614064335823, + 0.008648606948554516, + 0.020151684060692787, + 0.009632389061152935, + 0.00984309334307909, + 0.008475921116769314, + 0.007246851921081543, + 0.015701187774538994, + 0.026034438982605934, + 0.016771595925092697, + 0.01935001090168953, + 0.030240939930081367, + 0.02046176604926586, + 0.02330232970416546, + 0.0241712536662817, + 0.012584665790200233, + 0.019549833610653877, + 0.01678123138844967, + 0.01537014078348875, + 0.014796338975429535, + 0.014788545668125153, + 0.01995110511779785, + 0.013167032971978188, + 0.015419996343553066, + 0.014229684136807919, + 0.00820031575858593, + 0.011523403227329254, + 0.017195990309119225, + 0.012580483220517635, + 0.015330641530454159, + 0.019346149638295174, + 0.01316334679722786, + 0.010121788829565048, + 0.008041281253099442, + 0.011162939481437206, + 0.011257539503276348, + 0.015920596197247505, + 0.023696592077612877, + 0.011223910376429558, + 0.0021321887616068125, + 0.003470272058621049, + 0.001871167216449976, + 0.0026588921900838614, + 0.0025677981320768595, + 0.0025750163476914167, + 0.001755030476488173, + 0.0038290219381451607, + 0.00244697998277843, + 0.002907251939177513, + 0.0019652030896395445, + 0.00288494979031384, + 0.0026211454533040524, + 0.002912855241447687, + 0.0027174674905836582, + 0.0030193503480404615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.17997747659683228 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521497309207916 + ], + "opposite_logprob_tokens": [ + -1.5021497011184692 + ], + "logprob_reward": -0.2521497309207916, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521497309207916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08000592887401581, + 0.11285775154829025, + 0.18268334865570068, + 0.06125940382480621, + 0.10127976536750793, + 0.19465427100658417, + 0.1369517743587494, + 0.08881324529647827, + 0.01942567341029644, + 0.02206885814666748 + ], + "credit_map": [ + 0.008395219221711159, + 0.0122186578810215, + 0.007319380063563585, + 0.008011122234165668, + 0.011762787587940693, + 0.013267749920487404, + 0.008606944233179092, + 0.010424068197607994, + 0.011256281286478043, + 0.013697005808353424, + 0.01192467100918293, + 0.023307355120778084, + 0.014951500110328197, + 0.014384496957063675, + 0.01175372302532196, + 0.011582705192267895, + 0.020096927881240845, + 0.02387705072760582, + 0.016813740134239197, + 0.029864519834518433, + 0.016276361420750618, + 0.018242815509438515, + 0.032446593046188354, + 0.02506532520055771, + 0.008779855445027351, + 0.009351768530905247, + 0.00788665097206831, + 0.007557093631476164, + 0.0077160014770925045, + 0.00734160328283906, + 0.006487817037850618, + 0.006138609256595373, + 0.007985097356140614, + 0.013390625827014446, + 0.01024332083761692, + 0.016194114461541176, + 0.01460503414273262, + 0.012866007164120674, + 0.012282771058380604, + 0.013712786138057709, + 0.014455354772508144, + 0.019155194982886314, + 0.055471207946538925, + 0.01821257546544075, + 0.013192106038331985, + 0.04838908463716507, + 0.013802395202219486, + 0.011976335197687149, + 0.012613398022949696, + 0.012856051325798035, + 0.012875399552285671, + 0.040569763630628586, + 0.013661359436810017, + 0.015206494368612766, + 0.01294125895947218, + 0.016228031367063522, + 0.009702276438474655, + 0.009650089778006077, + 0.006916264537721872, + 0.009198730811476707, + 0.008487828075885773, + 0.014222764410078526, + 0.02043590322136879, + 0.010199381038546562, + 0.002441216493025422, + 0.0028838294092565775, + 0.001715076039545238, + 0.0028765525203198195, + 0.0024949528742581606, + 0.0023528016172349453, + 0.001795229152776301, + 0.002866014139726758, + 0.0016670646145939827, + 0.0031779983546584845, + 0.002076429547742009, + 0.00180545705370605, + 0.0030013981740921736, + 0.0031347712501883507, + 0.0020298068411648273, + 0.005175929982215166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.2521497309207916 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a chaotic handheld camera shot imbued with the iconic aesthetic of '80s anime, the frame shakes violently with the pounding wind, struggling to focus on a lone Sophora japonica tree. It's the epicenter of a ferocious storm, the sky churning into a maelstrom of deep purple and ominous magenta, while rough, high-contrast brushstrokes fill the background. The ancient Sophora japonica tree waged a desperate struggle, its stout trunk bent stubbornly to the brink of breaking under the relentless force of the hurricane, its twisted branches lashing wildly at the air like struggling arms. Torrential rain streaked across the frame like dense, white slashes, lashing the tree's gnarled, deeply cracked bark with a wet gleam in the dim light. Every shake of the camera captured the vivid details of the water coursing along the veins. Dark green leaves were ruthlessly torn off and danced wildly in the chaotic air currents. A dazzling flash of lightning instantly tore through the sky, illuminating this doomsday scene full of dynamic tension with a brief pale light, creating a primitive, intense and tragic visual atmosphere.", + "seed": 2078, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_5/index_3576_seed_2078_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.4501338452100754, + "vlm_logprob": -0.4501338452100754, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.4501338452100754, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.12975811958313 + ], + "opposite_logprob_tokens": [ + -0.1297580748796463 + ], + "logprob_reward": -2.12975811958313, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.12975811958313, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08859354257583618, + 0.20385321974754333, + 0.04564085230231285, + 0.021967828273773193, + 0.20547974109649658, + 0.04004913195967674, + 0.32196903228759766, + 0.03151463344693184, + 0.015310101211071014, + 0.025621969252824783 + ], + "credit_map": [ + 0.00906216911971569, + 0.008447736501693726, + 0.00874376017600298, + 0.01123614702373743, + 0.012965195812284946, + 0.01099060382694006, + 0.01331310998648405, + 0.013834814541041851, + 0.01810229942202568, + 0.013273254968225956, + 0.013739245012402534, + 0.009286251850426197, + 0.013272528536617756, + 0.07403959333896637, + 0.012440080754458904, + 0.04969993606209755, + 0.005296207964420319, + 0.006227605976164341, + 0.004788526333868504, + 0.004043017979711294, + 0.004857408348470926, + 0.010019158944487572, + 0.004303312394768, + 0.006105608772486448, + 0.0024000355042517185, + 0.0033912016078829765, + 0.0023571266792714596, + 0.002887505805119872, + 0.0031370946671813726, + 0.002773765940219164, + 0.0028553493320941925, + 0.0021657466422766447, + 0.02024836279451847, + 0.04212556406855583, + 0.0151832215487957, + 0.01591501571238041, + 0.022069426253437996, + 0.03734792396426201, + 0.017010394483804703, + 0.03557981178164482, + 0.004605684895068407, + 0.006166841834783554, + 0.004083297681063414, + 0.0043494622223079205, + 0.00657539302483201, + 0.004657610319554806, + 0.0056053875014185905, + 0.004005451686680317, + 0.05043265223503113, + 0.011602729558944702, + 0.011545722372829914, + 0.009207529947161674, + 0.015177431516349316, + 0.07316634804010391, + 0.020021427422761917, + 0.13081516325473785, + 0.005616023205220699, + 0.0036028821486979723, + 0.003340422874316573, + 0.003429155331104994, + 0.0036981524899601936, + 0.004062614869326353, + 0.004067445173859596, + 0.00369793432764709, + 0.0025757320690900087, + 0.0019851955585181713, + 0.001738031511195004, + 0.0015192676801234484, + 0.00181019795127213, + 0.0018488520290702581, + 0.0016019530594348907, + 0.0022308703046292067, + 0.0029193093068897724, + 0.006602142471820116, + 0.0029093988705426455, + 0.0022583368699997663, + 0.0032818259205669165, + 0.0030259941704571247, + 0.0021316411439329386, + 0.002493317238986492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -2.12975811958313 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033759646117687225 + ], + "opposite_logprob_tokens": [ + -3.408759593963623 + ], + "logprob_reward": -0.033759646117687225, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033759646117687225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08051398396492004, + 0.13550236821174622, + 0.12121807783842087, + 0.028531786054372787, + 0.07742077112197876, + 0.0662168338894844, + 0.3912639021873474, + 0.04452328756451607, + 0.013327525928616524, + 0.041481584310531616 + ], + "credit_map": [ + 0.009265211410820484, + 0.008343852125108242, + 0.0071830591186881065, + 0.01075779553502798, + 0.00948997400701046, + 0.010736887343227863, + 0.015949951484799385, + 0.008787248283624649, + 0.01061739306896925, + 0.009026357904076576, + 0.0074872481636703014, + 0.006427774205803871, + 0.009328015148639679, + 0.048828765749931335, + 0.00900384970009327, + 0.03478294983506203, + 0.00830613262951374, + 0.009240975603461266, + 0.010681009851396084, + 0.008927827700972557, + 0.00968991033732891, + 0.03723575547337532, + 0.008695142343640327, + 0.028441311791539192, + 0.003001617733389139, + 0.0027441561687737703, + 0.002832035068422556, + 0.004129764623939991, + 0.002840908244252205, + 0.003973709885030985, + 0.005778128281235695, + 0.003231461625546217, + 0.007019747979938984, + 0.013792756013572216, + 0.005604415666311979, + 0.00558645511046052, + 0.008250150829553604, + 0.01593957468867302, + 0.005770975723862648, + 0.01545669324696064, + 0.008222260512411594, + 0.010504040867090225, + 0.006613981444388628, + 0.0070579624734818935, + 0.009876987896859646, + 0.008709602989256382, + 0.006831380072981119, + 0.008400609716773033, + 0.05465182662010193, + 0.0150356600061059, + 0.013655929826200008, + 0.011019032448530197, + 0.019525211304426193, + 0.09533116221427917, + 0.024323146790266037, + 0.1577218770980835, + 0.006785871926695108, + 0.004780993331223726, + 0.004442739300429821, + 0.005744767840951681, + 0.005925267469137907, + 0.004696923308074474, + 0.008175482973456383, + 0.003971237223595381, + 0.0020463569089770317, + 0.0018874653615057468, + 0.0013159550726413727, + 0.001554847345687449, + 0.001440875930711627, + 0.002487846650183201, + 0.0011454798514023423, + 0.0014486964792013168, + 0.006337855011224747, + 0.006854282226413488, + 0.003944209776818752, + 0.002650695387274027, + 0.003161906497552991, + 0.006592597346752882, + 0.0030273315496742725, + 0.008912702091038227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.033759646117687225 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648322895169258 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.008648322895169258, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648322895169258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0699443593621254, + 0.2121371477842331, + 0.1570921093225479, + 0.024021435528993607, + 0.07232274860143661, + 0.057862281799316406, + 0.3017531931400299, + 0.0432325080037117, + 0.016648264601826668, + 0.04498597979545593 + ], + "credit_map": [ + 0.009818892925977707, + 0.006779753603041172, + 0.011878971010446548, + 0.008607186377048492, + 0.008032682351768017, + 0.0074508353136479855, + 0.006468525156378746, + 0.010907512158155441, + 0.02125987596809864, + 0.012422895058989525, + 0.011100305244326591, + 0.00825534574687481, + 0.014614845626056194, + 0.08381222933530807, + 0.010278481990098953, + 0.05039317160844803, + 0.01560082659125328, + 0.016532322391867638, + 0.015812980011105537, + 0.011451007798314095, + 0.011528711766004562, + 0.04086116701364517, + 0.01203289907425642, + 0.03327219560742378, + 0.0023194411769509315, + 0.002568397670984268, + 0.0023852598387748003, + 0.003744665766134858, + 0.0030183603521436453, + 0.0032781187910586596, + 0.004303705412894487, + 0.002403487451374531, + 0.008198066614568233, + 0.017121760174632072, + 0.006222368683665991, + 0.006390199065208435, + 0.009383087046444416, + 0.008504294790327549, + 0.0083576375618577, + 0.00814533606171608, + 0.0055136969313025475, + 0.007990035228431225, + 0.005388815887272358, + 0.0073666865937411785, + 0.009858022443950176, + 0.006644485983997583, + 0.008912354707717896, + 0.00618818448856473, + 0.023087382316589355, + 0.012790407054126263, + 0.010775317437946796, + 0.008181553333997726, + 0.016955159604549408, + 0.09372109919786453, + 0.017097489908337593, + 0.11914479732513428, + 0.006222428288310766, + 0.003630475141108036, + 0.007765449583530426, + 0.00441101286560297, + 0.00420373072847724, + 0.005478180944919586, + 0.005914393812417984, + 0.005606833379715681, + 0.002555932616814971, + 0.0025753274094313383, + 0.0015620035119354725, + 0.0017631748924031854, + 0.0021101769525557756, + 0.002005780814215541, + 0.0022636328358203173, + 0.0018122355686500669, + 0.0050250692293047905, + 0.004020975902676582, + 0.003128848969936371, + 0.0024072087835520506, + 0.0049536144360899925, + 0.01018216647207737, + 0.0037352940998971462, + 0.011532803997397423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.008648322895169258 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868850111961365 + ], + "opposite_logprob_tokens": [ + -3.0486884117126465 + ], + "logprob_reward": -0.04868850111961365, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868850111961365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07522805035114288, + 0.032832786440849304, + 0.060566045343875885, + 0.019217776134610176, + 0.32945504784584045, + 0.18830272555351257, + 0.18707342445850372, + 0.027282804250717163, + 0.014263490214943886, + 0.06577780842781067 + ], + "credit_map": [ + 0.008773774839937687, + 0.010341184213757515, + 0.0104859359562397, + 0.007016390562057495, + 0.009218729101121426, + 0.010331190191209316, + 0.00854865275323391, + 0.010512197390198708, + 0.0030132231768220663, + 0.003252673428505659, + 0.0027870263438671827, + 0.0035163178108632565, + 0.0026121679693460464, + 0.007582108490169048, + 0.003413143800571561, + 0.0066561270505189896, + 0.005970468278974295, + 0.006311831064522266, + 0.009664206765592098, + 0.00692326994612813, + 0.0071305884048342705, + 0.007888336665928364, + 0.007498682010918856, + 0.009178660809993744, + 0.002131086541339755, + 0.002100359182804823, + 0.0023200151044875383, + 0.0031169182620942593, + 0.002075755735859275, + 0.0021337117068469524, + 0.0036381410900503397, + 0.0017017896752804518, + 0.034260910004377365, + 0.07152363657951355, + 0.02519802935421467, + 0.027391130104660988, + 0.03577977418899536, + 0.0565301813185215, + 0.027590619400143623, + 0.05118076130747795, + 0.02238752692937851, + 0.027899786829948425, + 0.014203343540430069, + 0.019305987283587456, + 0.027205320075154305, + 0.028279270976781845, + 0.025439118966460228, + 0.023582370951771736, + 0.009596097283065319, + 0.010898640379309654, + 0.006881760898977518, + 0.006922619417309761, + 0.012049932032823563, + 0.062347229570150375, + 0.012324515730142593, + 0.0660526305437088, + 0.004473002161830664, + 0.0030322247184813023, + 0.004002188332378864, + 0.002679104683920741, + 0.002782774157822132, + 0.003489629365503788, + 0.002622659085318446, + 0.004201221279799938, + 0.0026194157544523478, + 0.0012380486587062478, + 0.0016723263543099165, + 0.0013702345313504338, + 0.0012099938467144966, + 0.002187936333939433, + 0.0013606817228719592, + 0.002604853129014373, + 0.011207638308405876, + 0.008143056184053421, + 0.0055689262226223946, + 0.005799716338515282, + 0.006299278698861599, + 0.011367885395884514, + 0.005345758982002735, + 0.012045546434819698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.04868850111961365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot handheld and shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a lone Sophora japonica tree with a stout, bent trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a maelstrom of deep purple and ominous magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there dark green leaves being torn off and dancing wildly in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a dazzling flash of lightning briefly illuminate the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981463633477688 + ], + "opposite_logprob_tokens": [ + -3.5298147201538086 + ], + "logprob_reward": -0.02981463633477688, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981463633477688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12145210802555084, + 0.04035243019461632, + 0.0413796603679657, + 0.02698858641088009, + 0.03896767646074295, + 0.057387061417102814, + 0.24269390106201172, + 0.04457531496882439, + 0.01785176806151867, + 0.368351548910141 + ], + "credit_map": [ + 0.015759436413645744, + 0.012465724721550941, + 0.016012877225875854, + 0.009150690399110317, + 0.009348967112600803, + 0.01951153203845024, + 0.008628273382782936, + 0.030574603006243706, + 0.0070726582780480385, + 0.005449894815683365, + 0.005304924212396145, + 0.004483215510845184, + 0.0037325285375118256, + 0.007235113997012377, + 0.004251414909958839, + 0.002822677604854107, + 0.005103664007037878, + 0.005027836188673973, + 0.0031566780526190996, + 0.003963634837418795, + 0.003150888718664646, + 0.007095285691320896, + 0.0053047738038003445, + 0.008576895110309124, + 0.0031195448245853186, + 0.004387483466416597, + 0.0031930545810610056, + 0.0035518945660442114, + 0.0035605845041573048, + 0.0027330960147082806, + 0.003583180485293269, + 0.002859747037291527, + 0.0039575546979904175, + 0.007507892791181803, + 0.0037827775813639164, + 0.00396076450124383, + 0.005002512596547604, + 0.004013019613921642, + 0.007165755145251751, + 0.0035773986019194126, + 0.006775280460715294, + 0.007920118048787117, + 0.006317049264907837, + 0.006944586057215929, + 0.0073808166198432446, + 0.006848298013210297, + 0.009715783409774303, + 0.005485131870955229, + 0.037900906056165695, + 0.013472825288772583, + 0.012930495664477348, + 0.008375884965062141, + 0.016609719023108482, + 0.086571604013443, + 0.012832644395530224, + 0.053999826312065125, + 0.0055198585614562035, + 0.0077850306406617165, + 0.006330623757094145, + 0.00565944891422987, + 0.003934127744287252, + 0.005167081952095032, + 0.005400720052421093, + 0.004778427537530661, + 0.0022076929453760386, + 0.0019817522261291742, + 0.0021909428760409355, + 0.002428770763799548, + 0.0023729235399514437, + 0.0020072758197784424, + 0.0021929729264229536, + 0.0024694360326975584, + 0.032322827726602554, + 0.072806216776371, + 0.03818809241056442, + 0.024367140606045723, + 0.0456216000020504, + 0.059748247265815735, + 0.03482813015580177, + 0.06046925485134125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.02981463633477688 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_6/index_tensor([7955, 7955])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_6/index_tensor([7955, 7955])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..80b184c539105679dd11d9405efa6a938a25e9f1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_6/index_tensor([7955, 7955])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful overhead shot, shot in high-frame-rate slow motion, captures a breathtaking moment: a massive, arrow-shaped metallic alien spacecraft violently tears through a thick, roiling storm cloud. Composed of a dark, non-reflective alloy, its surface covered in intricate geometric paneling and scarred by atmospheric passage, its massive bulk carves a vast chasm in the sea of clouds as it plummets downward. Filmed at high frame rate, every wisp of torn cloud and vaporized debris tumbles and dissipates around the vessel's edge in exquisite detail. Along its flank, energy conduits pulse with eerie blue and crimson light, casting an ominous halo on the surrounding clouds. As the camera tilts steadily downward, we see the spacecraft plummeting with terrifying precision towards the shadowed expanse of gleaming skyscrapers below. The entire scene is imbued with an overwhelming sense of scale and apocalyptic tension.", + "seed": 31116, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_6/index_7955_seed_31116_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03827402284368873, + "vlm_logprob": -0.03827402284368873, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03827402284368873, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010843047872185707 + ], + "opposite_logprob_tokens": [ + -4.760843276977539 + ], + "logprob_reward": -0.010843047872185707, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010843047872185707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08750598132610321, + 0.07271183282136917, + 0.0865384116768837, + 0.038700733333826065, + 0.1675654947757721, + 0.07223064452409744, + 0.11943154036998749, + 0.11598644405603409, + 0.08160760998725891, + 0.15772123634815216 + ], + "credit_map": [ + 0.007030867040157318, + 0.018099823966622353, + 0.009030771441757679, + 0.012557045556604862, + 0.011483089998364449, + 0.007530886214226484, + 0.01310445461422205, + 0.0086690504103899, + 0.008019905537366867, + 0.010154097341001034, + 0.008430990390479565, + 0.011187056079506874, + 0.006743631791323423, + 0.00699740182608366, + 0.010664425790309906, + 0.010514331050217152, + 0.011182663030922413, + 0.008470168337225914, + 0.010019469074904919, + 0.013349635526537895, + 0.010123783722519875, + 0.011045304127037525, + 0.012949001044034958, + 0.009398393332958221, + 0.002255551517009735, + 0.0029146193992346525, + 0.008928717114031315, + 0.004343521781265736, + 0.0032204745803028345, + 0.004162948578596115, + 0.004568910226225853, + 0.008305991068482399, + 0.014466985128819942, + 0.031217139214277267, + 0.018784813582897186, + 0.020679747685790062, + 0.018979307264089584, + 0.019873257726430893, + 0.02906779758632183, + 0.014496446587145329, + 0.0066761900670826435, + 0.013994378037750721, + 0.008539298549294472, + 0.008387475274503231, + 0.009679792448878288, + 0.0073801372200250626, + 0.009606228210031986, + 0.007967143319547176, + 0.009773114696145058, + 0.011575448326766491, + 0.014700280502438545, + 0.020588871091604233, + 0.01206874754279852, + 0.026957398280501366, + 0.013973107561469078, + 0.009794574230909348, + 0.006777659989893436, + 0.015559406019747257, + 0.009343735873699188, + 0.01071163360029459, + 0.012220454402267933, + 0.02618510276079178, + 0.009250677190721035, + 0.025937771424651146, + 0.004862584639340639, + 0.005200984422117472, + 0.011844748631119728, + 0.012400449253618717, + 0.011332608759403229, + 0.014318658038973808, + 0.007173031102865934, + 0.014474546536803246, + 0.01656767539680004, + 0.019725073128938675, + 0.02804308384656906, + 0.020437948405742645, + 0.028450313955545425, + 0.01902860403060913, + 0.014228134416043758, + 0.011240405030548573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.010843047872185707 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026362450793385506 + ], + "opposite_logprob_tokens": [ + -3.651362419128418 + ], + "logprob_reward": -0.026362450793385506, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026362450793385506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06095931679010391, + 0.07750498503446579, + 0.05886293947696686, + 0.1142851784825325, + 0.02573821134865284, + 0.044812560081481934, + 0.26442718505859375, + 0.08747556805610657, + 0.1342962086200714, + 0.13163785636425018 + ], + "credit_map": [ + 0.005320548079907894, + 0.009327172301709652, + 0.00889053475111723, + 0.008649765513837337, + 0.007643943186849356, + 0.006875547114759684, + 0.00808777753263712, + 0.006164026912301779, + 0.007581125013530254, + 0.014079124666750431, + 0.00811222568154335, + 0.008743354119360447, + 0.00794574897736311, + 0.009494625963270664, + 0.014517342671751976, + 0.007031445857137442, + 0.007439936511218548, + 0.006450512446463108, + 0.006939110346138477, + 0.006302684545516968, + 0.008238234557211399, + 0.006920947693288326, + 0.008016480132937431, + 0.0085550332441926, + 0.006058007013052702, + 0.012069380842149258, + 0.02908814139664173, + 0.011964957229793072, + 0.0051172045059502125, + 0.007230838295072317, + 0.014774920418858528, + 0.027981730177998543, + 0.002755574183538556, + 0.0034697898663580418, + 0.0031828307546675205, + 0.0035697645507752895, + 0.00300119468010962, + 0.003304033540189266, + 0.003807700239121914, + 0.002647322602570057, + 0.003566276514902711, + 0.005391237325966358, + 0.009098966605961323, + 0.0059476569294929504, + 0.005399862304329872, + 0.005585004109889269, + 0.004927234724164009, + 0.004896317142993212, + 0.0200706385076046, + 0.029172595590353012, + 0.02980685792863369, + 0.04554044455289841, + 0.03174833208322525, + 0.04406698793172836, + 0.04421841353178024, + 0.019802914932370186, + 0.0054071699269115925, + 0.010224485769867897, + 0.00953944493085146, + 0.010916654951870441, + 0.008124291896820068, + 0.012508287094533443, + 0.010942842811346054, + 0.019812392070889473, + 0.008697539567947388, + 0.011181354522705078, + 0.022592047229409218, + 0.015709541738033295, + 0.018124688416719437, + 0.023691410198807716, + 0.01098280493170023, + 0.023316819220781326, + 0.016298234462738037, + 0.02130020223557949, + 0.02255333960056305, + 0.014231665059924126, + 0.019421888515353203, + 0.016991062089800835, + 0.01191701926290989, + 0.008924443274736404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.026362450793385506 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463993102312088 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.00463993102312088, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463993102312088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05976957082748413, + 0.04487479850649834, + 0.04692450538277626, + 0.10639446973800659, + 0.089865542948246, + 0.0935920849442482, + 0.04717973619699478, + 0.21654291450977325, + 0.17884434759616852, + 0.11601199209690094 + ], + "credit_map": [ + 0.005540676414966583, + 0.00863699521869421, + 0.009878463111817837, + 0.007838801480829716, + 0.00707703223451972, + 0.004133629146963358, + 0.006798182614147663, + 0.009865788742899895, + 0.00478371512144804, + 0.006864228751510382, + 0.004942171275615692, + 0.008158034645020962, + 0.0046537695452570915, + 0.0036372242029756308, + 0.007470913231372833, + 0.004364744294434786, + 0.006032587960362434, + 0.004424797371029854, + 0.005709563847631216, + 0.0058237286284565926, + 0.005538972560316324, + 0.004923180211335421, + 0.00865230429917574, + 0.005819374229758978, + 0.007850747555494308, + 0.011010161601006985, + 0.023023031651973724, + 0.011794628575444221, + 0.0080499816685915, + 0.01111253909766674, + 0.011215664446353912, + 0.022337719798088074, + 0.006766865495592356, + 0.015128327533602715, + 0.009942689910531044, + 0.011963275261223316, + 0.010746500454843044, + 0.010457081720232964, + 0.017849428579211235, + 0.007011383771896362, + 0.007747821044176817, + 0.011691207066178322, + 0.01633440889418125, + 0.01183687336742878, + 0.011903545819222927, + 0.010943930596113205, + 0.011184990406036377, + 0.011949310079216957, + 0.004392907023429871, + 0.005246839486062527, + 0.004359091632068157, + 0.007153917104005814, + 0.005607019644230604, + 0.00739732151851058, + 0.008540540933609009, + 0.004482104908674955, + 0.012252149172127247, + 0.03038107417523861, + 0.020117932930588722, + 0.02198614552617073, + 0.019837871193885803, + 0.038090452551841736, + 0.0234739538282156, + 0.05040336772799492, + 0.011118218302726746, + 0.01366517785936594, + 0.023672794923186302, + 0.02459723688662052, + 0.02235928177833557, + 0.021009892225265503, + 0.04107413813471794, + 0.021347615867853165, + 0.012607915326952934, + 0.012959342449903488, + 0.01658874936401844, + 0.016487207263708115, + 0.009673996828496456, + 0.01910964772105217, + 0.012363155372440815, + 0.016221988946199417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.00463993102312088 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732400041073561 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.006732400041073561, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732400041073561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0670844316482544, + 0.03400149568915367, + 0.03832309693098068, + 0.06778769940137863, + 0.06015094742178917, + 0.0933254137635231, + 0.1184864342212677, + 0.06663231551647186, + 0.1721489280462265, + 0.2820592522621155 + ], + "credit_map": [ + 0.0061738197691738605, + 0.012091654352843761, + 0.007603578269481659, + 0.01110677421092987, + 0.007957433350384235, + 0.0065436228178441525, + 0.00881083495914936, + 0.00679670600220561, + 0.003845933824777603, + 0.004438255913555622, + 0.003211932722479105, + 0.0033754741307348013, + 0.0033906949684023857, + 0.00404298584908247, + 0.00788783933967352, + 0.0038083763793110847, + 0.005550282076001167, + 0.005368614569306374, + 0.003878084011375904, + 0.004487161058932543, + 0.003973477520048618, + 0.004211194347590208, + 0.006273337174206972, + 0.004580946639180183, + 0.004620433319360018, + 0.00822160579264164, + 0.013354914262890816, + 0.007585379760712385, + 0.005522410850971937, + 0.006782237906008959, + 0.010087473317980766, + 0.011613244190812111, + 0.004528920166194439, + 0.00777773279696703, + 0.007502087391912937, + 0.010940550826489925, + 0.0057044136337935925, + 0.0066325245425105095, + 0.012422263622283936, + 0.004642455372959375, + 0.007916353642940521, + 0.020674392580986023, + 0.010478784330189228, + 0.012042338028550148, + 0.016361359506845474, + 0.007582276593893766, + 0.011659180745482445, + 0.00661073112860322, + 0.008279449306428432, + 0.015128080733120441, + 0.014347662217915058, + 0.02091914787888527, + 0.011579467914998531, + 0.02695615030825138, + 0.010911067016422749, + 0.010365411639213562, + 0.006116175092756748, + 0.010880163870751858, + 0.006866772193461657, + 0.007859150879085064, + 0.006321167107671499, + 0.010139334946870804, + 0.006511450279504061, + 0.011938100680708885, + 0.010845102369785309, + 0.018415343016386032, + 0.03140159323811531, + 0.021690772846341133, + 0.020841609686613083, + 0.03401969000697136, + 0.013859156519174576, + 0.021075662225484848, + 0.016671953722834587, + 0.01651078462600708, + 0.04008731618523598, + 0.04220454394817352, + 0.04795327037572861, + 0.06322307884693146, + 0.018793635070323944, + 0.03661464899778366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.006732400041073561 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279228448867798 + ], + "opposite_logprob_tokens": [ + -2.017792224884033 + ], + "logprob_reward": -0.14279228448867798, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14279228448867798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08789743483066559, + 0.03705936670303345, + 0.06391244381666183, + 0.1445126235485077, + 0.27578479051589966, + 0.11742531508207321, + 0.10448286682367325, + 0.06574002653360367, + 0.046325601637363434, + 0.0568595789372921 + ], + "credit_map": [ + 0.008076286874711514, + 0.014083345420658588, + 0.011778222396969795, + 0.011169896461069584, + 0.010533809661865234, + 0.00755185866728425, + 0.014628137461841106, + 0.01007587555795908, + 0.0036087888292968273, + 0.005571909248828888, + 0.0044158329255878925, + 0.004404541105031967, + 0.003265672130510211, + 0.003904585959389806, + 0.00625729700550437, + 0.005630739498883486, + 0.00597686693072319, + 0.005940604023635387, + 0.010165467858314514, + 0.006822654046118259, + 0.008120534010231495, + 0.007467405870556831, + 0.008639391511678696, + 0.01077952329069376, + 0.008244412951171398, + 0.013820074498653412, + 0.03365891054272652, + 0.011931349523365498, + 0.010650496929883957, + 0.01260191947221756, + 0.022215958684682846, + 0.0313895009458065, + 0.021156270056962967, + 0.051192086189985275, + 0.030059635639190674, + 0.034439872950315475, + 0.0309187900274992, + 0.03277676925063133, + 0.05207732692360878, + 0.02316402830183506, + 0.01018630899488926, + 0.014679772779345512, + 0.017407264560461044, + 0.015624616295099258, + 0.013474192470312119, + 0.013248903676867485, + 0.016240619122982025, + 0.01656363345682621, + 0.007645492907613516, + 0.01285229530185461, + 0.012737142853438854, + 0.01135308388620615, + 0.011011511087417603, + 0.024654168635606766, + 0.014885632321238518, + 0.009343534708023071, + 0.003351778956130147, + 0.007805980276316404, + 0.006067380309104919, + 0.005774199962615967, + 0.007097617257386446, + 0.014224800281226635, + 0.0071617779321968555, + 0.014256490394473076, + 0.0035199280828237534, + 0.005316686350852251, + 0.006675613112747669, + 0.00738200219348073, + 0.006321742199361324, + 0.005742714274674654, + 0.005190665367990732, + 0.00617625005543232, + 0.004509674850851297, + 0.004792132880538702, + 0.00965428352355957, + 0.008564497344195843, + 0.006572049111127853, + 0.01128352154046297, + 0.005258643999695778, + 0.0062247756868600845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.14279228448867798 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful overhead shot, shot in high-frame-rate slow motion, captures a breathtaking moment: a massive, arrow-shaped metallic alien spacecraft violently tears through a thick, roiling storm cloud. Composed of a dark, non-reflective alloy, its surface covered in intricate geometric paneling and scarred by atmospheric passage, its massive bulk carves a vast chasm in the sea of clouds as it plummets downward. Filmed at high frame rate, every wisp of torn cloud and vaporized debris tumbles and dissipates around the vessel's edge in exquisite detail. Along its flank, energy conduits pulse with eerie blue and crimson light, casting an ominous halo on the surrounding clouds. As the camera tilts steadily downward, we see the spacecraft plummeting with terrifying precision towards the shadowed expanse of gleaming skyscrapers below. The entire scene is imbued with an overwhelming sense of scale and apocalyptic tension.", + "seed": 31117, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_6/index_7955_seed_31117_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05971326921135187, + "vlm_logprob": -0.05971326921135187, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.05971326921135187, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009146574884653091 + ], + "opposite_logprob_tokens": [ + -5.134146690368652 + ], + "logprob_reward": -0.009146574884653091, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009146574884653091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2020755559206009, + 0.1091107428073883, + 0.05418005585670471, + 0.12209723144769669, + 0.1141984686255455, + 0.10296982526779175, + 0.047124430537223816, + 0.07530089467763901, + 0.04765394330024719, + 0.12528885900974274 + ], + "credit_map": [ + 0.016538390889763832, + 0.03409770876169205, + 0.016217254102230072, + 0.03407861664891243, + 0.023233691230416298, + 0.023775042966008186, + 0.019793817773461342, + 0.03434103727340698, + 0.009991800412535667, + 0.018100084736943245, + 0.013812215998768806, + 0.010821939446032047, + 0.01484909001737833, + 0.014833991415798664, + 0.01262506004422903, + 0.01407655794173479, + 0.003935922868549824, + 0.006423488259315491, + 0.006406343542039394, + 0.007824758999049664, + 0.005587258841842413, + 0.008374212309718132, + 0.009287474676966667, + 0.00634059077128768, + 0.008753434754908085, + 0.009470519609749317, + 0.03444439917802811, + 0.016362274065613747, + 0.008923497051000595, + 0.011213429272174835, + 0.008867990225553513, + 0.024061691015958786, + 0.007590778172016144, + 0.013776886276900768, + 0.012398209422826767, + 0.012147600762546062, + 0.008767244406044483, + 0.009728046134114265, + 0.019878290593624115, + 0.029911410063505173, + 0.005788966082036495, + 0.00959932804107666, + 0.006086783483624458, + 0.015343881212174892, + 0.005087350960820913, + 0.008777990937232971, + 0.03519928827881813, + 0.017086243256926537, + 0.004462060984224081, + 0.005355661269277334, + 0.007876464165747166, + 0.007961110211908817, + 0.004989859648048878, + 0.00484843272715807, + 0.005877855233848095, + 0.005752986762672663, + 0.005275475326925516, + 0.007498125545680523, + 0.012701718136668205, + 0.015558885410428047, + 0.007810283452272415, + 0.010419169440865517, + 0.009815261699259281, + 0.00622197613120079, + 0.003483550390228629, + 0.007749814540147781, + 0.004417726770043373, + 0.005897625349462032, + 0.005873123649507761, + 0.00657119182869792, + 0.00827887561172247, + 0.005382033996284008, + 0.012947676703333855, + 0.016517290845513344, + 0.01684596762061119, + 0.016114594414830208, + 0.019813494756817818, + 0.024764930829405785, + 0.010573544539511204, + 0.007711362559348345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.009146574884653091 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308418184518814 + ], + "opposite_logprob_tokens": [ + -3.168084144592285 + ], + "logprob_reward": -0.04308418184518814, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308418184518814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2813574969768524, + 0.21399950981140137, + 0.0259256511926651, + 0.08270394802093506, + 0.10710040479898453, + 0.08638715744018555, + 0.06204495579004288, + 0.04219581559300423, + 0.04157674312591553, + 0.05670827999711037 + ], + "credit_map": [ + 0.021017078310251236, + 0.059365641325712204, + 0.02447315864264965, + 0.048743925988674164, + 0.03431858494877815, + 0.033198755234479904, + 0.0287183690816164, + 0.031521987169981, + 0.015248346142470837, + 0.05353542044758797, + 0.02017931267619133, + 0.017490893602371216, + 0.037866875529289246, + 0.02524823695421219, + 0.025208577513694763, + 0.019221842288970947, + 0.0028907335363328457, + 0.004114823415875435, + 0.002945923712104559, + 0.0033202460035681725, + 0.003541406709700823, + 0.0026916288770735264, + 0.003554795403033495, + 0.0028660940006375313, + 0.006728156004101038, + 0.0031053933780640364, + 0.02849441021680832, + 0.005596601869910955, + 0.003442003857344389, + 0.008904755115509033, + 0.006341428030282259, + 0.02009120211005211, + 0.005918207578361034, + 0.008364737033843994, + 0.010886256583034992, + 0.009052466601133347, + 0.0073110731318593025, + 0.005896790884435177, + 0.030520115047693253, + 0.029150765389204025, + 0.004103475250303745, + 0.00505203427746892, + 0.005050085484981537, + 0.007304879371076822, + 0.0036597619764506817, + 0.0052339243702590466, + 0.03173806518316269, + 0.024244925007224083, + 0.006524945609271526, + 0.00799243152141571, + 0.009663118049502373, + 0.009540574625134468, + 0.006754626054316759, + 0.007007784675806761, + 0.008510565385222435, + 0.006050910335034132, + 0.0030919008422642946, + 0.004788696300238371, + 0.006165946368128061, + 0.009024979546666145, + 0.004472174681723118, + 0.005440914072096348, + 0.004084278363734484, + 0.005126922391355038, + 0.0028994332533329725, + 0.0052019814029335976, + 0.004605211317539215, + 0.004961278289556503, + 0.004861877299845219, + 0.005413427483290434, + 0.007698471657931805, + 0.0059350598603487015, + 0.006297604646533728, + 0.006451649125665426, + 0.0072689722292125225, + 0.007239802740514278, + 0.008748294785618782, + 0.012284284457564354, + 0.004410642199218273, + 0.004007026087492704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04308418184518814 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078749783337116 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078749783337116, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078749783337116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07597606629133224, + 0.047062478959560394, + 0.019991569221019745, + 0.0451531745493412, + 0.6667397022247314, + 0.04327470064163208, + 0.024560101330280304, + 0.027340613305568695, + 0.023383164778351784, + 0.02651849575340748 + ], + "credit_map": [ + 0.006947945337742567, + 0.009184413589537144, + 0.00738446693867445, + 0.01247057132422924, + 0.011393370106816292, + 0.008482658304274082, + 0.00651985639706254, + 0.013592774048447609, + 0.0035400493070483208, + 0.009334756061434746, + 0.005409894045442343, + 0.004028188064694405, + 0.008480589836835861, + 0.00532960519194603, + 0.005664007272571325, + 0.0052753835916519165, + 0.0015844963490962982, + 0.0020188200287520885, + 0.0025124549865722656, + 0.003437839448451996, + 0.002268532756716013, + 0.0027838083915412426, + 0.0032605240121483803, + 0.002125091152265668, + 0.0028552215080708265, + 0.0034634368494153023, + 0.009934630244970322, + 0.010182680562138557, + 0.002631715964525938, + 0.0033778147771954536, + 0.004103414248675108, + 0.008604257367551327, + 0.02192338928580284, + 0.02778126485645771, + 0.03115418180823326, + 0.01989646628499031, + 0.022016974166035652, + 0.02356301248073578, + 0.25225430727005005, + 0.26815006136894226, + 0.001982651650905609, + 0.003769514849409461, + 0.0020556035451591015, + 0.004025527276098728, + 0.0020754856523126364, + 0.0023215110413730145, + 0.013114769011735916, + 0.013929632492363453, + 0.0019208056619390845, + 0.0023503119591623545, + 0.003357743378728628, + 0.004523695912212133, + 0.0020221739541739225, + 0.003184178378432989, + 0.004821484442800283, + 0.002379703801125288, + 0.002075684955343604, + 0.002897744532674551, + 0.004530044738203287, + 0.005024214740842581, + 0.0031512391287833452, + 0.0033287559635937214, + 0.003791780909523368, + 0.00254114530980587, + 0.0017657895805314183, + 0.0033461658749729395, + 0.003078281180933118, + 0.0025024055503308773, + 0.0028900697361677885, + 0.0035697470884770155, + 0.0033925434108823538, + 0.0028381620068103075, + 0.002747260732576251, + 0.0035949619486927986, + 0.004335371777415276, + 0.003607057500630617, + 0.002810915932059288, + 0.0030221014749258757, + 0.0038390301633626223, + 0.002561794826760888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.011078749783337116 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336868055164814 + ], + "opposite_logprob_tokens": [ + -3.408686876296997 + ], + "logprob_reward": -0.0336868055164814, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336868055164814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10637114197015762, + 0.09853699803352356, + 0.04240686073899269, + 0.15793560445308685, + 0.17712749540805817, + 0.05837910994887352, + 0.03741247206926346, + 0.050219032913446426, + 0.042605720460414886, + 0.2290055751800537 + ], + "credit_map": [ + 0.006988649256527424, + 0.009766923263669014, + 0.007989813573658466, + 0.013803757727146149, + 0.021511662751436234, + 0.013220621272921562, + 0.017299402505159378, + 0.015790313482284546, + 0.008246530778706074, + 0.020434094592928886, + 0.010529087856411934, + 0.010659224353730679, + 0.014692876487970352, + 0.011554963886737823, + 0.010442977771162987, + 0.0119772432371974, + 0.0033626698423177004, + 0.004574809689074755, + 0.004909692797809839, + 0.007024209015071392, + 0.004575853236019611, + 0.005484702531248331, + 0.005801025312393904, + 0.006673900410532951, + 0.012769931927323341, + 0.013076224364340305, + 0.043037280440330505, + 0.02057829685509205, + 0.010640951804816723, + 0.014656235463917255, + 0.01463799923658371, + 0.02853867970407009, + 0.006964885629713535, + 0.008744364604353905, + 0.010152168571949005, + 0.009665963239967823, + 0.007364753633737564, + 0.008178521879017353, + 0.062225982546806335, + 0.06383085250854492, + 0.006688948255032301, + 0.005963797215372324, + 0.006864501629024744, + 0.005960480310022831, + 0.003673430997878313, + 0.005284344777464867, + 0.01571565866470337, + 0.008227949030697346, + 0.0026558139361441135, + 0.004180144518613815, + 0.005445428658276796, + 0.006880376487970352, + 0.003259364515542984, + 0.003825549501925707, + 0.007141805719584227, + 0.004023984540253878, + 0.003942861687391996, + 0.005188474897295237, + 0.007830430753529072, + 0.008883923292160034, + 0.005560895428061485, + 0.007646679878234863, + 0.006799033842980862, + 0.00436673266813159, + 0.0031149389687925577, + 0.004496010486036539, + 0.005150848068296909, + 0.004548538476228714, + 0.005376780871301889, + 0.0077814459800720215, + 0.005889019463211298, + 0.006248136050999165, + 0.02440462075173855, + 0.024600859731435776, + 0.03762977942824364, + 0.030825531110167503, + 0.030598435550928116, + 0.04259955883026123, + 0.021024877205491066, + 0.01732192002236843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.0336868055164814 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015700340270996 + ], + "opposite_logprob_tokens": [ + -1.7015700340270996 + ], + "logprob_reward": -0.2015700340270996, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015700340270996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12253234535455704, + 0.12540163099765778, + 0.06390029937028885, + 0.0514204278588295, + 0.15308670699596405, + 0.0994422510266304, + 0.09847040474414825, + 0.05849772319197655, + 0.16185738146305084, + 0.06539079546928406 + ], + "credit_map": [ + 0.00881222914904356, + 0.020334981381893158, + 0.0118380356580019, + 0.011957398615777493, + 0.022898515686392784, + 0.015245707705616951, + 0.017040520906448364, + 0.01440495066344738, + 0.012467772699892521, + 0.033214371651411057, + 0.012538835406303406, + 0.00875019934028387, + 0.013471538200974464, + 0.018612084910273552, + 0.01209193654358387, + 0.01425488106906414, + 0.006834006402641535, + 0.009675429202616215, + 0.006982191000133753, + 0.0071778264828026295, + 0.007326876744627953, + 0.008436858654022217, + 0.009331748820841312, + 0.008135363459587097, + 0.0036086835898458958, + 0.007387180346995592, + 0.010115879587829113, + 0.008323971182107925, + 0.003741353051736951, + 0.004908024333417416, + 0.007039344869554043, + 0.006295990664511919, + 0.009059668518602848, + 0.013146624900400639, + 0.02378288470208645, + 0.01770874485373497, + 0.011229321360588074, + 0.012706922367215157, + 0.026716576889157295, + 0.03873596712946892, + 0.010365946218371391, + 0.008063288405537605, + 0.007222238462418318, + 0.023689383640885353, + 0.00857089925557375, + 0.009209876880049706, + 0.018681034445762634, + 0.01363958790898323, + 0.010694132186472416, + 0.009440477006137371, + 0.011340657249093056, + 0.01915922947227955, + 0.008549651131033897, + 0.008226901292800903, + 0.015263191424310207, + 0.01579616591334343, + 0.004540329799056053, + 0.007007375825196505, + 0.008312322199344635, + 0.007835693657398224, + 0.006773314438760281, + 0.010337157174944878, + 0.006880921311676502, + 0.006810605991631746, + 0.011913114227354527, + 0.017397774383425713, + 0.018706105649471283, + 0.013037015683948994, + 0.02308933064341545, + 0.03302784636616707, + 0.023741988465189934, + 0.020944202318787575, + 0.005734336096793413, + 0.006924352142959833, + 0.006965530104935169, + 0.009640754200518131, + 0.009868876077234745, + 0.01133976224809885, + 0.008177416399121284, + 0.0067397733218967915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.2015700340270996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful overhead shot, shot in high-frame-rate slow motion, captures a breathtaking moment: a massive, arrow-shaped metallic alien spacecraft violently tears through a thick, roiling storm cloud. Composed of a dark, non-reflective alloy, its surface covered in intricate geometric paneling and scarred by atmospheric passage, its massive bulk carves a vast chasm in the sea of clouds as it plummets downward. Filmed at high frame rate, every wisp of torn cloud and vaporized debris tumbles and dissipates around the vessel's edge in exquisite detail. Along its flank, energy conduits pulse with eerie blue and crimson light, casting an ominous halo on the surrounding clouds. As the camera tilts steadily downward, we see the spacecraft plummeting with terrifying precision towards the shadowed expanse of gleaming skyscrapers below. The entire scene is imbued with an overwhelming sense of scale and apocalyptic tension.", + "seed": 31118, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_6/index_7955_seed_31118_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03536637369543314, + "vlm_logprob": -0.03536637369543314, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.03536637369543314, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007457631640136242 + ], + "opposite_logprob_tokens": [ + -5.257457733154297 + ], + "logprob_reward": -0.007457631640136242, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.007457631640136242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13477592170238495, + 0.23145076632499695, + 0.03876727446913719, + 0.08922963589429855, + 0.07616105675697327, + 0.21676509082317352, + 0.036069344729185104, + 0.07632362842559814, + 0.0278993621468544, + 0.07255787402391434 + ], + "credit_map": [ + 0.010947405360639095, + 0.01952315866947174, + 0.020830675959587097, + 0.012603381648659706, + 0.021076373755931854, + 0.023213762789964676, + 0.012558938935399055, + 0.014022242277860641, + 0.00686325877904892, + 0.004815347492694855, + 0.008031236939132214, + 0.0029876823537051678, + 0.009165079332888126, + 0.09835540503263474, + 0.0061376034282147884, + 0.09509516507387161, + 0.0033998903818428516, + 0.006414599716663361, + 0.005000502336770296, + 0.004799981601536274, + 0.007108602672815323, + 0.004129198379814625, + 0.004453073721379042, + 0.003461427753791213, + 0.00791738461703062, + 0.00960397906601429, + 0.013015561737120152, + 0.013879767619073391, + 0.009067830629646778, + 0.011447235010564327, + 0.009620904922485352, + 0.01467698160558939, + 0.008233058266341686, + 0.012457862496376038, + 0.01127476803958416, + 0.013994197361171246, + 0.0059729372151196, + 0.009477493353188038, + 0.00777832604944706, + 0.006972425617277622, + 0.022917041555047035, + 0.03511326387524605, + 0.03335178643465042, + 0.021450893953442574, + 0.02727377414703369, + 0.01923057809472084, + 0.03630940988659859, + 0.021118339151144028, + 0.0036175136920064688, + 0.004342695232480764, + 0.005452604498714209, + 0.00468845060095191, + 0.0048486110754311085, + 0.005422391463071108, + 0.004061965271830559, + 0.003635117784142494, + 0.007798574864864349, + 0.008721550926566124, + 0.008315720595419407, + 0.008017498068511486, + 0.013075475580990314, + 0.008230216801166534, + 0.014186284504830837, + 0.007978306151926517, + 0.0037483859341591597, + 0.002546685514971614, + 0.004451601300388575, + 0.0044805873185396194, + 0.0019472858402878046, + 0.002888834336772561, + 0.004698014818131924, + 0.0031379687134176493, + 0.005443653557449579, + 0.00517019210383296, + 0.0068297008983790874, + 0.00669853063300252, + 0.00952606275677681, + 0.012541361153125763, + 0.010537032037973404, + 0.015811344608664513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.007457631640136242 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331160008907318 + ], + "opposite_logprob_tokens": [ + -3.7733116149902344 + ], + "logprob_reward": -0.02331160008907318, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331160008907318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6303351521492004, + 0.1695781946182251, + 0.014538888819515705, + 0.024851350113749504, + 0.025275129824876785, + 0.06075631454586983, + 0.020126523450016975, + 0.02326269820332527, + 0.0105752469971776, + 0.020700519904494286 + ], + "credit_map": [ + 0.016663014888763428, + 0.021495141088962555, + 0.027390191331505775, + 0.01713240146636963, + 0.023456895723938942, + 0.26194706559181213, + 0.029405254870653152, + 0.2328452318906784, + 0.004593210760504007, + 0.002487711375579238, + 0.004530023783445358, + 0.0017999433912336826, + 0.006701508536934853, + 0.07396601140499115, + 0.003917319234460592, + 0.07158247381448746, + 0.001340332324616611, + 0.0014491910114884377, + 0.0025741548743098974, + 0.0011898785596713424, + 0.0032956525683403015, + 0.001797366188839078, + 0.0016684994334354997, + 0.001223814208060503, + 0.0021309133153408766, + 0.003205498680472374, + 0.003370477817952633, + 0.0030329758301377296, + 0.0027243029326200485, + 0.0035273258108645678, + 0.0029217982664704323, + 0.0039380560629069805, + 0.002557603642344475, + 0.003225624095648527, + 0.00441054766997695, + 0.0038937535136938095, + 0.0020844629034399986, + 0.0027620894834399223, + 0.003802128601819277, + 0.0025389213114976883, + 0.005631058942526579, + 0.008389877155423164, + 0.008417562581598759, + 0.004971605259925127, + 0.010393007658421993, + 0.005484485533088446, + 0.01171958725899458, + 0.005749127361923456, + 0.0016844093333929777, + 0.0023221438750624657, + 0.002974968170747161, + 0.002565812785178423, + 0.003008874598890543, + 0.0033025520388036966, + 0.0023890433367341757, + 0.001878717914223671, + 0.002303689019754529, + 0.0025222566910088062, + 0.0025514012668281794, + 0.0027189033571630716, + 0.004006108734756708, + 0.003115119645372033, + 0.0034391400404274464, + 0.0026060768868774176, + 0.001331518986262381, + 0.0011618912685662508, + 0.0013296017423272133, + 0.0026190464850515127, + 0.0010376130230724812, + 0.0010417583398520947, + 0.0011212844401597977, + 0.0009325326536782086, + 0.001238894066773355, + 0.00161093776114285, + 0.0029184564482420683, + 0.002387390937656164, + 0.0020554850343614817, + 0.003959666471928358, + 0.002571956254541874, + 0.003957733046263456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.02331160008907318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545948848128319 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.012545948848128319, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545948848128319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4896126985549927, + 0.270988792181015, + 0.01364953350275755, + 0.013301264494657516, + 0.02045711688697338, + 0.106254942715168, + 0.023581158369779587, + 0.016949163749814034, + 0.00964763667434454, + 0.03555769473314285 + ], + "credit_map": [ + 0.013625982217490673, + 0.016519760712981224, + 0.02250903844833374, + 0.01529800146818161, + 0.019558072090148926, + 0.20211565494537354, + 0.023144040256738663, + 0.17684218287467957, + 0.007563420571386814, + 0.004366918001323938, + 0.0068455711007118225, + 0.003250308334827423, + 0.011826056987047195, + 0.11778682470321655, + 0.006274519022554159, + 0.11307517439126968, + 0.0014337997417896986, + 0.001336001674644649, + 0.0031102080829441547, + 0.0012814337387681007, + 0.0019280686974525452, + 0.0015628671972081065, + 0.001518737175501883, + 0.001478418125770986, + 0.0014179161516949534, + 0.0014403491513803601, + 0.0015480973524972796, + 0.002355943899601698, + 0.002244287170469761, + 0.001226070336997509, + 0.0015268318820744753, + 0.0015417688991874456, + 0.0018541031749919057, + 0.004561414010822773, + 0.0032845321111381054, + 0.002953418530523777, + 0.0018689646385610104, + 0.0016791329253464937, + 0.00291200028732419, + 0.0013435520231723785, + 0.010324888862669468, + 0.016572045162320137, + 0.015784434974193573, + 0.008096370846033096, + 0.01885153539478779, + 0.009925385937094688, + 0.017775457352399826, + 0.008924824185669422, + 0.00199323333799839, + 0.0021995154675096273, + 0.002959508216008544, + 0.0031198179349303246, + 0.0033941580913960934, + 0.0030504553578794003, + 0.003514778334647417, + 0.003349692327901721, + 0.0015109704108908772, + 0.001741493702866137, + 0.0018599717877805233, + 0.0018363253911957145, + 0.0027080767322331667, + 0.0022255468647927046, + 0.0037723113782703876, + 0.0012944666668772697, + 0.001348172198049724, + 0.0010139181977137923, + 0.0013687448808923364, + 0.001406847732141614, + 0.0010885142255574465, + 0.00099589629098773, + 0.001274398877285421, + 0.001151144620962441, + 0.0029497062787413597, + 0.0025210995227098465, + 0.006030992139130831, + 0.004623644053936005, + 0.005216876044869423, + 0.005297111347317696, + 0.002722780220210552, + 0.006195484660565853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.012545948848128319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057763561606407 + ], + "opposite_logprob_tokens": [ + -3.8955776691436768 + ], + "logprob_reward": -0.02057763561606407, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057763561606407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5504079461097717, + 0.2657553553581238, + 0.014846405014395714, + 0.020298555493354797, + 0.014676953665912151, + 0.029033998027443886, + 0.01646105758845806, + 0.016742942854762077, + 0.006488576531410217, + 0.065288245677948 + ], + "credit_map": [ + 0.015213094651699066, + 0.018662787973880768, + 0.03396700322628021, + 0.019461864605545998, + 0.024157444015145302, + 0.22393427789211273, + 0.023760125041007996, + 0.19125130772590637, + 0.0071282535791397095, + 0.004049554467201233, + 0.00667081493884325, + 0.0030227117240428925, + 0.01059537474066019, + 0.11610279232263565, + 0.006184326019138098, + 0.11200150847434998, + 0.0014055058127269149, + 0.0014279313618317246, + 0.0023710629902780056, + 0.0012806669110432267, + 0.0023271730169653893, + 0.0018335068598389626, + 0.0024872294161468744, + 0.001713328412733972, + 0.0018304999684914947, + 0.001962244277819991, + 0.002673626411706209, + 0.004038033075630665, + 0.0019005516078323126, + 0.002999719465151429, + 0.0019837270956486464, + 0.0029101516120135784, + 0.0019336416153237224, + 0.002471222775056958, + 0.0021404854487627745, + 0.0027083384338766336, + 0.0010175873758271337, + 0.001164137851446867, + 0.002133066300302744, + 0.0011084717698395252, + 0.0033423856366425753, + 0.0047154733911156654, + 0.004291915334761143, + 0.002137600677087903, + 0.0042183962650597095, + 0.0033871892374008894, + 0.004527198616415262, + 0.002413836307823658, + 0.0014199893921613693, + 0.0016349831130355597, + 0.002932863077148795, + 0.002314044861122966, + 0.0017823970410972834, + 0.0019126783590763807, + 0.0023470406886190176, + 0.0021170603577047586, + 0.0014055317733436823, + 0.001569266663864255, + 0.0018131736433133483, + 0.0022610831074416637, + 0.0019032175187021494, + 0.0020107540767639875, + 0.00393744558095932, + 0.0018424693262204528, + 0.000910420436412096, + 0.0007956069312058389, + 0.001050940714776516, + 0.0007190536125563085, + 0.000538872554898262, + 0.0008658284787088633, + 0.0008912949124351144, + 0.0007165583083406091, + 0.005173178389668465, + 0.0056435177102684975, + 0.010392636060714722, + 0.01224369928240776, + 0.0067336950451135635, + 0.008275305852293968, + 0.006619143299758434, + 0.010207066312432289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.02057763561606407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293905228376389 + ], + "opposite_logprob_tokens": [ + -2.2379391193389893 + ], + "logprob_reward": -0.11293905228376389, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293905228376389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6246782541275024, + 0.1747889220714569, + 0.03592035919427872, + 0.016313184052705765, + 0.026827963069081306, + 0.031106989830732346, + 0.02546253614127636, + 0.031733497977256775, + 0.01694752648472786, + 0.01622079685330391 + ], + "credit_map": [ + 0.01684359274804592, + 0.01857435517013073, + 0.031343359500169754, + 0.017654404044151306, + 0.022187739610671997, + 0.26026487350463867, + 0.02703053690493107, + 0.23077943921089172, + 0.004783728159964085, + 0.0025783516466617584, + 0.004187426064163446, + 0.0019725181628018618, + 0.007029223721474409, + 0.07644219696521759, + 0.003933444619178772, + 0.07386204600334167, + 0.00297142518684268, + 0.003102918853983283, + 0.007915474474430084, + 0.0028255609795451164, + 0.005477956961840391, + 0.004530941601842642, + 0.005133840721100569, + 0.00396224157884717, + 0.0015351887559518218, + 0.0015764478594064713, + 0.002504603937268257, + 0.003529902547597885, + 0.0014319049660116434, + 0.0017712959088385105, + 0.002047726884484291, + 0.0019161157542839646, + 0.003148737596347928, + 0.003960563801229, + 0.004562784917652607, + 0.0028442339971661568, + 0.0015347599983215332, + 0.0015646986430510879, + 0.005771328695118427, + 0.0034408580977469683, + 0.005361526273190975, + 0.004115494433790445, + 0.004121693316847086, + 0.004494179040193558, + 0.0023917125072330236, + 0.003616345813497901, + 0.004283502697944641, + 0.0027225378435105085, + 0.002471282845363021, + 0.002869588788598776, + 0.003992823883891106, + 0.004307142924517393, + 0.003135944250971079, + 0.0017235273262485862, + 0.0034302149433642626, + 0.003532013390213251, + 0.0028449525125324726, + 0.004497845657169819, + 0.003519065212458372, + 0.0032274548429995775, + 0.004304620902985334, + 0.003836072050035, + 0.007257778663188219, + 0.002245710464194417, + 0.003042029682546854, + 0.0018201112980023026, + 0.002672189148142934, + 0.002201934577897191, + 0.0014665579656139016, + 0.001634360640309751, + 0.0021967969369143248, + 0.0019135460024699569, + 0.0011497708037495613, + 0.001841870485804975, + 0.0019503093790262938, + 0.0022171922028064728, + 0.002160176634788513, + 0.0025998333003371954, + 0.0017386481631547213, + 0.0025629966985434294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.11293905228376389 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful overhead shot, shot in high-frame-rate slow motion, captures a breathtaking moment: a massive, arrow-shaped metallic alien spacecraft violently tears through a thick, roiling storm cloud. Composed of a dark, non-reflective alloy, its surface covered in intricate geometric paneling and scarred by atmospheric passage, its massive bulk carves a vast chasm in the sea of clouds as it plummets downward. Filmed at high frame rate, every wisp of torn cloud and vaporized debris tumbles and dissipates around the vessel's edge in exquisite detail. Along its flank, energy conduits pulse with eerie blue and crimson light, casting an ominous halo on the surrounding clouds. As the camera tilts steadily downward, we see the spacecraft plummeting with terrifying precision towards the shadowed expanse of gleaming skyscrapers below. The entire scene is imbued with an overwhelming sense of scale and apocalyptic tension.", + "seed": 31119, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_6/index_7955_seed_31119_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.030324048455804588, + "vlm_logprob": -0.030324048455804588, + "vlm_yes_no_margin": 3.9999999523162844, + "vlm_token_credit": -0.030324048455804588, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013400605879724026 + ], + "opposite_logprob_tokens": [ + -4.638400554656982 + ], + "logprob_reward": -0.013400605879724026, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013400605879724026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09647373855113983, + 0.08227845281362534, + 0.05406734719872475, + 0.08912433683872223, + 0.29162952303886414, + 0.04198404401540756, + 0.052018653601408005, + 0.061641283333301544, + 0.11262791603803635, + 0.11815472692251205 + ], + "credit_map": [ + 0.010443318635225296, + 0.014215025119483471, + 0.01560945063829422, + 0.008962612599134445, + 0.011144320480525494, + 0.014871126972138882, + 0.014057796448469162, + 0.007170075550675392, + 0.0071455636061728, + 0.013065243139863014, + 0.010905381292104721, + 0.01272502914071083, + 0.011740929447114468, + 0.009753298945724964, + 0.009882023558020592, + 0.0070609766989946365, + 0.004797224421054125, + 0.009367430582642555, + 0.006750915665179491, + 0.007844013161957264, + 0.006361108738929033, + 0.007140756584703922, + 0.0063658724538981915, + 0.005440026056021452, + 0.006737984716892242, + 0.007571902591735125, + 0.012317772954702377, + 0.011641851626336575, + 0.010556207969784737, + 0.010722953826189041, + 0.019112803041934967, + 0.01046285405755043, + 0.020794671028852463, + 0.018826588988304138, + 0.08164027333259583, + 0.02061060257256031, + 0.026299072429537773, + 0.022253239527344704, + 0.0170421339571476, + 0.08416292816400528, + 0.004579581320285797, + 0.003203614382073283, + 0.005626754369586706, + 0.005062723532319069, + 0.005394045729190111, + 0.007305834908038378, + 0.006131615489721298, + 0.004679868929088116, + 0.004612335469573736, + 0.006637692917138338, + 0.006033099256455898, + 0.005352208390831947, + 0.010411623865365982, + 0.007779974490404129, + 0.004285804461687803, + 0.006905907299369574, + 0.005020490847527981, + 0.0045798965729773045, + 0.006640007719397545, + 0.01288264337927103, + 0.006441619712859392, + 0.010162434540688992, + 0.005970850121229887, + 0.009943334385752678, + 0.012576374225318432, + 0.010883758775889874, + 0.01696876250207424, + 0.018692612648010254, + 0.012093582190573215, + 0.013116179034113884, + 0.013836379162967205, + 0.014460258185863495, + 0.007754439953714609, + 0.016675401479005814, + 0.01894432120025158, + 0.017543885856866837, + 0.014878314919769764, + 0.013858994469046593, + 0.01460589561611414, + 0.013893465511500835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.013400605879724026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026368487626314163 + ], + "opposite_logprob_tokens": [ + -3.6513683795928955 + ], + "logprob_reward": -0.026368487626314163, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026368487626314163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15459023416042328, + 0.09318876266479492, + 0.06771602481603622, + 0.04325955733656883, + 0.25165802240371704, + 0.05407901480793953, + 0.0595114529132843, + 0.06150449439883232, + 0.11606094986200333, + 0.0984315276145935 + ], + "credit_map": [ + 0.011230665259063244, + 0.019922416657209396, + 0.015011887066066265, + 0.024006463587284088, + 0.019424261525273323, + 0.028691820800304413, + 0.021127643063664436, + 0.01517509389668703, + 0.008888193406164646, + 0.01493664737790823, + 0.00941177736967802, + 0.013707658275961876, + 0.01686767116189003, + 0.011117206886410713, + 0.010393085889518261, + 0.007866522297263145, + 0.004282072186470032, + 0.012707243673503399, + 0.006347293499857187, + 0.008074558340013027, + 0.012153995223343372, + 0.007465758360922337, + 0.009945500642061234, + 0.006739597301930189, + 0.003219193546101451, + 0.0043204305693507195, + 0.006390980910509825, + 0.005096448585391045, + 0.006230535916984081, + 0.004437149036675692, + 0.009770780801773071, + 0.0037940365727990866, + 0.019361665472388268, + 0.009909859858453274, + 0.08275525271892548, + 0.020435404032468796, + 0.01523812860250473, + 0.016814738512039185, + 0.009311798959970474, + 0.07783117145299911, + 0.00520629808306694, + 0.0068622613325715065, + 0.0055825007148087025, + 0.005841731559485197, + 0.009012226946651936, + 0.006915080826729536, + 0.0077322497963905334, + 0.006926666479557753, + 0.008393584750592709, + 0.004344246815890074, + 0.007929438725113869, + 0.012202205136418343, + 0.008968187496066093, + 0.008217137306928635, + 0.003504108637571335, + 0.0059525445103645325, + 0.005322431679815054, + 0.005218958016484976, + 0.009086567908525467, + 0.011546904221177101, + 0.007182778790593147, + 0.010058361105620861, + 0.005860115867108107, + 0.007228377740830183, + 0.009377955459058285, + 0.010802880860865116, + 0.017662636935710907, + 0.02007896825671196, + 0.011315124109387398, + 0.014366568997502327, + 0.017589012160897255, + 0.014867798425257206, + 0.00724450359120965, + 0.013518914580345154, + 0.021563440561294556, + 0.011214048601686954, + 0.013530493713915348, + 0.010600278154015541, + 0.009430412203073502, + 0.011329438537359238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.026368487626314163 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004641829524189234 + ], + "opposite_logprob_tokens": [ + -5.379642009735107 + ], + "logprob_reward": -0.004641829524189234, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004641829524189234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09276837110519409, + 0.03878443315625191, + 0.03391231223940849, + 0.08622092008590698, + 0.20286056399345398, + 0.0640067309141159, + 0.0982750877737999, + 0.09143644571304321, + 0.06390560418367386, + 0.22782954573631287 + ], + "credit_map": [ + 0.011511710472404957, + 0.010320886969566345, + 0.013918602839112282, + 0.014378613792359829, + 0.011356845498085022, + 0.012687904760241508, + 0.009622077457606792, + 0.008971738629043102, + 0.0034535236191004515, + 0.005928257945924997, + 0.0047050765715539455, + 0.005188183393329382, + 0.005030140280723572, + 0.004967200569808483, + 0.005188171751797199, + 0.004323879256844521, + 0.001930835540406406, + 0.004736419767141342, + 0.003997605759650469, + 0.005865807645022869, + 0.005152582656592131, + 0.004056497011333704, + 0.005263317842036486, + 0.0029092442709952593, + 0.006201738957315683, + 0.01565808244049549, + 0.010560956783592701, + 0.011838576756417751, + 0.0092349573969841, + 0.009641901589930058, + 0.013461757451295853, + 0.009622955694794655, + 0.015794437378644943, + 0.008417824283242226, + 0.06760141998529434, + 0.015411521308124065, + 0.009256287477910519, + 0.012859330512583256, + 0.008317827247083187, + 0.06520191580057144, + 0.007417627144604921, + 0.004827149212360382, + 0.007160016801208258, + 0.007278618402779102, + 0.010433157905936241, + 0.010175298899412155, + 0.008366046473383904, + 0.008348810486495495, + 0.013255765661597252, + 0.010447188280522823, + 0.013980879448354244, + 0.018213532865047455, + 0.010560570284724236, + 0.013455545529723167, + 0.008487300015985966, + 0.009874306619167328, + 0.007057329174131155, + 0.011511210352182388, + 0.00979703851044178, + 0.012984507717192173, + 0.01723739691078663, + 0.014005721546709538, + 0.008789053186774254, + 0.01005418598651886, + 0.0064164926297962666, + 0.007298621814697981, + 0.008753250353038311, + 0.009270559065043926, + 0.007227931171655655, + 0.008573236875236034, + 0.00854609627276659, + 0.007819416001439095, + 0.014498402364552021, + 0.03181503340601921, + 0.033792752772569656, + 0.0328243151307106, + 0.029072649776935577, + 0.03578001633286476, + 0.02202673815190792, + 0.028019635006785393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.004641829524189234 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180983141064644 + ], + "opposite_logprob_tokens": [ + -4.018180847167969 + ], + "logprob_reward": -0.018180983141064644, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018180983141064644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10223373770713806, + 0.04205964133143425, + 0.020413482561707497, + 0.035355061292648315, + 0.28829339146614075, + 0.07313358783721924, + 0.059411827474832535, + 0.03516924008727074, + 0.07213681936264038, + 0.2717932164669037 + ], + "credit_map": [ + 0.010046808049082756, + 0.012702427804470062, + 0.012545385397970676, + 0.015346964821219444, + 0.010046609677374363, + 0.016296034678816795, + 0.014286376535892487, + 0.010963136330246925, + 0.0033970412332564592, + 0.005628306418657303, + 0.0057613421231508255, + 0.006393967662006617, + 0.005525283049792051, + 0.005610269494354725, + 0.005345999263226986, + 0.004397430457174778, + 0.0013599845115095377, + 0.0025450114626437426, + 0.0027442695572972298, + 0.0027529371436685324, + 0.003584526712074876, + 0.0027148108929395676, + 0.002681254642084241, + 0.002030686940997839, + 0.0028071890119463205, + 0.004931612405925989, + 0.004674198571592569, + 0.004846296738833189, + 0.004534339066594839, + 0.004974814131855965, + 0.004813244566321373, + 0.003773367265239358, + 0.023132022470235825, + 0.010916645638644695, + 0.09540124982595444, + 0.023081718012690544, + 0.01586480624973774, + 0.019367879256606102, + 0.012614458799362183, + 0.0879145935177803, + 0.008576717227697372, + 0.007606348022818565, + 0.009671099483966827, + 0.009295820258557796, + 0.007754037156701088, + 0.01064968015998602, + 0.009272041730582714, + 0.010307841934263706, + 0.004792294930666685, + 0.006034939084202051, + 0.006075686309486628, + 0.006995258387178183, + 0.01406790129840374, + 0.00776544539257884, + 0.005550271365791559, + 0.00813002698123455, + 0.0025969359558075666, + 0.003129994962364435, + 0.004400112200528383, + 0.007214962504804134, + 0.0036656588781625032, + 0.005351179279386997, + 0.0031876862049102783, + 0.005622709635645151, + 0.005522825755178928, + 0.00612615654245019, + 0.011017386801540852, + 0.010454674251377583, + 0.009504176676273346, + 0.010439404286444187, + 0.009209680370986462, + 0.00986251700669527, + 0.021067067980766296, + 0.03126286342740059, + 0.0558645986020565, + 0.03901529312133789, + 0.0282822884619236, + 0.026551811024546623, + 0.04044939577579498, + 0.029299888759851456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.018180983141064644 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, arrow-shaped metallic alien spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft's surface dark, non-reflective, and covered in intricate geometric paneling with scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are energy conduits visible along the spacecraft's flank pulsing with blue and crimson light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt steadily downward to show the spacecraft plummeting toward skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video shot in high-frame-rate slow motion showing torn clouds and debris tumbling in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902833610773087 + ], + "opposite_logprob_tokens": [ + -2.4640283584594727 + ], + "logprob_reward": -0.08902833610773087, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902833610773087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10524109750986099, + 0.05443544313311577, + 0.057553451508283615, + 0.046818096190690994, + 0.4901733100414276, + 0.05867147818207741, + 0.027793334797024727, + 0.02106533572077751, + 0.0697590708732605, + 0.06848941743373871 + ], + "credit_map": [ + 0.01108970120549202, + 0.0151401087641716, + 0.01970653608441353, + 0.011897224001586437, + 0.010909197852015495, + 0.012333768419921398, + 0.012789642438292503, + 0.011374927125871181, + 0.004849347751587629, + 0.009791437536478043, + 0.005406699609011412, + 0.007929275743663311, + 0.010645652189850807, + 0.005743080284446478, + 0.0057051582261919975, + 0.004364790860563517, + 0.0037351204082369804, + 0.010227099992334843, + 0.006758905481547117, + 0.008640803396701813, + 0.009893587790429592, + 0.006380595266819, + 0.006653339136391878, + 0.005264002829790115, + 0.003651753067970276, + 0.006866768002510071, + 0.006569303572177887, + 0.005304819904267788, + 0.004244043026119471, + 0.0069447290152311325, + 0.009128919802606106, + 0.004107761662453413, + 0.04059455916285515, + 0.01647612266242504, + 0.1673383265733719, + 0.04198092967271805, + 0.021879898384213448, + 0.03077130764722824, + 0.017177639529109, + 0.15395450592041016, + 0.007291668094694614, + 0.004675588104873896, + 0.008690417744219303, + 0.009909544140100479, + 0.005171128548681736, + 0.009593451395630836, + 0.005828628782182932, + 0.007511048577725887, + 0.0036143476609140635, + 0.003181990934535861, + 0.003443626221269369, + 0.0031182130333036184, + 0.002854611724615097, + 0.0052847424522042274, + 0.003072642255574465, + 0.00322315888479352, + 0.001939826994203031, + 0.0022920698393136263, + 0.0024208021350204945, + 0.003932427614927292, + 0.0034229420125484467, + 0.002891944255679846, + 0.0016724991146475077, + 0.0024928248021751642, + 0.006959978956729174, + 0.007116006687283516, + 0.009885173290967941, + 0.012926241382956505, + 0.007726631127297878, + 0.010385622270405293, + 0.0068386271595954895, + 0.007920791395008564, + 0.005444094073027372, + 0.009158905595541, + 0.0136590376496315, + 0.007437790744006634, + 0.00884866900742054, + 0.009077265858650208, + 0.007759410422295332, + 0.007104243151843548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.08902833610773087 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_7/index_tensor([6678, 6678])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_7/index_tensor([6678, 6678])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..544c604995971b0570e6ac5ed82aa9d7a8e272cd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_7/index_tensor([6678, 6678])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera, in a smooth, 360-degree panning shot, slowly pans around a man in a courtyard suspended above a sea of clouds. In the center of the courtyard, this handsome man with an ethereal presence, his waist-length jet-black hair gently brushes his chiseled features. He wears a rich, smooth silk robe of a deep midnight blue. The wide cuffs and lapels are meticulously embroidered with intricate patterns of stars and flowing clouds in silver thread, shimmering with a subtle luster in the dim light. With focused, careful hands, he holds a delicate, crescent-shaped silver vessel. Steadily and deliberately, he tilts it, and a dazzling stream of light, composed of countless shimmering particles, like a liquid galaxy, slowly flows out, pouring down upon a unique flower growing in a white marble pot before him. The flower is composed entirely of transparent crystal, its petals perfectly faceted. As the starlight touches it, the entire blossom instantly illuminates from within, emitting a rainbow-like, shifting halo. As the camera rotates, the background scene unfolds: the courtyard is bordered by an intricately carved white jade railing, its texture as gentle as jade. Organically entwined with exotic vines emitting a soft blue glow, their leaves seemingly composed of solidified light. Beyond the railing lies an endless, swirling sea of clouds, dyed a lavender by celestial light. The deep night sky is dotted with distant nebulae and stars, creating an atmosphere of tranquility, sacredness, and a sense of wonder.", + "seed": 31784, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_7/index_6678_seed_31784_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08086488991975785, + "vlm_logprob": -0.08086488991975785, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.08086488991975785, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32519879937171936 + ], + "opposite_logprob_tokens": [ + -1.325198769569397 + ], + "logprob_reward": -0.32519879937171936, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32519879937171936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09512042999267578, + 0.09904950857162476, + 0.11500591039657593, + 0.1679254025220871, + 0.13626720011234283, + 0.15978428721427917, + 0.052523959428071976, + 0.03178662806749344, + 0.09154178202152252, + 0.05099491402506828 + ], + "credit_map": [ + 0.010641973465681076, + 0.0099262073636055, + 0.00909073743969202, + 0.0125584052875638, + 0.00952635332942009, + 0.009376066736876965, + 0.017972296103835106, + 0.016028396785259247, + 0.011004257947206497, + 0.013359935022890568, + 0.0077773150987923145, + 0.018535329028964043, + 0.012245679274201393, + 0.016359969973564148, + 0.010145057924091816, + 0.009621965698897839, + 0.008760892786085606, + 0.018748579546809196, + 0.010098782368004322, + 0.01454873476177454, + 0.02205781079828739, + 0.020183011889457703, + 0.011372116394340992, + 0.009235979057848454, + 0.011535638011991978, + 0.019195396453142166, + 0.023565689101815224, + 0.021958259865641594, + 0.02441711165010929, + 0.028461826965212822, + 0.017091691493988037, + 0.021699784323573112, + 0.016695396974682808, + 0.014177575707435608, + 0.014577188529074192, + 0.012542111799120903, + 0.021286901086568832, + 0.021339979022741318, + 0.02134336344897747, + 0.014304682612419128, + 0.020285723730921745, + 0.012758725322782993, + 0.019467908889055252, + 0.021500442177057266, + 0.022684089839458466, + 0.030289076268672943, + 0.018504606559872627, + 0.01429371815174818, + 0.004805644508451223, + 0.006386715453118086, + 0.007800383027642965, + 0.009985804557800293, + 0.004654865246266127, + 0.010360891930758953, + 0.0033816855866461992, + 0.005147966556251049, + 0.003966344054788351, + 0.006653050892055035, + 0.0033408068120479584, + 0.004775709006935358, + 0.004195862449705601, + 0.0031957258470356464, + 0.0024781664833426476, + 0.0031809641513973475, + 0.029326066374778748, + 0.036257728934288025, + 0.006884670816361904, + 0.006999914068728685, + 0.002838707761839032, + 0.0034175142645835876, + 0.003703750902786851, + 0.002113434486091137, + 0.011438713409006596, + 0.015523736365139484, + 0.00536548625677824, + 0.005495359655469656, + 0.004013387486338615, + 0.002261368092149496, + 0.005064203403890133, + 0.0018326605204492807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.32519879937171936 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328958734869957 + ], + "opposite_logprob_tokens": [ + -3.773289680480957 + ], + "logprob_reward": -0.02328958734869957, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328958734869957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21189962327480316, + 0.07166851311922073, + 0.07630924135446548, + 0.07332602888345718, + 0.06799962371587753, + 0.0695507600903511, + 0.04018932208418846, + 0.3124752342700958, + 0.026171980425715446, + 0.050409652292728424 + ], + "credit_map": [ + 0.03712339326739311, + 0.028302986174821854, + 0.02475765347480774, + 0.034553758800029755, + 0.020541291683912277, + 0.020885154604911804, + 0.02540653571486473, + 0.02032887004315853, + 0.0077730887569487095, + 0.008752452209591866, + 0.004118218552321196, + 0.009221206419169903, + 0.009977092035114765, + 0.011411481536924839, + 0.013116969726979733, + 0.007298005744814873, + 0.007719823159277439, + 0.015862172469496727, + 0.007501391693949699, + 0.009633492678403854, + 0.012062059715390205, + 0.009128409437835217, + 0.008206315338611603, + 0.006195577792823315, + 0.005351160187274218, + 0.012624604627490044, + 0.009546034969389439, + 0.010630463249981403, + 0.010364449582993984, + 0.007182952482253313, + 0.009232875891029835, + 0.008393487893044949, + 0.007893898524343967, + 0.008982188999652863, + 0.008226797915995121, + 0.008482931181788445, + 0.008831260725855827, + 0.008748280815780163, + 0.008395809680223465, + 0.008438455872237682, + 0.013029281049966812, + 0.004690987057983875, + 0.009197991341352463, + 0.005168502684682608, + 0.006972521543502808, + 0.018743785098195076, + 0.006159469019621611, + 0.005588220898061991, + 0.0036072907969355583, + 0.006229744758456945, + 0.004841103684157133, + 0.004521390423178673, + 0.006682149600237608, + 0.006619870662689209, + 0.0038926599081605673, + 0.0037951136473566294, + 0.031975846737623215, + 0.10016655176877975, + 0.02687934786081314, + 0.0947432592511177, + 0.017423629760742188, + 0.016240455210208893, + 0.013057376258075237, + 0.011988792568445206, + 0.005981738213449717, + 0.007398654241114855, + 0.003138655796647072, + 0.0033226816449314356, + 0.0013837877195328474, + 0.0012858071131631732, + 0.001979889813810587, + 0.0016807663487270474, + 0.009120299480855465, + 0.013283154927194118, + 0.006343707907944918, + 0.005826358217746019, + 0.0053388201631605625, + 0.0035953805781900883, + 0.0038950960151851177, + 0.0030068361666053534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.02328958734869957 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107827853411436 + ], + "opposite_logprob_tokens": [ + -4.511078357696533 + ], + "logprob_reward": -0.01107827853411436, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107827853411436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1442415565252304, + 0.1458398699760437, + 0.1176333799958229, + 0.1308492124080658, + 0.1195712611079216, + 0.08963317424058914, + 0.04427558183670044, + 0.06108508259057999, + 0.06848553568124771, + 0.0783853754401207 + ], + "credit_map": [ + 0.01955929398536682, + 0.018479682505130768, + 0.017798809334635735, + 0.021637191995978355, + 0.014128262177109718, + 0.014186650514602661, + 0.020560022443532944, + 0.017891641706228256, + 0.015147031284868717, + 0.016322709619998932, + 0.012365782633423805, + 0.03618187457323074, + 0.015129207633435726, + 0.01852405071258545, + 0.01699877344071865, + 0.015170431695878506, + 0.008724604733288288, + 0.013949536718428135, + 0.012779525481164455, + 0.015450295992195606, + 0.025466229766607285, + 0.014511877670884132, + 0.015730975195765495, + 0.011020335368812084, + 0.01220864150673151, + 0.01742139458656311, + 0.015627380460500717, + 0.014044472016394138, + 0.01994183659553528, + 0.02224266715347767, + 0.014866049401462078, + 0.014496770687401295, + 0.013336404226720333, + 0.013095933943986893, + 0.018089881166815758, + 0.009667863138020039, + 0.01950552873313427, + 0.015829773619771004, + 0.015287373214960098, + 0.014758502133190632, + 0.008697733283042908, + 0.007559000980108976, + 0.01538728829473257, + 0.011923986487090588, + 0.012477451004087925, + 0.01343400776386261, + 0.00987154059112072, + 0.010282169096171856, + 0.003562872763723135, + 0.006311601493507624, + 0.0059213582426309586, + 0.007661763112992048, + 0.003943135961890221, + 0.00861645769327879, + 0.0037037632428109646, + 0.004554628394544125, + 0.005518596153706312, + 0.011097853071987629, + 0.007089550141245127, + 0.014447907917201519, + 0.0067069036886096, + 0.004470047075301409, + 0.005431331694126129, + 0.006322894711047411, + 0.019621608778834343, + 0.0230507031083107, + 0.00634376984089613, + 0.0074488022364676, + 0.002443720819428563, + 0.0037299655377864838, + 0.0034147300757467747, + 0.0024322341196238995, + 0.016254976391792297, + 0.024779371917247772, + 0.009348548948764801, + 0.009059125557541847, + 0.005607771687209606, + 0.004525146447122097, + 0.00543187465518713, + 0.0033785649575293064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.01107827853411436 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336914137005806 + ], + "opposite_logprob_tokens": [ + -3.40869140625 + ], + "logprob_reward": -0.0336914137005806, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336914137005806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11660563200712204, + 0.10415563732385635, + 0.0931231677532196, + 0.0657123327255249, + 0.2387993335723877, + 0.05925631895661354, + 0.04105420410633087, + 0.07369515299797058, + 0.11567214131355286, + 0.09192609041929245 + ], + "credit_map": [ + 0.010080158710479736, + 0.012165887281298637, + 0.009773995727300644, + 0.017873795703053474, + 0.011120766401290894, + 0.014034596271812916, + 0.02321995608508587, + 0.018336480483412743, + 0.008515224792063236, + 0.01223560981452465, + 0.007204996421933174, + 0.009855316020548344, + 0.015027801506221294, + 0.0212031789124012, + 0.017553050071001053, + 0.012560462579131126, + 0.007659658323973417, + 0.010709508322179317, + 0.007233424112200737, + 0.008117961697280407, + 0.021222926676273346, + 0.015681413933634758, + 0.011745801195502281, + 0.010752472095191479, + 0.004896171391010284, + 0.010009629651904106, + 0.010247081518173218, + 0.00892410147935152, + 0.010087964124977589, + 0.008888812735676765, + 0.005995086394250393, + 0.006663480773568153, + 0.02441837638616562, + 0.01896052248775959, + 0.03967302665114403, + 0.018799258396029472, + 0.03493517264723778, + 0.028846193104982376, + 0.03862517699599266, + 0.034541621804237366, + 0.0054433876648545265, + 0.004683355800807476, + 0.010268921963870525, + 0.009798254817724228, + 0.008186486549675465, + 0.008362299762666225, + 0.006912552285939455, + 0.005601060576736927, + 0.0036053818184882402, + 0.010504595004022121, + 0.00453542685136199, + 0.004952088929712772, + 0.0031972804572433233, + 0.008777043782174587, + 0.002487104618921876, + 0.002995280781760812, + 0.008665399625897408, + 0.020359210669994354, + 0.007206982001662254, + 0.012195782735943794, + 0.007489335723221302, + 0.006705695763230324, + 0.004731644876301289, + 0.006341107189655304, + 0.03322987258434296, + 0.04186990484595299, + 0.010559633374214172, + 0.012594501487910748, + 0.003428080817684531, + 0.00510118342936039, + 0.005374319851398468, + 0.0035146395675837994, + 0.016542932018637657, + 0.028691809624433517, + 0.013506201095879078, + 0.012049208395183086, + 0.005462976638227701, + 0.004965518601238728, + 0.006656050216406584, + 0.004051393363624811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.0336914137005806 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066370643675327 + ], + "opposite_logprob_tokens": [ + -4.511066436767578 + ], + "logprob_reward": -0.011066370643675327, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066370643675327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10737036168575287, + 0.2987707257270813, + 0.07780438661575317, + 0.0639985054731369, + 0.029145462438464165, + 0.04101823270320892, + 0.034257445484399796, + 0.07486382126808167, + 0.11920610070228577, + 0.1535649448633194 + ], + "credit_map": [ + 0.01115359552204609, + 0.012929915450513363, + 0.009981700219213963, + 0.013639879412949085, + 0.010728641413152218, + 0.015335923060774803, + 0.016357405111193657, + 0.01724330149590969, + 0.027997851371765137, + 0.034287068992853165, + 0.017382366582751274, + 0.030584435909986496, + 0.04491071403026581, + 0.06128600239753723, + 0.04676340892910957, + 0.035558853298425674, + 0.004760650917887688, + 0.008645082823932171, + 0.005332675762474537, + 0.006575300358235836, + 0.020762324333190918, + 0.01907046139240265, + 0.0052597662433981895, + 0.0073981210589408875, + 0.006118530873209238, + 0.009317263960838318, + 0.005956205073744059, + 0.008362406864762306, + 0.009813967160880566, + 0.010555490851402283, + 0.007166420575231314, + 0.0067082238383591175, + 0.0038545478601008654, + 0.0019391195382922888, + 0.004193427041172981, + 0.0024227232206612825, + 0.0033638221211731434, + 0.007231742609292269, + 0.0027931039221584797, + 0.003346974728628993, + 0.00764564098790288, + 0.003037608228623867, + 0.005592924077063799, + 0.003817490767687559, + 0.004859846085309982, + 0.00730782700702548, + 0.004109738394618034, + 0.004647158086299896, + 0.003185616573318839, + 0.004198689013719559, + 0.00503502506762743, + 0.006744578946381807, + 0.003262546146288514, + 0.006166344508528709, + 0.0023961965925991535, + 0.003268449567258358, + 0.009130213409662247, + 0.025820592418313026, + 0.006008295342326164, + 0.01819099858403206, + 0.004277682863175869, + 0.0048354570753872395, + 0.0032298557925969362, + 0.0033707318361848593, + 0.031229214742779732, + 0.040495868772268295, + 0.014193027280271053, + 0.015262085944414139, + 0.003580756252631545, + 0.004735247697681189, + 0.005246673244982958, + 0.00446323212236166, + 0.022019391879439354, + 0.037929169833660126, + 0.025856008753180504, + 0.025335267186164856, + 0.010078893043100834, + 0.009189664386212826, + 0.014809256419539452, + 0.008347301743924618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.011066370643675327 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera, in a smooth, 360-degree panning shot, slowly pans around a man in a courtyard suspended above a sea of clouds. In the center of the courtyard, this handsome man with an ethereal presence, his waist-length jet-black hair gently brushes his chiseled features. He wears a rich, smooth silk robe of a deep midnight blue. The wide cuffs and lapels are meticulously embroidered with intricate patterns of stars and flowing clouds in silver thread, shimmering with a subtle luster in the dim light. With focused, careful hands, he holds a delicate, crescent-shaped silver vessel. Steadily and deliberately, he tilts it, and a dazzling stream of light, composed of countless shimmering particles, like a liquid galaxy, slowly flows out, pouring down upon a unique flower growing in a white marble pot before him. The flower is composed entirely of transparent crystal, its petals perfectly faceted. As the starlight touches it, the entire blossom instantly illuminates from within, emitting a rainbow-like, shifting halo. As the camera rotates, the background scene unfolds: the courtyard is bordered by an intricately carved white jade railing, its texture as gentle as jade. Organically entwined with exotic vines emitting a soft blue glow, their leaves seemingly composed of solidified light. Beyond the railing lies an endless, swirling sea of clouds, dyed a lavender by celestial light. The deep night sky is dotted with distant nebulae and stars, creating an atmosphere of tranquility, sacredness, and a sense of wonder.", + "seed": 31785, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_7/index_6678_seed_31785_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10963228065520525, + "vlm_logprob": -0.10963228065520525, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.10963228065520525, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4884224534034729 + ], + "opposite_logprob_tokens": [ + -0.9884224534034729 + ], + "logprob_reward": -0.4884224534034729, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4884224534034729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.249534010887146, + 0.054915353655815125, + 0.10157345235347748, + 0.0680105909705162, + 0.08441577851772308, + 0.12777090072631836, + 0.08189470320940018, + 0.03561525046825409, + 0.12615038454532623, + 0.07011952996253967 + ], + "credit_map": [ + 0.023295003920793533, + 0.04257535934448242, + 0.025917934253811836, + 0.029612895101308823, + 0.046547938138246536, + 0.02717275172472, + 0.03534179925918579, + 0.019070344045758247, + 0.0067878179252147675, + 0.00815578456968069, + 0.006417127326130867, + 0.006933709606528282, + 0.006535491440445185, + 0.006406699772924185, + 0.007591338362544775, + 0.006087386514991522, + 0.008115904405713081, + 0.009881016798317432, + 0.00972962100058794, + 0.009282599203288555, + 0.012975617311894894, + 0.011128580197691917, + 0.021041925996541977, + 0.01941818930208683, + 0.005231432616710663, + 0.00922806654125452, + 0.007047886028885841, + 0.007316497154533863, + 0.007526459638029337, + 0.006670563016086817, + 0.008203010074794292, + 0.016786683350801468, + 0.0064364769496023655, + 0.00768879521638155, + 0.01134810596704483, + 0.008489176630973816, + 0.009724804200232029, + 0.020008722320199013, + 0.010028762742877007, + 0.01069093868136406, + 0.009895450435578823, + 0.01887493208050728, + 0.009687530808150768, + 0.018834399059414864, + 0.01210156548768282, + 0.02530791610479355, + 0.01666978932917118, + 0.01639932207763195, + 0.007017994299530983, + 0.00954371690750122, + 0.008328016847372055, + 0.00990572851151228, + 0.012401780113577843, + 0.009752118960022926, + 0.01200828980654478, + 0.012937061488628387, + 0.003582162084057927, + 0.004619206767529249, + 0.005600710399448872, + 0.0038502488750964403, + 0.0037140422500669956, + 0.005302696023136377, + 0.005129690282046795, + 0.003816493321210146, + 0.009134544059634209, + 0.005162225104868412, + 0.045659441500902176, + 0.006252153776586056, + 0.005203781183809042, + 0.011041433550417423, + 0.010584523901343346, + 0.03311228007078171, + 0.008405864238739014, + 0.005571440327912569, + 0.012613451108336449, + 0.006271884310990572, + 0.007305310573428869, + 0.013063473626971245, + 0.006155952345579863, + 0.010732152499258518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.4884224534034729 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545241974294186 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545241974294186, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545241974294186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19477425515651703, + 0.04357235133647919, + 0.15010729432106018, + 0.05971142277121544, + 0.08394396305084229, + 0.06497365981340408, + 0.04934119060635567, + 0.12867994606494904, + 0.1208682507276535, + 0.104027658700943 + ], + "credit_map": [ + 0.018303442746400833, + 0.03039981983602047, + 0.026202920824289322, + 0.022824222221970558, + 0.026735549792647362, + 0.019825685769319534, + 0.029362104833126068, + 0.021120494231581688, + 0.004283589776605368, + 0.007907712832093239, + 0.004829137586057186, + 0.0057500144466757774, + 0.005803431384265423, + 0.003634036984294653, + 0.006581265013664961, + 0.004783160984516144, + 0.011888655833899975, + 0.01503473799675703, + 0.012442251667380333, + 0.014338239096105099, + 0.019788162782788277, + 0.03244760259985924, + 0.023900974541902542, + 0.020266659557819366, + 0.004429992754012346, + 0.00642471294850111, + 0.0040125600062310696, + 0.0051527321338653564, + 0.007199128624051809, + 0.012607746757566929, + 0.005372933577746153, + 0.014511617831885815, + 0.00865749642252922, + 0.010540608316659927, + 0.015205726958811283, + 0.009662661701440811, + 0.007144962903112173, + 0.010969537310302258, + 0.010341593064367771, + 0.011421380564570427, + 0.005994622129946947, + 0.008877129293978214, + 0.0072343237698078156, + 0.007563374470919371, + 0.005618614610284567, + 0.012201924808323383, + 0.008117681369185448, + 0.009365989826619625, + 0.004648065660148859, + 0.0059393225237727165, + 0.008169303648173809, + 0.005543810781091452, + 0.0076122889295220375, + 0.005573135334998369, + 0.005483035463839769, + 0.006372228264808655, + 0.007600520737469196, + 0.007798611186444759, + 0.011515376158058643, + 0.00669516809284687, + 0.011318132281303406, + 0.0380888506770134, + 0.008370824158191681, + 0.03729245439171791, + 0.013389639556407928, + 0.013996216468513012, + 0.02946368418633938, + 0.014722968451678753, + 0.008597814477980137, + 0.012745666317641735, + 0.015455089509487152, + 0.012497179210186005, + 0.008215677924454212, + 0.006789666134864092, + 0.015440518036484718, + 0.005958527326583862, + 0.011445912532508373, + 0.03073536790907383, + 0.011432128958404064, + 0.014009867794811726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.012545241974294186 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978199765086174 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.00978199765086174, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978199765086174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2437896877527237, + 0.07668174058198929, + 0.10731089860200882, + 0.1735125184059143, + 0.08007953315973282, + 0.028869647532701492, + 0.05097107216715813, + 0.07441803812980652, + 0.11140697449445724, + 0.052959855645895004 + ], + "credit_map": [ + 0.02341413125395775, + 0.04215690866112709, + 0.02522970736026764, + 0.02662396989762783, + 0.04473301023244858, + 0.02596883848309517, + 0.035626642405986786, + 0.020036490634083748, + 0.0049941870383918285, + 0.009199567139148712, + 0.012531726621091366, + 0.01128909457474947, + 0.006410261616110802, + 0.0065752374939620495, + 0.015168245881795883, + 0.01051341649144888, + 0.007647530175745487, + 0.00822378695011139, + 0.00945017859339714, + 0.007933035492897034, + 0.013836773112416267, + 0.025716042146086693, + 0.014758383855223656, + 0.01974516548216343, + 0.008782843127846718, + 0.006630412768572569, + 0.005954687017947435, + 0.009787841700017452, + 0.014646979980170727, + 0.04881101846694946, + 0.01338344905525446, + 0.06551527976989746, + 0.007454398088157177, + 0.007816607132554054, + 0.014324748888611794, + 0.008479666896164417, + 0.008792273700237274, + 0.01065501943230629, + 0.011300701647996902, + 0.01125611737370491, + 0.0022494501899927855, + 0.003612091764807701, + 0.0030160045716911554, + 0.0037389735225588083, + 0.003464351175352931, + 0.005055942572653294, + 0.0043035331182181835, + 0.003429299220442772, + 0.004831009078770876, + 0.006561075337231159, + 0.007015910465270281, + 0.005931010935455561, + 0.009916882961988449, + 0.005371840205043554, + 0.0053940871730446815, + 0.005949256010353565, + 0.004958901088684797, + 0.004257567226886749, + 0.009966087527573109, + 0.003963679540902376, + 0.005272555164992809, + 0.021753987297415733, + 0.004581204149872065, + 0.01966405287384987, + 0.008611888624727726, + 0.005105048418045044, + 0.03662034869194031, + 0.00493128877133131, + 0.004701205529272556, + 0.010869914665818214, + 0.00658699544146657, + 0.03398028016090393, + 0.00413900800049305, + 0.0026243822649121284, + 0.00807116087526083, + 0.0034942931961268187, + 0.005361550487577915, + 0.01500802394002676, + 0.004738317336887121, + 0.0095231207087636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.00978199765086174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790569096803665 + ], + "opposite_logprob_tokens": [ + -3.5297906398773193 + ], + "logprob_reward": -0.029790569096803665, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790569096803665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22765842080116272, + 0.035022228956222534, + 0.068199023604393, + 0.07305024564266205, + 0.06843404471874237, + 0.031699493527412415, + 0.05448716878890991, + 0.1607150137424469, + 0.1732272207736969, + 0.1075071468949318 + ], + "credit_map": [ + 0.02055761218070984, + 0.036525238305330276, + 0.02232547290623188, + 0.02375359646975994, + 0.04057106375694275, + 0.023827042430639267, + 0.03605625778436661, + 0.024042122066020966, + 0.0028388313949108124, + 0.004332669544965029, + 0.003233373397961259, + 0.004711688496172428, + 0.004135826602578163, + 0.00430360808968544, + 0.00579463317990303, + 0.005671597085893154, + 0.006009398028254509, + 0.0057706343941390514, + 0.007326567079871893, + 0.005563956219702959, + 0.007510658819228411, + 0.013685825280845165, + 0.01117688324302435, + 0.011155092157423496, + 0.007021146826446056, + 0.006803615018725395, + 0.005310654174536467, + 0.007108435034751892, + 0.017438769340515137, + 0.009537866339087486, + 0.009665003046393394, + 0.010164749808609486, + 0.006399018224328756, + 0.005024009849876165, + 0.013656942173838615, + 0.006291778292506933, + 0.007923867553472519, + 0.011065714992582798, + 0.010301350615918636, + 0.0077713606879115105, + 0.0025086221285164356, + 0.004118680022656918, + 0.005028094630688429, + 0.003547141794115305, + 0.0034573934972286224, + 0.005601933691650629, + 0.0037084848154336214, + 0.0037291382905095816, + 0.006074419245123863, + 0.00741849048063159, + 0.009330452419817448, + 0.005814855452626944, + 0.008557511493563652, + 0.005826801061630249, + 0.006296808365732431, + 0.005167821422219276, + 0.009596407413482666, + 0.006992901209741831, + 0.017167285084724426, + 0.008345137350261211, + 0.011122596450150013, + 0.05099684000015259, + 0.010628266260027885, + 0.04586556553840637, + 0.014601517468690872, + 0.0066008977591991425, + 0.06983691453933716, + 0.007436491549015045, + 0.00596577487885952, + 0.026919754222035408, + 0.01105364691466093, + 0.03081221505999565, + 0.00893881730735302, + 0.0050349547527730465, + 0.01710495725274086, + 0.005981207359582186, + 0.011162815615534782, + 0.03199002519249916, + 0.009151308797299862, + 0.018143048509955406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.029790569096803665 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007621141150593758 + ], + "opposite_logprob_tokens": [ + -4.8826212882995605 + ], + "logprob_reward": -0.007621141150593758, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007621141150593758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18966315686702728, + 0.037293050438165665, + 0.07540019601583481, + 0.05990525335073471, + 0.06658007949590683, + 0.09676588326692581, + 0.02662854827940464, + 0.28387102484703064, + 0.12013031542301178, + 0.04376249760389328 + ], + "credit_map": [ + 0.014085771515965462, + 0.026669371873140335, + 0.017978504300117493, + 0.017200544476509094, + 0.04333047196269035, + 0.024516252800822258, + 0.029908519238233566, + 0.01597372628748417, + 0.00344830472022295, + 0.005214419681578875, + 0.003327649552375078, + 0.005213563330471516, + 0.005325488746166229, + 0.004855360370129347, + 0.005429356824606657, + 0.004478909075260162, + 0.005144710186868906, + 0.005498312413692474, + 0.006259093526750803, + 0.006619890220463276, + 0.008618420921266079, + 0.015704162418842316, + 0.01713666506111622, + 0.010418935678899288, + 0.0035807311069220304, + 0.003300092415884137, + 0.0032560862600803375, + 0.004660321399569511, + 0.011332650668919086, + 0.012501876801252365, + 0.005658287089318037, + 0.015615208074450493, + 0.0072967116720974445, + 0.00525412242859602, + 0.013510867021977901, + 0.0062098754569888115, + 0.007631431799381971, + 0.010517696850001812, + 0.006320866756141186, + 0.009838512167334557, + 0.007257115561515093, + 0.008365687914192677, + 0.010647542774677277, + 0.007508067414164543, + 0.01398453488945961, + 0.02341378480195999, + 0.01091521605849266, + 0.014673935249447823, + 0.002437329851090908, + 0.0033752871677279472, + 0.004523843992501497, + 0.003004995407536626, + 0.003662334056571126, + 0.0031373559031635523, + 0.003103930037468672, + 0.0033834718633443117, + 0.020302901044487953, + 0.0072341542690992355, + 0.01961715705692768, + 0.012788733467459679, + 0.015187814831733704, + 0.10095805674791336, + 0.016567349433898926, + 0.0912148654460907, + 0.011937180534005165, + 0.006219889968633652, + 0.0346541628241539, + 0.005778215825557709, + 0.005949326790869236, + 0.030454447492957115, + 0.010188119485974312, + 0.01494897436350584, + 0.0034567886032164097, + 0.0027063682209700346, + 0.007616837974637747, + 0.0026012789458036423, + 0.004194909706711769, + 0.012941445223987103, + 0.004355599172413349, + 0.005889275576919317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.007621141150593758 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera, in a smooth, 360-degree panning shot, slowly pans around a man in a courtyard suspended above a sea of clouds. In the center of the courtyard, this handsome man with an ethereal presence, his waist-length jet-black hair gently brushes his chiseled features. He wears a rich, smooth silk robe of a deep midnight blue. The wide cuffs and lapels are meticulously embroidered with intricate patterns of stars and flowing clouds in silver thread, shimmering with a subtle luster in the dim light. With focused, careful hands, he holds a delicate, crescent-shaped silver vessel. Steadily and deliberately, he tilts it, and a dazzling stream of light, composed of countless shimmering particles, like a liquid galaxy, slowly flows out, pouring down upon a unique flower growing in a white marble pot before him. The flower is composed entirely of transparent crystal, its petals perfectly faceted. As the starlight touches it, the entire blossom instantly illuminates from within, emitting a rainbow-like, shifting halo. As the camera rotates, the background scene unfolds: the courtyard is bordered by an intricately carved white jade railing, its texture as gentle as jade. Organically entwined with exotic vines emitting a soft blue glow, their leaves seemingly composed of solidified light. Beyond the railing lies an endless, swirling sea of clouds, dyed a lavender by celestial light. The deep night sky is dotted with distant nebulae and stars, creating an atmosphere of tranquility, sacredness, and a sense of wonder.", + "seed": 31786, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_7/index_6678_seed_31786_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2145763972774148, + "vlm_logprob": -0.2145763972774148, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.2145763972774148, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9863691329956055 + ], + "opposite_logprob_tokens": [ + -0.48636913299560547 + ], + "logprob_reward": -0.9863691329956055, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9863691329956055, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17609648406505585, + 0.07699259370565414, + 0.1447826772928238, + 0.055657219141721725, + 0.0936676636338234, + 0.05142281949520111, + 0.08875265717506409, + 0.1455140858888626, + 0.0846814513206482, + 0.08243226259946823 + ], + "credit_map": [ + 0.024219946935772896, + 0.011210785247385502, + 0.02372438833117485, + 0.026548195630311966, + 0.012729397974908352, + 0.03955012187361717, + 0.014465458691120148, + 0.02364819310605526, + 0.011650098487734795, + 0.004347176291048527, + 0.008757554925978184, + 0.01576773077249527, + 0.004116471856832504, + 0.014418133534491062, + 0.005844872910529375, + 0.012090550735592842, + 0.016657754778862, + 0.016612354665994644, + 0.03003290481865406, + 0.016662059351801872, + 0.013071049004793167, + 0.017945194616913795, + 0.01522708311676979, + 0.01857428066432476, + 0.009417271241545677, + 0.0052343737334012985, + 0.010892105288803577, + 0.006437846925109625, + 0.004138740710914135, + 0.005611013155430555, + 0.006087652873247862, + 0.007838217541575432, + 0.01322187576442957, + 0.0064295027405023575, + 0.012345514260232449, + 0.013464710675179958, + 0.006500633433461189, + 0.012057532556355, + 0.016438625752925873, + 0.013209269382059574, + 0.007708652410656214, + 0.005991167388856411, + 0.007508821785449982, + 0.005328724626451731, + 0.004223284311592579, + 0.005786393769085407, + 0.00882051233202219, + 0.006055263336747885, + 0.007073019631206989, + 0.008756374940276146, + 0.012997743673622608, + 0.009763374924659729, + 0.009285143576562405, + 0.007771576289087534, + 0.021970627829432487, + 0.011134800501167774, + 0.014545855112373829, + 0.01863192766904831, + 0.02686196193099022, + 0.01349654607474804, + 0.012634293176233768, + 0.020994484424591064, + 0.01988227106630802, + 0.018466755747795105, + 0.008169031701982021, + 0.01257130317389965, + 0.013003023341298103, + 0.010268223471939564, + 0.00836433470249176, + 0.01191522553563118, + 0.011402291245758533, + 0.008988016285002232, + 0.007791571319103241, + 0.013239661231637001, + 0.010037568397819996, + 0.009602449834346771, + 0.011372866109013557, + 0.009916839189827442, + 0.00946098007261753, + 0.011010331101715565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.9863691329956055 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981058694422245 + ], + "opposite_logprob_tokens": [ + -3.529810667037964 + ], + "logprob_reward": -0.02981058694422245, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981058694422245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18733449280261993, + 0.15535277128219604, + 0.0893799439072609, + 0.0783601701259613, + 0.05686832219362259, + 0.11024237424135208, + 0.060632843524217606, + 0.07722870260477066, + 0.13092999160289764, + 0.05367037653923035 + ], + "credit_map": [ + 0.019495779648423195, + 0.01933209039270878, + 0.023971719667315483, + 0.0373866967856884, + 0.022835012525320053, + 0.01847885362803936, + 0.02115659974515438, + 0.02467774972319603, + 0.027105994522571564, + 0.011229660362005234, + 0.016353821381926537, + 0.01731053739786148, + 0.00995771773159504, + 0.03634902834892273, + 0.01375804003328085, + 0.023287970572710037, + 0.010649289935827255, + 0.00872739963233471, + 0.014378379099071026, + 0.010786211118102074, + 0.007281729951500893, + 0.017384039238095284, + 0.008832880295813084, + 0.011340010911226273, + 0.008386160247027874, + 0.004359513986855745, + 0.005237022880464792, + 0.006467460189014673, + 0.002595100086182356, + 0.022930480539798737, + 0.00895900372415781, + 0.019425423815846443, + 0.006836700718849897, + 0.006061465013772249, + 0.010439740493893623, + 0.0055662184022367, + 0.004253799095749855, + 0.006756665650755167, + 0.007043600082397461, + 0.009910134598612785, + 0.009966692887246609, + 0.016748661175370216, + 0.012293204665184021, + 0.015723323449492455, + 0.009025109000504017, + 0.010583690367639065, + 0.022810155525803566, + 0.013091528788208961, + 0.00657020416110754, + 0.009243208914995193, + 0.010570974089205265, + 0.00726888095960021, + 0.006771740969270468, + 0.0049968017265200615, + 0.008553712628781796, + 0.006657317280769348, + 0.007188520859926939, + 0.009402903728187084, + 0.012909656390547752, + 0.008238629437983036, + 0.006403403822332621, + 0.014523347839713097, + 0.008987491019070148, + 0.009574748575687408, + 0.010347984731197357, + 0.017291249707341194, + 0.02248489297926426, + 0.010853869840502739, + 0.012539496645331383, + 0.019065404310822487, + 0.023623032495379448, + 0.014724062755703926, + 0.00514182960614562, + 0.009222017601132393, + 0.008134614676237106, + 0.005787933710962534, + 0.006270988378673792, + 0.007165294140577316, + 0.005657265894114971, + 0.00629042973741889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.02981058694422245 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978530291467905 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.00978530291467905, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978530291467905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16695848107337952, + 0.11738387495279312, + 0.06140255928039551, + 0.1974164843559265, + 0.06233733147382736, + 0.10768651962280273, + 0.04662921652197838, + 0.046538688242435455, + 0.12427756190299988, + 0.06936925649642944 + ], + "credit_map": [ + 0.02547859400510788, + 0.015621221624314785, + 0.015509692020714283, + 0.029673058539628983, + 0.01645231619477272, + 0.027963688597083092, + 0.014603249728679657, + 0.021656667813658714, + 0.019333453848958015, + 0.008114313706755638, + 0.013265644200146198, + 0.018876170739531517, + 0.00723724951967597, + 0.02349449321627617, + 0.009583648294210434, + 0.017478909343481064, + 0.007486139424145222, + 0.005535481031984091, + 0.012767001986503601, + 0.007802191190421581, + 0.004559798631817102, + 0.00826024729758501, + 0.006872430443763733, + 0.008119277656078339, + 0.01954149268567562, + 0.010857204906642437, + 0.013056757859885693, + 0.01402703020721674, + 0.0075128464959561825, + 0.06356295198202133, + 0.01882350817322731, + 0.05003470554947853, + 0.007120536174625158, + 0.005438983906060457, + 0.007657721173018217, + 0.008328689262270927, + 0.005482544656842947, + 0.007376362103968859, + 0.011890328489243984, + 0.009042168967425823, + 0.011597979813814163, + 0.01174717117100954, + 0.01099167950451374, + 0.01348178181797266, + 0.010848484933376312, + 0.009684913791716099, + 0.027325093746185303, + 0.012009425088763237, + 0.004448376130312681, + 0.004105293191969395, + 0.006892318371683359, + 0.0038387677632272243, + 0.00353417475707829, + 0.006468617357313633, + 0.011558689177036285, + 0.0057829818688333035, + 0.004815832711756229, + 0.004264011047780514, + 0.008743837475776672, + 0.003396065440028906, + 0.0037054908461868763, + 0.010436049662530422, + 0.0047637661918997765, + 0.006413637660443783, + 0.010449264198541641, + 0.016302118077874184, + 0.021629037335515022, + 0.009550684131681919, + 0.012339519336819649, + 0.01786871813237667, + 0.022208692505955696, + 0.01392954122275114, + 0.007647762540727854, + 0.007203410845249891, + 0.012046371586620808, + 0.004781932570040226, + 0.008738995529711246, + 0.010539221577346325, + 0.008915039710700512, + 0.00949652586132288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.00978530291467905 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026342246681451797 + ], + "opposite_logprob_tokens": [ + -3.6513421535491943 + ], + "logprob_reward": -0.026342246681451797, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342246681451797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09575746208429337, + 0.08873115479946136, + 0.0450289249420166, + 0.10201098769903183, + 0.06755475699901581, + 0.08449900150299072, + 0.1692742109298706, + 0.05595996230840683, + 0.1032259464263916, + 0.18795756995677948 + ], + "credit_map": [ + 0.014703157357871532, + 0.007846883498132229, + 0.008783209137618542, + 0.019990330561995506, + 0.008059809915721416, + 0.015229839831590652, + 0.008820678107440472, + 0.01232355460524559, + 0.012809916399419308, + 0.006409475579857826, + 0.009436149150133133, + 0.011679122224450111, + 0.0063678002916276455, + 0.018192943185567856, + 0.007884879596531391, + 0.015950877219438553, + 0.005798653233796358, + 0.004647060297429562, + 0.009252685122191906, + 0.005334076937288046, + 0.0034227969590574503, + 0.005511941388249397, + 0.0055269645527005196, + 0.005534750409424305, + 0.010656769387423992, + 0.003876349190250039, + 0.008429605513811111, + 0.0062060789205133915, + 0.004562776070088148, + 0.03383328393101692, + 0.009078732691705227, + 0.02536739781498909, + 0.009639614261686802, + 0.00466942461207509, + 0.010229481384158134, + 0.007186305243521929, + 0.004341063089668751, + 0.011288823559880257, + 0.009848098270595074, + 0.01035195030272007, + 0.007698824163526297, + 0.007970215752720833, + 0.013853233307600021, + 0.009163053706288338, + 0.0076530734077095985, + 0.00979637261480093, + 0.018154755234718323, + 0.010209484957158566, + 0.015510112047195435, + 0.012507908046245575, + 0.021578248590230942, + 0.015810998156666756, + 0.014112976379692554, + 0.020277559757232666, + 0.04918739199638367, + 0.02028903365135193, + 0.0050195506773889065, + 0.004263701848685741, + 0.010003523901104927, + 0.0036516760010272264, + 0.005202587693929672, + 0.013982417993247509, + 0.005708177573978901, + 0.008128331042826176, + 0.01154372002929449, + 0.009383384138345718, + 0.013340801000595093, + 0.006034106947481632, + 0.013504049740731716, + 0.025899194180965424, + 0.01017305999994278, + 0.013347647152841091, + 0.018952015787363052, + 0.014742055907845497, + 0.026016047224402428, + 0.01121753640472889, + 0.02636684477329254, + 0.032519418746232986, + 0.031211290508508682, + 0.026932384818792343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.026342246681451797 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020574716851115227 + ], + "opposite_logprob_tokens": [ + -3.8955748081207275 + ], + "logprob_reward": -0.020574716851115227, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020574716851115227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22838252782821655, + 0.06417430192232132, + 0.043006837368011475, + 0.07088392972946167, + 0.15032035112380981, + 0.029413102194666862, + 0.16828884184360504, + 0.11246521025896072, + 0.06440676003694534, + 0.06865815073251724 + ], + "credit_map": [ + 0.041600845754146576, + 0.013862214051187038, + 0.03021869994699955, + 0.019792255014181137, + 0.017408840358257294, + 0.05657783895730972, + 0.018032288178801537, + 0.030889542773365974, + 0.010399001650512218, + 0.004119653254747391, + 0.00705371517688036, + 0.009585116058588028, + 0.003869707230478525, + 0.01297009363770485, + 0.005865706596523523, + 0.010311308316886425, + 0.004707894288003445, + 0.004183951765298843, + 0.0060973139479756355, + 0.005402522627264261, + 0.003722407855093479, + 0.007519951090216637, + 0.005404267460107803, + 0.005968528799712658, + 0.009204257279634476, + 0.005539024248719215, + 0.011058925651013851, + 0.00562707707285881, + 0.0038426737301051617, + 0.015406076796352863, + 0.007121967617422342, + 0.013083932921290398, + 0.020561063662171364, + 0.010956758633255959, + 0.020413653925061226, + 0.010474027134478092, + 0.010973364114761353, + 0.03524589166045189, + 0.014266995713114738, + 0.02742859162390232, + 0.0031101207714527845, + 0.0033119567669928074, + 0.0033812420442700386, + 0.0028507383540272713, + 0.002782449359074235, + 0.003575278678908944, + 0.006884017027914524, + 0.0035172998905181885, + 0.021104861050844193, + 0.01753099076449871, + 0.02910376526415348, + 0.014252365566790104, + 0.016212033107876778, + 0.017354553565382957, + 0.0321958027780056, + 0.020534470677375793, + 0.009714843705296516, + 0.008979625068604946, + 0.020016873255372047, + 0.006230033468455076, + 0.009955092333257198, + 0.02863530069589615, + 0.012086397036910057, + 0.016847038641572, + 0.006487414240837097, + 0.00565734226256609, + 0.006233432795852423, + 0.006196419708430767, + 0.005288669839501381, + 0.015345760621130466, + 0.009507943876087666, + 0.00968977902084589, + 0.00725384010002017, + 0.0068927668035030365, + 0.009779560379683971, + 0.005778337363153696, + 0.00831209123134613, + 0.012850663624703884, + 0.008913361467421055, + 0.008877536281943321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.020574716851115227 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera, in a smooth, 360-degree panning shot, slowly pans around a man in a courtyard suspended above a sea of clouds. In the center of the courtyard, this handsome man with an ethereal presence, his waist-length jet-black hair gently brushes his chiseled features. He wears a rich, smooth silk robe of a deep midnight blue. The wide cuffs and lapels are meticulously embroidered with intricate patterns of stars and flowing clouds in silver thread, shimmering with a subtle luster in the dim light. With focused, careful hands, he holds a delicate, crescent-shaped silver vessel. Steadily and deliberately, he tilts it, and a dazzling stream of light, composed of countless shimmering particles, like a liquid galaxy, slowly flows out, pouring down upon a unique flower growing in a white marble pot before him. The flower is composed entirely of transparent crystal, its petals perfectly faceted. As the starlight touches it, the entire blossom instantly illuminates from within, emitting a rainbow-like, shifting halo. As the camera rotates, the background scene unfolds: the courtyard is bordered by an intricately carved white jade railing, its texture as gentle as jade. Organically entwined with exotic vines emitting a soft blue glow, their leaves seemingly composed of solidified light. Beyond the railing lies an endless, swirling sea of clouds, dyed a lavender by celestial light. The deep night sky is dotted with distant nebulae and stars, creating an atmosphere of tranquility, sacredness, and a sense of wonder.", + "seed": 31787, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_7/index_6678_seed_31787_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.21914470810443162, + "vlm_logprob": -0.21914470810443162, + "vlm_yes_no_margin": 2.8500000059604647, + "vlm_token_credit": -0.21914470810443162, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9831266403198242 + ], + "opposite_logprob_tokens": [ + -0.4831266701221466 + ], + "logprob_reward": -0.9831266403198242, + "yes_no_margin_reward": -0.4999999701976776, + "token_credit_reward": -0.9831266403198242, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18642812967300415, + 0.05314404144883156, + 0.057127878069877625, + 0.18492667376995087, + 0.026727011427283287, + 0.0873221755027771, + 0.02794039621949196, + 0.1827303171157837, + 0.15537244081497192, + 0.03828094154596329 + ], + "credit_map": [ + 0.0267038531601429, + 0.016824962571263313, + 0.025027461349964142, + 0.014360569417476654, + 0.028739022091031075, + 0.03281858563423157, + 0.02014290541410446, + 0.021810783073306084, + 0.006057278718799353, + 0.010657917708158493, + 0.004894308280199766, + 0.006660389248281717, + 0.007040489464998245, + 0.007918098010122776, + 0.005180449225008488, + 0.004735119640827179, + 0.00796618964523077, + 0.004173515364527702, + 0.005863019730895758, + 0.0042386348359286785, + 0.005925944074988365, + 0.018513932824134827, + 0.004108395893126726, + 0.006338246166706085, + 0.027970340102910995, + 0.010525630787014961, + 0.01830076053738594, + 0.008825878612697124, + 0.008880539797246456, + 0.0627979189157486, + 0.03306233137845993, + 0.014563278295099735, + 0.003129594260826707, + 0.002628037706017494, + 0.003686619224026799, + 0.0022625550627708435, + 0.0029021010268479586, + 0.0031422264873981476, + 0.0030816064681857824, + 0.005894274450838566, + 0.02685556933283806, + 0.006701374892145395, + 0.014567967504262924, + 0.007482395973056555, + 0.006077914964407682, + 0.008596091531217098, + 0.004594213794916868, + 0.01244665402919054, + 0.0031298219691962004, + 0.0025018395390361547, + 0.003449496580287814, + 0.003278463613241911, + 0.0037993607111275196, + 0.004117102827876806, + 0.0024877949617803097, + 0.005176517181098461, + 0.03126595541834831, + 0.02336304262280464, + 0.02420162409543991, + 0.02017565816640854, + 0.031205078586935997, + 0.02007036842405796, + 0.01483440026640892, + 0.01761418581008911, + 0.010770970955491066, + 0.05439819023013115, + 0.012016773223876953, + 0.05148204043507576, + 0.008417543023824692, + 0.005868060979992151, + 0.006838764529675245, + 0.0055801100097596645, + 0.004563392139971256, + 0.005603884346783161, + 0.006449534557759762, + 0.006517559755593538, + 0.004133164416998625, + 0.003915288019925356, + 0.0035853195004165173, + 0.0035128025338053703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.9831266403198242 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033700522035360336 + ], + "opposite_logprob_tokens": [ + -3.408700466156006 + ], + "logprob_reward": -0.033700522035360336, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033700522035360336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1427694708108902, + 0.01830601505935192, + 0.06742949038743973, + 0.49467554688453674, + 0.02590608410537243, + 0.05061660334467888, + 0.021588709205389023, + 0.018401682376861572, + 0.12531191110610962, + 0.03499453887343407 + ], + "credit_map": [ + 0.030465280637145042, + 0.0102342888712883, + 0.015044275671243668, + 0.011051264591515064, + 0.01206994242966175, + 0.03656814247369766, + 0.014767071232199669, + 0.012569187209010124, + 0.0022039986215531826, + 0.0031691209878772497, + 0.0024347142316401005, + 0.002033891621977091, + 0.001971033401787281, + 0.0017929773312062025, + 0.0027746004052460194, + 0.0019256770610809326, + 0.012653442099690437, + 0.005617825780063868, + 0.004551246762275696, + 0.003332900581881404, + 0.008071979507803917, + 0.025392945855855942, + 0.002780635142698884, + 0.0050285132601857185, + 0.07275756448507309, + 0.027070172131061554, + 0.05126316845417023, + 0.02982269786298275, + 0.0240013524889946, + 0.16392186284065247, + 0.09032171964645386, + 0.035516951233148575, + 0.002737157279625535, + 0.0027289572171866894, + 0.0038385563530027866, + 0.0028833632823079824, + 0.003483002306893468, + 0.0026728971861302853, + 0.0023134557995945215, + 0.005248694214969873, + 0.005790919065475464, + 0.0050034355372190475, + 0.009121405892074108, + 0.005680419504642487, + 0.00649664644151926, + 0.005414743907749653, + 0.004852028097957373, + 0.00825700256973505, + 0.0021869014017283916, + 0.002643751911818981, + 0.0036178689915686846, + 0.003477728459984064, + 0.002812930615618825, + 0.0021713627502322197, + 0.001622148440219462, + 0.00305601442232728, + 0.0027741335798054934, + 0.0033843459095805883, + 0.0025601957459002733, + 0.002321528270840645, + 0.0028321228455752134, + 0.001644673990085721, + 0.0013076337054371834, + 0.001577045419253409, + 0.008582687936723232, + 0.04667120426893234, + 0.008785412646830082, + 0.04285810515284538, + 0.006190383341163397, + 0.0036626835353672504, + 0.0053354972042143345, + 0.0032259291037917137, + 0.004550485871732235, + 0.0038833965081721544, + 0.007843281142413616, + 0.005795994773507118, + 0.003460734151303768, + 0.0031020529568195343, + 0.003784500528126955, + 0.002574089216068387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.033700522035360336 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067076474428177 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.016067076474428177, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067076474428177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2538606822490692, + 0.033013664186000824, + 0.08665314316749573, + 0.1354946494102478, + 0.05330699682235718, + 0.13529177010059357, + 0.10298747569322586, + 0.0960695818066597, + 0.04453130438923836, + 0.058790791779756546 + ], + "credit_map": [ + 0.027953309938311577, + 0.024164794012904167, + 0.0365787036716938, + 0.023300355300307274, + 0.02152031846344471, + 0.02810853347182274, + 0.0374620258808136, + 0.05477263405919075, + 0.0033548620995134115, + 0.00647599995136261, + 0.0041810027323663235, + 0.004016299732029438, + 0.0033598204609006643, + 0.003952209837734699, + 0.0034745833836495876, + 0.004198884125798941, + 0.010842340998351574, + 0.013146133162081242, + 0.010010105557739735, + 0.0082250339910388, + 0.007641088217496872, + 0.022407058626413345, + 0.006750525441020727, + 0.007630856242030859, + 0.019796162843704224, + 0.008121510967612267, + 0.014508619904518127, + 0.00940689817070961, + 0.0070898025296628475, + 0.04419085755944252, + 0.02238452434539795, + 0.009996279142796993, + 0.005497659556567669, + 0.004859020933508873, + 0.007072755601257086, + 0.005023486912250519, + 0.006393055431544781, + 0.005138141568750143, + 0.005221948493272066, + 0.014100925996899605, + 0.03852567821741104, + 0.007553372066468, + 0.018569491803646088, + 0.0055073886178433895, + 0.006558695808053017, + 0.010129280388355255, + 0.004413621034473181, + 0.044034238904714584, + 0.011826383881270885, + 0.008449823595583439, + 0.019570639356970787, + 0.014317266643047333, + 0.01120900921523571, + 0.01213366724550724, + 0.01019345223903656, + 0.015287238173186779, + 0.01616816408932209, + 0.014118745923042297, + 0.012938224710524082, + 0.011703170835971832, + 0.015431204810738564, + 0.00996336992830038, + 0.006815316621214151, + 0.008931380696594715, + 0.007220297586172819, + 0.004934505093842745, + 0.010620519518852234, + 0.00616665231063962, + 0.0036995327100157738, + 0.003263358026742935, + 0.003640519455075264, + 0.004985919687896967, + 0.006967878434807062, + 0.00650860695168376, + 0.012699647806584835, + 0.008971631526947021, + 0.006440356373786926, + 0.004966189619153738, + 0.00760563462972641, + 0.004630846902728081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.016067076474428177 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048643991351127625 + ], + "opposite_logprob_tokens": [ + -3.0486440658569336 + ], + "logprob_reward": -0.048643991351127625, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048643991351127625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19142472743988037, + 0.02066509984433651, + 0.07262951135635376, + 0.36917534470558167, + 0.04353737831115723, + 0.06010329723358154, + 0.06866701692342758, + 0.0653526559472084, + 0.08375956118106842, + 0.024685407057404518 + ], + "credit_map": [ + 0.04405613988637924, + 0.011574411764740944, + 0.015087611973285675, + 0.009315690957009792, + 0.024689311161637306, + 0.06031090393662453, + 0.013384213671088219, + 0.013006444089114666, + 0.002522155875340104, + 0.0025071660056710243, + 0.002510250546038151, + 0.0014994790544733405, + 0.0027379035018384457, + 0.00359825580380857, + 0.0022836970165371895, + 0.0030061928555369377, + 0.009228806011378765, + 0.009723251685500145, + 0.009677033871412277, + 0.005682940594851971, + 0.00638546934351325, + 0.020183632150292397, + 0.004966089501976967, + 0.006782297044992447, + 0.057444363832473755, + 0.02014388144016266, + 0.03861444070935249, + 0.020394880324602127, + 0.01799529232084751, + 0.12524722516536713, + 0.06313490122556686, + 0.026200365275144577, + 0.00393273402005434, + 0.0032695375848561525, + 0.00681972224265337, + 0.004162346012890339, + 0.004845034796744585, + 0.005346976686269045, + 0.004015384707599878, + 0.011145645752549171, + 0.019145723432302475, + 0.0039856573566794395, + 0.007379570044577122, + 0.002667587948963046, + 0.0027098562568426132, + 0.004825285170227289, + 0.001888594008050859, + 0.01750102825462818, + 0.006658022291958332, + 0.008420681580901146, + 0.015235567465424538, + 0.0085308151319623, + 0.007252454292029142, + 0.007336079142987728, + 0.006252622231841087, + 0.008980774320662022, + 0.008530889637768269, + 0.008386099711060524, + 0.009043346159160137, + 0.007605755235999823, + 0.012447579763829708, + 0.00725980382412672, + 0.005224096123129129, + 0.006855081766843796, + 0.00793601293116808, + 0.02571910247206688, + 0.008737741969525814, + 0.02418956160545349, + 0.004854824393987656, + 0.003694993443787098, + 0.004258778411895037, + 0.00436854874715209, + 0.0035080357920378447, + 0.002459742361679673, + 0.006441755220293999, + 0.004334266763180494, + 0.002489744918420911, + 0.002233424223959446, + 0.0017450520535930991, + 0.0014733867719769478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.048643991351127625 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree panning shot around the central subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a man with waist-length jet-black hair wearing a deep midnight blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a delicate crescent-shaped silver vessel from which a stream of shimmering particles flows out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a flower made of transparent crystal in a white marble pot that illuminates with a rainbow-like halo when touched by the light stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background environment a courtyard with white jade railings, glowing blue vines, and a lavender sea of clouds under a starry night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418531034141779 + ], + "opposite_logprob_tokens": [ + -4.264185428619385 + ], + "logprob_reward": -0.01418531034141779, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418531034141779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16272708773612976, + 0.12931300699710846, + 0.05359560251235962, + 0.054918982088565826, + 0.04277293384075165, + 0.0410708524286747, + 0.061923038214445114, + 0.09055634588003159, + 0.31438860297203064, + 0.04873352870345116 + ], + "credit_map": [ + 0.028739385306835175, + 0.012708044610917568, + 0.021246066316962242, + 0.011569536291062832, + 0.01804937794804573, + 0.028448542580008507, + 0.019544348120689392, + 0.022421784698963165, + 0.015558578073978424, + 0.021631816402077675, + 0.013691084459424019, + 0.009755022823810577, + 0.016676368191838264, + 0.01830601692199707, + 0.014773176982998848, + 0.018920935690402985, + 0.0075589995831251144, + 0.006001645233482122, + 0.009110691025853157, + 0.004128753207623959, + 0.004388522822409868, + 0.013372025452554226, + 0.003801944199949503, + 0.0052330223843455315, + 0.00466348510235548, + 0.004352498333901167, + 0.005271431989967823, + 0.005644922610372305, + 0.002819361165165901, + 0.014540988020598888, + 0.012193755246698856, + 0.005432538688182831, + 0.003964703995734453, + 0.0036208732053637505, + 0.006068170070648193, + 0.00424363324418664, + 0.004443597048521042, + 0.005241791717708111, + 0.004170279949903488, + 0.011019882746040821, + 0.0031763766892254353, + 0.0022893876302987337, + 0.009447162970900536, + 0.004467961844056845, + 0.0034590992145240307, + 0.004354854580014944, + 0.0023690147791057825, + 0.011506995186209679, + 0.0063316491432487965, + 0.006260757800191641, + 0.011516258120536804, + 0.008757988922297955, + 0.007138882298022509, + 0.008368142880499363, + 0.004401945509016514, + 0.009147414937615395, + 0.011016030795872211, + 0.011982908472418785, + 0.01511920802295208, + 0.010815665125846863, + 0.015467362478375435, + 0.0100651104003191, + 0.007346091326326132, + 0.008743972517549992, + 0.02281692624092102, + 0.11753582209348679, + 0.023006679490208626, + 0.10539980232715607, + 0.015031606890261173, + 0.008663715794682503, + 0.012428643181920052, + 0.009505406953394413, + 0.007653558626770973, + 0.00710351811721921, + 0.00943418126553297, + 0.008446584455668926, + 0.0047714305110275745, + 0.004297861363738775, + 0.0038420355413109064, + 0.003184356028214097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.01418531034141779 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_8/index_tensor([7843, 7843])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_8/index_tensor([7843, 7843])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5c212db94a477fb547b795fcb7a3b3407f36d52f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_8/index_tensor([7843, 7843])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic slow-motion shot begins from an upward perspective, then slowly pulls back to reveal a grand scene. Atop a precipitous peak, a knight stands braving the fierce wind. He wears a wide-sleeved, dark cyan silk and linen blend Hanfu (Chinese traditional clothing). The wide sleeves and layered hem are lifted high in the fierce mountain breeze, like wings. His inky black hair dances wildly back in the wind, a few strands brushing against his chiseled, resolute features. His feet are firmly planted on the weathered gray-black rock, his body leaning back slightly, his brow furrowed as he roars heavenward with all his might. The veins in his neck are clearly visible from the effort. As the camera pulls out, a swirling sea of clouds appears beneath him. Thick clouds drift slowly across the mountain range. The sky is a gloomy, leaden gray, with only a hint of pale light piercing the distant horizon, casting a cold, even, diffuse light over the entire scene. The main colors of the entire picture are cold blue, gray and black, full of epic loneliness and indomitable heroism, which fully expresses the ultimate tension and majestic momentum in martial arts aesthetics.", + "seed": 21861, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_8/index_7843_seed_21861_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11601075865328311, + "vlm_logprob": -0.11601075865328311, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.11601075865328311, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031616777181625366 + ], + "opposite_logprob_tokens": [ + -3.906616687774658 + ], + "logprob_reward": -0.031616777181625366, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.031616777181625366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14496304094791412, + 0.08444485068321228, + 0.07921266555786133, + 0.2716253399848938, + 0.09027106314897537, + 0.05348413437604904, + 0.05441348999738693, + 0.08177334815263748, + 0.07918597757816315, + 0.06062610447406769 + ], + "credit_map": [ + 0.01602502353489399, + 0.01664418913424015, + 0.01599535346031189, + 0.022151922807097435, + 0.010314268991351128, + 0.015184564515948296, + 0.027595125138759613, + 0.021052580326795578, + 0.007480778731405735, + 0.014300454407930374, + 0.0060152048245072365, + 0.013902876526117325, + 0.007240727078169584, + 0.008581607602536678, + 0.015218421816825867, + 0.011704771779477596, + 0.008374352939426899, + 0.008431293070316315, + 0.005636467598378658, + 0.0053353626281023026, + 0.005268426146358252, + 0.011961161158978939, + 0.01290514413267374, + 0.02130044996738434, + 0.03197048231959343, + 0.024185143411159515, + 0.020814435556530952, + 0.010424938052892685, + 0.02022651396691799, + 0.04411910101771355, + 0.022735655307769775, + 0.0971490740776062, + 0.011477954685688019, + 0.010031426325440407, + 0.011652681045234203, + 0.008535162545740604, + 0.0066298493184149265, + 0.011071532964706421, + 0.009100615046918392, + 0.021771833300590515, + 0.008282494731247425, + 0.006638243328779936, + 0.01116066612303257, + 0.0031528430990874767, + 0.004072132054716349, + 0.00536041334271431, + 0.004686345811933279, + 0.010130995884537697, + 0.006501208059489727, + 0.005704852752387524, + 0.010713698342442513, + 0.004931669216603041, + 0.0041371481493115425, + 0.005921655800193548, + 0.00403881398960948, + 0.012464439496397972, + 0.013293083757162094, + 0.009320490062236786, + 0.01037912629544735, + 0.006222244817763567, + 0.007184108253568411, + 0.011814410798251629, + 0.005408701486885548, + 0.018151182681322098, + 0.011509235017001629, + 0.005567400716245174, + 0.00889610219746828, + 0.006662705913186073, + 0.006623786874115467, + 0.011865808628499508, + 0.005473485682159662, + 0.022587444633245468, + 0.005276838783174753, + 0.004749289248138666, + 0.004849283955991268, + 0.005109038669615984, + 0.011300970800220966, + 0.011966360732913017, + 0.00425544660538435, + 0.013118871487677097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.031616777181625366 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06210116669535637 + ], + "opposite_logprob_tokens": [ + -2.812101125717163 + ], + "logprob_reward": -0.06210116669535637, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06210116669535637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18750068545341492, + 0.2341587394475937, + 0.15754187107086182, + 0.06296750903129578, + 0.08013167977333069, + 0.044192392379045486, + 0.046184372156858444, + 0.11952001601457596, + 0.0394962914288044, + 0.028306446969509125 + ], + "credit_map": [ + 0.02223876304924488, + 0.020075825974345207, + 0.0187454205006361, + 0.025070343166589737, + 0.013112207874655724, + 0.02132302150130272, + 0.03387630358338356, + 0.03305879235267639, + 0.02574995346367359, + 0.032542720437049866, + 0.014533725567162037, + 0.02824890986084938, + 0.022970153018832207, + 0.02593209035694599, + 0.05723243206739426, + 0.026948753744363785, + 0.01576117053627968, + 0.022218722850084305, + 0.015984151512384415, + 0.018192119896411896, + 0.015155211091041565, + 0.02046450972557068, + 0.020065225660800934, + 0.029700767248868942, + 0.008842334151268005, + 0.0073308940045535564, + 0.00686841132119298, + 0.0055948770605027676, + 0.004357625264674425, + 0.008446602150797844, + 0.004542436916381121, + 0.016984334215521812, + 0.008956032805144787, + 0.01663718931376934, + 0.008652429096400738, + 0.01387836504727602, + 0.007261104416102171, + 0.007663834374397993, + 0.005943248979747295, + 0.01113947480916977, + 0.006207896396517754, + 0.005921978037804365, + 0.008695258758962154, + 0.003302572760730982, + 0.0035578852985054255, + 0.005105676129460335, + 0.0034021197352558374, + 0.007999003864824772, + 0.005227585788816214, + 0.0067061432637274265, + 0.006739094387739897, + 0.0052624852396547794, + 0.0034291830379515886, + 0.005468233488500118, + 0.0034925222862511873, + 0.009859124198555946, + 0.015306499786674976, + 0.01310337521135807, + 0.008049866184592247, + 0.0090916333720088, + 0.007636959198862314, + 0.017606519162654877, + 0.006532520055770874, + 0.04219263419508934, + 0.0054020024836063385, + 0.004255670588463545, + 0.005379846319556236, + 0.0043250215239822865, + 0.0037021622993052006, + 0.005573983304202557, + 0.002480386523529887, + 0.008377214893698692, + 0.003598875366151333, + 0.0022482983767986298, + 0.003991348668932915, + 0.003183809807524085, + 0.005009503569453955, + 0.004248994868248701, + 0.0016921747010201216, + 0.004333443474024534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.06210116669535637 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553011998534203 + ], + "opposite_logprob_tokens": [ + -4.3875532150268555 + ], + "logprob_reward": -0.012553011998534203, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553011998534203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17836135625839233, + 0.15005521476268768, + 0.10539010912179947, + 0.11105313897132874, + 0.11740428954362869, + 0.05462860316038132, + 0.0639045387506485, + 0.12818710505962372, + 0.04384659230709076, + 0.04716898500919342 + ], + "credit_map": [ + 0.01892484724521637, + 0.021393798291683197, + 0.02206745184957981, + 0.02843482606112957, + 0.013990755192935467, + 0.018754545599222183, + 0.028874671086668968, + 0.025920450687408447, + 0.016374621540308, + 0.012524977326393127, + 0.01177874393761158, + 0.01047902088612318, + 0.011867382563650608, + 0.018719498068094254, + 0.03348923474550247, + 0.03482174128293991, + 0.011420166119933128, + 0.01207171194255352, + 0.01095523964613676, + 0.009055291302502155, + 0.005739087238907814, + 0.014617519453167915, + 0.0156889446079731, + 0.025842154398560524, + 0.01210922934114933, + 0.020738985389471054, + 0.012646229937672615, + 0.012720993719995022, + 0.012550057843327522, + 0.012749827466905117, + 0.010585052892565727, + 0.016952764242887497, + 0.014445023611187935, + 0.013308131136000156, + 0.016416655853390694, + 0.009526997804641724, + 0.009737364947795868, + 0.015408560633659363, + 0.010313055478036404, + 0.028248505666851997, + 0.006779666058719158, + 0.008102305233478546, + 0.009744294919073582, + 0.00416179234161973, + 0.004609068855643272, + 0.005799081642180681, + 0.0047073266468942165, + 0.010725065134465694, + 0.007725253701210022, + 0.007986949756741524, + 0.010256077162921429, + 0.009069778025150299, + 0.003937308210879564, + 0.006469253450632095, + 0.005539126694202423, + 0.012920793145895004, + 0.01570495404303074, + 0.008198202587664127, + 0.01566760614514351, + 0.006671820301562548, + 0.008969934657216072, + 0.02017928846180439, + 0.007820993661880493, + 0.044974301010370255, + 0.005731537938117981, + 0.0040750750340521336, + 0.005884924903512001, + 0.006985747255384922, + 0.003967040218412876, + 0.005063407123088837, + 0.0033310989383608103, + 0.008807756938040257, + 0.006713738199323416, + 0.0042469697073102, + 0.0059082224033772945, + 0.004964393097907305, + 0.005964905023574829, + 0.006304214708507061, + 0.004026147536933422, + 0.00904039666056633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.012553011998534203 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036048531532288 + ], + "opposite_logprob_tokens": [ + -1.9103604555130005 + ], + "logprob_reward": -0.16036048531532288, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036048531532288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20431499183177948, + 0.11044076085090637, + 0.11870294064283371, + 0.17082779109477997, + 0.10055681318044662, + 0.035551466047763824, + 0.05572327971458435, + 0.08129985630512238, + 0.058478958904743195, + 0.0641031339764595 + ], + "credit_map": [ + 0.028185304254293442, + 0.02077038213610649, + 0.026434291154146194, + 0.023502223193645477, + 0.017976241186261177, + 0.024114292114973068, + 0.03917952999472618, + 0.02415272779762745, + 0.009303013794124126, + 0.01595464162528515, + 0.006747795734554529, + 0.01297430694103241, + 0.008951499126851559, + 0.01250908151268959, + 0.021458284929394722, + 0.022542141377925873, + 0.011206935159862041, + 0.013891460373997688, + 0.015037965029478073, + 0.01335749588906765, + 0.009840399026870728, + 0.01648075506091118, + 0.01574929989874363, + 0.02313862554728985, + 0.023799629881978035, + 0.025565944612026215, + 0.02168651856482029, + 0.019070299342274666, + 0.012620684690773487, + 0.019123610109090805, + 0.013667798601090908, + 0.035293299704790115, + 0.01164847519248724, + 0.015286660753190517, + 0.01637054793536663, + 0.009636708535254002, + 0.007930351421236992, + 0.010605602525174618, + 0.008115156553685665, + 0.020963303744792938, + 0.00650718342512846, + 0.003840136108919978, + 0.0045830486342310905, + 0.0020713075064122677, + 0.0029389099217951298, + 0.004921909887343645, + 0.0029766091611236334, + 0.007712364196777344, + 0.007182097528129816, + 0.00496937520802021, + 0.013357710093259811, + 0.005638851318508387, + 0.0033538357820361853, + 0.0059449016116559505, + 0.004318046849220991, + 0.010958464816212654, + 0.01160344947129488, + 0.00769791379570961, + 0.015832001343369484, + 0.004768699873238802, + 0.004959995858371258, + 0.009483635425567627, + 0.00584242632612586, + 0.021111732348799706, + 0.006805839482694864, + 0.004662546794861555, + 0.007006408646702766, + 0.007863419130444527, + 0.005757605656981468, + 0.008963847532868385, + 0.004073226358741522, + 0.013346067629754543, + 0.007416608277708292, + 0.00450373999774456, + 0.008220620453357697, + 0.004381793085485697, + 0.00767284631729126, + 0.010831386782228947, + 0.004637048579752445, + 0.016439087688922882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.16036048531532288 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134223520755768 + ], + "opposite_logprob_tokens": [ + -1.3134223222732544 + ], + "logprob_reward": -0.3134223520755768, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134223520755768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22996142506599426, + 0.1066015213727951, + 0.0925857275724411, + 0.09190479665994644, + 0.09019199758768082, + 0.05373680964112282, + 0.15483741462230682, + 0.08801933377981186, + 0.040750619024038315, + 0.05141032487154007 + ], + "credit_map": [ + 0.03260229527950287, + 0.028892716392874718, + 0.018547944724559784, + 0.03305673971772194, + 0.021668460220098495, + 0.02811768837273121, + 0.04148392006754875, + 0.025591645389795303, + 0.009491955861449242, + 0.015000897459685802, + 0.009741660207509995, + 0.011236260645091534, + 0.010634398087859154, + 0.012997976504266262, + 0.023283587768673897, + 0.014214780181646347, + 0.014171362854540348, + 0.016755452379584312, + 0.007116542663425207, + 0.010655011981725693, + 0.004889581352472305, + 0.011992646381258965, + 0.009549041278660297, + 0.017456086352467537, + 0.010907895863056183, + 0.01510652806609869, + 0.011668930761516094, + 0.0072906240820884705, + 0.009883702732622623, + 0.011924064718186855, + 0.0061705405823886395, + 0.018952500075101852, + 0.012021019123494625, + 0.021996913477778435, + 0.00822377484291792, + 0.01114092767238617, + 0.005408169236034155, + 0.009884331375360489, + 0.005902077071368694, + 0.015614781528711319, + 0.008445123210549355, + 0.009606442414224148, + 0.006119438447058201, + 0.004235355649143457, + 0.005120286718010902, + 0.0063705407083034515, + 0.003991521429270506, + 0.009848096407949924, + 0.020763030275702477, + 0.020372919738292694, + 0.015919899567961693, + 0.020653463900089264, + 0.010906598530709743, + 0.019483700394630432, + 0.013500389643013477, + 0.03323740139603615, + 0.012776151299476624, + 0.007181759458035231, + 0.013008617796003819, + 0.0044113583862781525, + 0.006298426538705826, + 0.011119351722300053, + 0.005488911177963018, + 0.027734756469726562, + 0.004506254568696022, + 0.004151433706283569, + 0.003875794354826212, + 0.004574146121740341, + 0.005087714176625013, + 0.005766686052083969, + 0.0032247682102024555, + 0.009563818573951721, + 0.005604845471680164, + 0.005112717859447002, + 0.005450351629406214, + 0.006001595873385668, + 0.007235497236251831, + 0.008200532756745815, + 0.004774460569024086, + 0.009030320681631565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.3134223520755768 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic slow-motion shot begins from an upward perspective, then slowly pulls back to reveal a grand scene. Atop a precipitous peak, a knight stands braving the fierce wind. He wears a wide-sleeved, dark cyan silk and linen blend Hanfu (Chinese traditional clothing). The wide sleeves and layered hem are lifted high in the fierce mountain breeze, like wings. His inky black hair dances wildly back in the wind, a few strands brushing against his chiseled, resolute features. His feet are firmly planted on the weathered gray-black rock, his body leaning back slightly, his brow furrowed as he roars heavenward with all his might. The veins in his neck are clearly visible from the effort. As the camera pulls out, a swirling sea of clouds appears beneath him. Thick clouds drift slowly across the mountain range. The sky is a gloomy, leaden gray, with only a hint of pale light piercing the distant horizon, casting a cold, even, diffuse light over the entire scene. The main colors of the entire picture are cold blue, gray and black, full of epic loneliness and indomitable heroism, which fully expresses the ultimate tension and majestic momentum in martial arts aesthetics.", + "seed": 21862, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_8/index_7843_seed_21862_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12399293836206197, + "vlm_logprob": -0.12399293836206197, + "vlm_yes_no_margin": 3.3999999940395353, + "vlm_token_credit": -0.12399293836206197, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009573856368660927 + ], + "opposite_logprob_tokens": [ + -5.634573936462402 + ], + "logprob_reward": -0.009573856368660927, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.009573856368660927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10215539485216141, + 0.09351486712694168, + 0.042908571660518646, + 0.02585277520120144, + 0.05484011396765709, + 0.022926274687051773, + 0.027836717665195465, + 0.5642547607421875, + 0.037828098982572556, + 0.027882367372512817 + ], + "credit_map": [ + 0.013417677022516727, + 0.013779525645077229, + 0.014483412727713585, + 0.016167767345905304, + 0.013571722432971, + 0.0145000871270895, + 0.007320066913962364, + 0.008915144950151443, + 0.010012693703174591, + 0.012714250013232231, + 0.02629886008799076, + 0.015586832538247108, + 0.007297576870769262, + 0.009170826524496078, + 0.005511065013706684, + 0.0069227684289216995, + 0.0038488181307911873, + 0.005715692415833473, + 0.007891187444329262, + 0.007004137150943279, + 0.008732160553336143, + 0.004106794483959675, + 0.0027062557637691498, + 0.0029035324696451426, + 0.0033078889828175306, + 0.0031473487615585327, + 0.004524777177721262, + 0.0039100549183785915, + 0.003288077423349023, + 0.0034439917653799057, + 0.0019077055621892214, + 0.002322934800758958, + 0.006184150464832783, + 0.005910737905651331, + 0.008370192721486092, + 0.010053458623588085, + 0.00616421177983284, + 0.008292371407151222, + 0.005021747201681137, + 0.004843247123062611, + 0.002622205065563321, + 0.0025259980466216803, + 0.004516844637691975, + 0.003980958368629217, + 0.0032694165129214525, + 0.0023903220426291227, + 0.0017969765467569232, + 0.001823555794544518, + 0.004773746244609356, + 0.0035552370827645063, + 0.003949108067899942, + 0.004136162344366312, + 0.0035611458588391542, + 0.003043654840439558, + 0.0027198991738259792, + 0.0020977684762328863, + 0.026815129444003105, + 0.22863903641700745, + 0.01038863230496645, + 0.011112010106444359, + 0.020208247005939484, + 0.24013909697532654, + 0.01208129245787859, + 0.014871419407427311, + 0.003689373377710581, + 0.005005048122256994, + 0.005543222185224295, + 0.007418064866214991, + 0.004297167062759399, + 0.0045833345502614975, + 0.003941450268030167, + 0.0033504455350339413, + 0.0029239614959806204, + 0.0035164218861609697, + 0.004305652808398008, + 0.004733254201710224, + 0.0029113951604813337, + 0.0037830849178135395, + 0.0032300218008458614, + 0.0024785781279206276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.009573856368660927 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029838357120752335 + ], + "opposite_logprob_tokens": [ + -3.5298383235931396 + ], + "logprob_reward": -0.029838357120752335, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029838357120752335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07446270436048508, + 0.06560087203979492, + 0.03799019008874893, + 0.017166471108794212, + 0.028673844411969185, + 0.022092150524258614, + 0.029956819489598274, + 0.6772373914718628, + 0.009951348416507244, + 0.03686828166246414 + ], + "credit_map": [ + 0.007457176689058542, + 0.008472826331853867, + 0.011488847434520721, + 0.00931592658162117, + 0.012626223266124725, + 0.008875780738890171, + 0.008085723035037518, + 0.008140203543007374, + 0.0054431878961622715, + 0.005940536968410015, + 0.008788641542196274, + 0.007205198518931866, + 0.01779855601489544, + 0.007675029803067446, + 0.006327132694423199, + 0.006422589533030987, + 0.0035595018416643143, + 0.004179935436695814, + 0.0039958590641617775, + 0.004749040585011244, + 0.006968610920011997, + 0.006611854303628206, + 0.003306333441287279, + 0.004619055427610874, + 0.0018818887183442712, + 0.0019298580009490252, + 0.003537387354299426, + 0.002853631740435958, + 0.0027937896084040403, + 0.0017308202804997563, + 0.001220804639160633, + 0.0012182920472696424, + 0.0036649401299655437, + 0.003849053056910634, + 0.0033991660457104445, + 0.0047108205035328865, + 0.004840364679694176, + 0.004250362981110811, + 0.001974779646843672, + 0.0019843587651848793, + 0.002417624229565263, + 0.0023919143714010715, + 0.004226610995829105, + 0.0036449143663048744, + 0.0026201813016086817, + 0.002106711035594344, + 0.0023045833222568035, + 0.0023796120658516884, + 0.004115342628210783, + 0.0038916512858122587, + 0.004191735293716192, + 0.005711665842682123, + 0.00316935358569026, + 0.0026303818449378014, + 0.003397672902792692, + 0.0028490140102803707, + 0.0343765951693058, + 0.27244871854782104, + 0.016562385484576225, + 0.01605578325688839, + 0.022753937169909477, + 0.28386104106903076, + 0.013755409978330135, + 0.017423508688807487, + 0.0010734735988080502, + 0.0015434096567332745, + 0.0011831166921183467, + 0.001372105092741549, + 0.0010995652992278337, + 0.0013942591613158584, + 0.001375960186123848, + 0.0009094576234929264, + 0.0028539637569338083, + 0.007271074689924717, + 0.003471182892099023, + 0.00486078392714262, + 0.0059178732335567474, + 0.006080924533307552, + 0.003526104148477316, + 0.002886372385546565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.029838357120752335 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958531983196735 + ], + "opposite_logprob_tokens": [ + -5.130958557128906 + ], + "logprob_reward": -0.005958531983196735, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958531983196735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12174838781356812, + 0.12100228667259216, + 0.12124239653348923, + 0.10721389204263687, + 0.06037221476435661, + 0.025800200179219246, + 0.036275241523981094, + 0.341511994600296, + 0.027339866384863853, + 0.037493545562028885 + ], + "credit_map": [ + 0.01102592796087265, + 0.017064984887838364, + 0.021313825622200966, + 0.019381504505872726, + 0.016582876443862915, + 0.01499106828123331, + 0.009990843012928963, + 0.01139734871685505, + 0.012314006686210632, + 0.01673448644578457, + 0.025128453969955444, + 0.019459770992398262, + 0.013440786860883236, + 0.010720144025981426, + 0.006658113095909357, + 0.01654651388525963, + 0.011505475267767906, + 0.014487244188785553, + 0.01677180454134941, + 0.017931990325450897, + 0.024088289588689804, + 0.01645101234316826, + 0.01044087577611208, + 0.009565691463649273, + 0.011669869534671307, + 0.013848059810698032, + 0.021211884915828705, + 0.016434915363788605, + 0.014887887984514236, + 0.009579997509717941, + 0.008831446059048176, + 0.010749825276434422, + 0.007667733822017908, + 0.008470200002193451, + 0.0051969606429338455, + 0.006476053036749363, + 0.009777205064892769, + 0.015291833318769932, + 0.0032240787986665964, + 0.004268147051334381, + 0.003364618867635727, + 0.0035854980815201998, + 0.0048037124797701836, + 0.0036188706289976835, + 0.003657995257526636, + 0.0028059910982847214, + 0.0018290936714038253, + 0.0021344185806810856, + 0.0051973732188344, + 0.006036018021404743, + 0.003950313199311495, + 0.0042359190993011, + 0.005351687781512737, + 0.0032711224630475044, + 0.004746965132653713, + 0.0034858430735766888, + 0.01812872476875782, + 0.12920096516609192, + 0.012713388539850712, + 0.010550202801823616, + 0.014220775105059147, + 0.13463875651359558, + 0.011688741855323315, + 0.010370430536568165, + 0.0024852892383933067, + 0.0029035136103630066, + 0.00419792952015996, + 0.002965181600302458, + 0.0038836034946143627, + 0.0036609473172575235, + 0.004185782745480537, + 0.003057616064324975, + 0.00345442071557045, + 0.00452785799279809, + 0.005366956815123558, + 0.0056214528158307076, + 0.004356995224952698, + 0.005137775093317032, + 0.004569250624626875, + 0.0044588325545191765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.005958531983196735 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003112941980362 + ], + "opposite_logprob_tokens": [ + -2.350311279296875 + ], + "logprob_reward": -0.1003112941980362, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003112941980362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21394360065460205, + 0.16956567764282227, + 0.11083745956420898, + 0.052099261432886124, + 0.0841670110821724, + 0.04484062269330025, + 0.05430493503808975, + 0.20034323632717133, + 0.03654951602220535, + 0.033348727971315384 + ], + "credit_map": [ + 0.025710629299283028, + 0.03315846621990204, + 0.040966540575027466, + 0.03050798736512661, + 0.02794678322970867, + 0.02767583727836609, + 0.013926326297223568, + 0.014051031321287155, + 0.01771652325987816, + 0.026529081165790558, + 0.04657651111483574, + 0.03031139262020588, + 0.01587989740073681, + 0.01256253570318222, + 0.006657231133431196, + 0.013332504779100418, + 0.012031369842588902, + 0.012407493777573109, + 0.011162403970956802, + 0.012365982867777348, + 0.023373641073703766, + 0.018785256892442703, + 0.008680772967636585, + 0.012030533514916897, + 0.004485459066927433, + 0.006225598976016045, + 0.009197658859193325, + 0.008169676177203655, + 0.010059359483420849, + 0.004659139085561037, + 0.003912721294909716, + 0.005389646627008915, + 0.009631728753447533, + 0.010557564906775951, + 0.011076068505644798, + 0.012483171187341213, + 0.013775773346424103, + 0.011946381069719791, + 0.00753139378502965, + 0.007164930924773216, + 0.004584754351526499, + 0.004784115124493837, + 0.008753420785069466, + 0.006374507676810026, + 0.005050331819802523, + 0.003658600617200136, + 0.006060454994440079, + 0.005574436392635107, + 0.007860533893108368, + 0.009266671724617481, + 0.004289588890969753, + 0.006831694394350052, + 0.007417046464979649, + 0.005141025874763727, + 0.006482110824435949, + 0.00701626343652606, + 0.014838634990155697, + 0.0483953133225441, + 0.02641584910452366, + 0.021846869960427284, + 0.013502161018550396, + 0.05879911780357361, + 0.008039016276597977, + 0.008506275713443756, + 0.0038031931035220623, + 0.005287724081426859, + 0.004879591520875692, + 0.005103184841573238, + 0.0038671644870191813, + 0.005403474438935518, + 0.004986175801604986, + 0.003219008445739746, + 0.003161671804264188, + 0.005441924091428518, + 0.002571935998275876, + 0.002843794645741582, + 0.007468791212886572, + 0.004866284318268299, + 0.004067583009600639, + 0.0029267417266964912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.1003112941980362 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742826521396637 + ], + "opposite_logprob_tokens": [ + -0.9742826223373413 + ], + "logprob_reward": -0.4742826521396637, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.4742826521396637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1506398767232895, + 0.1082662045955658, + 0.09872587025165558, + 0.12739206850528717, + 0.11949507147073746, + 0.06885753571987152, + 0.05592913553118706, + 0.1618054360151291, + 0.06390014290809631, + 0.044988613575696945 + ], + "credit_map": [ + 0.019466480240225792, + 0.018079401925206184, + 0.02478722110390663, + 0.029219118878245354, + 0.013878031633794308, + 0.015919556841254234, + 0.01223678607493639, + 0.017053280025720596, + 0.012297277338802814, + 0.01442746166139841, + 0.022758888080716133, + 0.015481033362448215, + 0.008543620817363262, + 0.011100798845291138, + 0.00853212270885706, + 0.015124993398785591, + 0.012000382877886295, + 0.012563657015562057, + 0.02062150090932846, + 0.01790674217045307, + 0.010332046076655388, + 0.011892254464328289, + 0.005861685145646334, + 0.007547601591795683, + 0.013497011736035347, + 0.015179687179625034, + 0.022023899480700493, + 0.018540363758802414, + 0.023344550281763077, + 0.012113434262573719, + 0.010951484553515911, + 0.011741643771529198, + 0.01041102223098278, + 0.012817444279789925, + 0.01940317451953888, + 0.0213810782879591, + 0.016129303723573685, + 0.017521193251013756, + 0.010290341451764107, + 0.011541512794792652, + 0.009823467582464218, + 0.009349235333502293, + 0.008674596436321735, + 0.010080539621412754, + 0.011271939612925053, + 0.006562329363077879, + 0.007044109515845776, + 0.006051312666386366, + 0.006803708616644144, + 0.006628850940614939, + 0.007450335193425417, + 0.00988913793116808, + 0.00750018609687686, + 0.005451611243188381, + 0.007020775694400072, + 0.0051845284178853035, + 0.012192594818770885, + 0.05269557237625122, + 0.011601344682276249, + 0.01053237821906805, + 0.010916231200098991, + 0.051338374614715576, + 0.007004713639616966, + 0.005524230655282736, + 0.006662724539637566, + 0.008987272158265114, + 0.010107209905982018, + 0.009908140636980534, + 0.005736034829169512, + 0.01104139443486929, + 0.0058209546841681, + 0.005636409390717745, + 0.004323679953813553, + 0.005923768505454063, + 0.006196269765496254, + 0.00841593835502863, + 0.004871396813541651, + 0.006978914607316256, + 0.004226075019687414, + 0.004052567295730114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.4742826521396637 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic slow-motion shot begins from an upward perspective, then slowly pulls back to reveal a grand scene. Atop a precipitous peak, a knight stands braving the fierce wind. He wears a wide-sleeved, dark cyan silk and linen blend Hanfu (Chinese traditional clothing). The wide sleeves and layered hem are lifted high in the fierce mountain breeze, like wings. His inky black hair dances wildly back in the wind, a few strands brushing against his chiseled, resolute features. His feet are firmly planted on the weathered gray-black rock, his body leaning back slightly, his brow furrowed as he roars heavenward with all his might. The veins in his neck are clearly visible from the effort. As the camera pulls out, a swirling sea of clouds appears beneath him. Thick clouds drift slowly across the mountain range. The sky is a gloomy, leaden gray, with only a hint of pale light piercing the distant horizon, casting a cold, even, diffuse light over the entire scene. The main colors of the entire picture are cold blue, gray and black, full of epic loneliness and indomitable heroism, which fully expresses the ultimate tension and majestic momentum in martial arts aesthetics.", + "seed": 21863, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_8/index_7843_seed_21863_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09423367213457823, + "vlm_logprob": -0.09423367213457823, + "vlm_yes_no_margin": 3.1500000238418577, + "vlm_token_credit": -0.09423367213457823, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014587887562811375 + ], + "opposite_logprob_tokens": [ + -4.764587879180908 + ], + "logprob_reward": -0.014587887562811375, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.014587887562811375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13898223638534546, + 0.10711533576250076, + 0.3558533787727356, + 0.051682740449905396, + 0.04508588835597038, + 0.060097452253103256, + 0.053339883685112, + 0.05453324317932129, + 0.06460423767566681, + 0.06870559602975845 + ], + "credit_map": [ + 0.022130312398076057, + 0.013472184538841248, + 0.025710729882121086, + 0.011726513504981995, + 0.01017697062343359, + 0.019441677257418633, + 0.015847517177462578, + 0.0204763300716877, + 0.011694584973156452, + 0.0063579268753528595, + 0.03153245523571968, + 0.008594765327870846, + 0.009239349514245987, + 0.020878257229924202, + 0.01126826275140047, + 0.0075497254729270935, + 0.04642321541905403, + 0.055807095021009445, + 0.06277179718017578, + 0.030062135308980942, + 0.04384667053818703, + 0.04781908541917801, + 0.03556101396679878, + 0.03356235474348068, + 0.006071100477129221, + 0.006634986959397793, + 0.008761203847825527, + 0.003675345564261079, + 0.00927650835365057, + 0.005752179771661758, + 0.004064461216330528, + 0.0074469540268182755, + 0.00982541125267744, + 0.005121413618326187, + 0.005876415874809027, + 0.005104581825435162, + 0.004444146528840065, + 0.005761453881859779, + 0.0038550053723156452, + 0.005097460933029652, + 0.008711311034858227, + 0.0070870802737772465, + 0.009341049939393997, + 0.006260848604142666, + 0.007398671004921198, + 0.0074427491053938866, + 0.005595358554273844, + 0.008260384202003479, + 0.007556016556918621, + 0.00602201372385025, + 0.009858508594334126, + 0.0026160208508372307, + 0.009694160893559456, + 0.006447570864111185, + 0.003617051290348172, + 0.007528539746999741, + 0.006277468986809254, + 0.006011928431689739, + 0.007329918909817934, + 0.004933657590299845, + 0.006961950566619635, + 0.005991689395159483, + 0.005521475337445736, + 0.01150515303015709, + 0.005357867106795311, + 0.00711526395753026, + 0.009539776481688023, + 0.00695364223793149, + 0.012685559689998627, + 0.006067968904972076, + 0.008031410165131092, + 0.008852743543684483, + 0.009033910930156708, + 0.009367121383547783, + 0.008227299898862839, + 0.005745092872530222, + 0.006957539357244968, + 0.006792272441089153, + 0.007203727029263973, + 0.01537863165140152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.014587887562811375 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283692836761475 + ], + "opposite_logprob_tokens": [ + -2.0178370475769043 + ], + "logprob_reward": -0.14283692836761475, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14283692836761475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3747401237487793, + 0.30366840958595276, + 0.11541542410850525, + 0.04030250012874603, + 0.012046522460877895, + 0.018178505823016167, + 0.036412984132766724, + 0.021878039464354515, + 0.04237041622400284, + 0.03498700261116028 + ], + "credit_map": [ + 0.02998906373977661, + 0.013424416072666645, + 0.13366736471652985, + 0.02394561655819416, + 0.018072813749313354, + 0.11055752635002136, + 0.02416650578379631, + 0.02091684378683567, + 0.03773612901568413, + 0.02776198275387287, + 0.07242476940155029, + 0.022932080551981926, + 0.020873814821243286, + 0.08641169965267181, + 0.015873128548264503, + 0.01965482160449028, + 0.03796083852648735, + 0.00795145146548748, + 0.00926285982131958, + 0.00435813469812274, + 0.009676831774413586, + 0.03539106994867325, + 0.00586502393707633, + 0.004949238151311874, + 0.0048608058132231236, + 0.004451194778084755, + 0.006472299806773663, + 0.0031158910132944584, + 0.009229335002601147, + 0.004489975981414318, + 0.003740731393918395, + 0.00394227122887969, + 0.0020966678857803345, + 0.0013694815570488572, + 0.0017970355693250895, + 0.0013361795572564006, + 0.0016075758030638099, + 0.0014867328573018312, + 0.0010475566377863288, + 0.0013052939902991056, + 0.0015711727319285274, + 0.003400708083063364, + 0.0014301748014986515, + 0.001814202987588942, + 0.005327671300619841, + 0.0015905153704807162, + 0.0016276779351755977, + 0.0014163852902129292, + 0.0036098090931773186, + 0.00444188853725791, + 0.0029455232433974743, + 0.003002273151651025, + 0.010826555080711842, + 0.0031762628350406885, + 0.003548694308847189, + 0.004861982073634863, + 0.0023911017924547195, + 0.002668713917955756, + 0.0024958536960184574, + 0.002105063060298562, + 0.004162440076470375, + 0.002512040315195918, + 0.0031467238441109657, + 0.0023961039260029793, + 0.004329558461904526, + 0.005164918955415487, + 0.003863252466544509, + 0.003684242255985737, + 0.012142814695835114, + 0.0037709937896579504, + 0.005030396394431591, + 0.004384244326502085, + 0.0043030004017055035, + 0.004699565004557371, + 0.0038524342235177755, + 0.002710438799113035, + 0.005203323904424906, + 0.0038050685543566942, + 0.003041124902665615, + 0.007372047286480665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.14283692836761475 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744714453816414 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.006744714453816414, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744714453816414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09492257237434387, + 0.4282374680042267, + 0.07545039057731628, + 0.10630816221237183, + 0.03502766788005829, + 0.03794176131486893, + 0.04070008918642998, + 0.08530232310295105, + 0.04132692515850067, + 0.05478266626596451 + ], + "credit_map": [ + 0.008567594923079014, + 0.011053347960114479, + 0.019986318424344063, + 0.010171105153858662, + 0.007734109181910753, + 0.016507690772414207, + 0.009121387265622616, + 0.011781014502048492, + 0.05478954315185547, + 0.04268534108996391, + 0.0976792499423027, + 0.035043444484472275, + 0.027679404243826866, + 0.11984413117170334, + 0.022328319028019905, + 0.028188053518533707, + 0.02212880179286003, + 0.004964686464518309, + 0.00811996590346098, + 0.004591756965965033, + 0.004824520088732243, + 0.021352706477046013, + 0.004447396844625473, + 0.005020558834075928, + 0.015893302857875824, + 0.014469358138740063, + 0.01189780980348587, + 0.008256719447672367, + 0.025576140731573105, + 0.013488879427313805, + 0.009151316247880459, + 0.00757463276386261, + 0.004964932799339294, + 0.003338608890771866, + 0.006323955021798611, + 0.004216312896460295, + 0.004160905722528696, + 0.004407030995935202, + 0.0034028496593236923, + 0.0042130714282393456, + 0.0030600863974541426, + 0.006781908683478832, + 0.0032592422794550657, + 0.003941576462239027, + 0.009454401209950447, + 0.004181882366538048, + 0.004198672249913216, + 0.003063988871872425, + 0.004808301106095314, + 0.006590300239622593, + 0.004214387387037277, + 0.002499106340110302, + 0.009114774875342846, + 0.005099870264530182, + 0.004192001186311245, + 0.004181348718702793, + 0.014248774386942387, + 0.013342206366360188, + 0.014471203088760376, + 0.0070135812275111675, + 0.008224116638302803, + 0.011033479124307632, + 0.006068411748856306, + 0.01090055052191019, + 0.004048280417919159, + 0.004982398822903633, + 0.005707274191081524, + 0.003979741595685482, + 0.009631948545575142, + 0.0037048347294330597, + 0.003578865434974432, + 0.005693583283573389, + 0.009403147734701633, + 0.008839678019285202, + 0.006115583702921867, + 0.005507167428731918, + 0.004397614859044552, + 0.005852942820638418, + 0.006915296893566847, + 0.0077512371353805065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.006744714453816414 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494527146220207 + ], + "opposite_logprob_tokens": [ + -2.929945230484009 + ], + "logprob_reward": -0.05494527146220207, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494527146220207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18384315073490143, + 0.12358907610177994, + 0.20308473706245422, + 0.06866864114999771, + 0.0647306814789772, + 0.0853106826543808, + 0.0631655678153038, + 0.05943155288696289, + 0.05884500965476036, + 0.08933089673519135 + ], + "credit_map": [ + 0.01700599305331707, + 0.01108598243445158, + 0.06083882600069046, + 0.013111291453242302, + 0.00881561916321516, + 0.05042199417948723, + 0.011846763081848621, + 0.010716686956584454, + 0.015303746797144413, + 0.0180231761187315, + 0.021089717745780945, + 0.014175519347190857, + 0.011361959390342236, + 0.025220345705747604, + 0.009356789290904999, + 0.009057827293872833, + 0.06106621399521828, + 0.01785716786980629, + 0.01469525508582592, + 0.0088548818603158, + 0.01962452009320259, + 0.05819087103009224, + 0.013648122549057007, + 0.009147725068032742, + 0.01168855745345354, + 0.008097345940768719, + 0.01022176630795002, + 0.0045262607745826244, + 0.010714799165725708, + 0.01022338680922985, + 0.006160663906484842, + 0.0070358579978346825, + 0.00863261241465807, + 0.005293074995279312, + 0.00836963765323162, + 0.006708207074552774, + 0.013180095702409744, + 0.006277826614677906, + 0.008627591654658318, + 0.007641634903848171, + 0.010036800988018513, + 0.010955682955682278, + 0.013661731034517288, + 0.008529670536518097, + 0.01357546728104353, + 0.009684663265943527, + 0.008099798113107681, + 0.010766874067485332, + 0.011728656478226185, + 0.006638635881245136, + 0.008702720515429974, + 0.005439136177301407, + 0.0066651334054768085, + 0.008421630598604679, + 0.006504375021904707, + 0.009065280668437481, + 0.007495094556361437, + 0.007262876722961664, + 0.009185494855046272, + 0.005216426681727171, + 0.009381875395774841, + 0.005928578320890665, + 0.00697745569050312, + 0.00798374880105257, + 0.007079383824020624, + 0.008061274886131287, + 0.005693253129720688, + 0.005966514348983765, + 0.012965977191925049, + 0.005828964523971081, + 0.005554364062845707, + 0.007695277687162161, + 0.014523312449455261, + 0.01311033871024847, + 0.0095510957762599, + 0.009552965871989727, + 0.0091372299939394, + 0.010525478981435299, + 0.010529843159019947, + 0.012400628067553043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.05494527146220207 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25205355882644653 + ], + "opposite_logprob_tokens": [ + -1.5020534992218018 + ], + "logprob_reward": -0.25205355882644653, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25205355882644653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2540041208267212, + 0.20990805327892303, + 0.06500544399023056, + 0.0678284764289856, + 0.05617658421397209, + 0.03987518325448036, + 0.0632578432559967, + 0.060892727226018906, + 0.15126226842403412, + 0.03178924694657326 + ], + "credit_map": [ + 0.017497461289167404, + 0.011665750294923782, + 0.09124500304460526, + 0.016262806951999664, + 0.012023801915347576, + 0.07393281161785126, + 0.016653506085276604, + 0.014722969383001328, + 0.02371375449001789, + 0.017141899093985558, + 0.05398548021912575, + 0.015048772096633911, + 0.015450780279934406, + 0.0584627203643322, + 0.012066544033586979, + 0.014038094319403172, + 0.022182166576385498, + 0.004088931251317263, + 0.00452965684235096, + 0.0025204969570040703, + 0.003331341315060854, + 0.021444188430905342, + 0.003968758508563042, + 0.0029399008490145206, + 0.009057585150003433, + 0.007184259127825499, + 0.012243923731148243, + 0.005615861155092716, + 0.014568304643034935, + 0.006232027895748615, + 0.006668069865554571, + 0.006258435547351837, + 0.007336780428886414, + 0.005579405929893255, + 0.007121879607439041, + 0.00629629660397768, + 0.01103617437183857, + 0.00594582362100482, + 0.0061447592452168465, + 0.00671546533703804, + 0.009210845455527306, + 0.004304941743612289, + 0.005733356811106205, + 0.0036882052663713694, + 0.004469351377338171, + 0.005719430278986692, + 0.0021352162584662437, + 0.004613836295902729, + 0.007056225091218948, + 0.006373693235218525, + 0.009042400866746902, + 0.004573407582938671, + 0.013270173221826553, + 0.006744209211319685, + 0.006049329414963722, + 0.010148408822715282, + 0.008655437268316746, + 0.00795621145516634, + 0.009803805500268936, + 0.006783378776162863, + 0.008235556073486805, + 0.00642104959115386, + 0.006444129161536694, + 0.0065931580029428005, + 0.029291464015841484, + 0.023177247494459152, + 0.02245253138244152, + 0.012197470292448997, + 0.019358200952410698, + 0.016988234594464302, + 0.009615999646484852, + 0.018181128427386284, + 0.0028414614498615265, + 0.0035552699118852615, + 0.0032675424590706825, + 0.003343813819810748, + 0.005930643528699875, + 0.0029301531612873077, + 0.0037674647755920887, + 0.006152900401502848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.25205355882644653 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic slow-motion shot begins from an upward perspective, then slowly pulls back to reveal a grand scene. Atop a precipitous peak, a knight stands braving the fierce wind. He wears a wide-sleeved, dark cyan silk and linen blend Hanfu (Chinese traditional clothing). The wide sleeves and layered hem are lifted high in the fierce mountain breeze, like wings. His inky black hair dances wildly back in the wind, a few strands brushing against his chiseled, resolute features. His feet are firmly planted on the weathered gray-black rock, his body leaning back slightly, his brow furrowed as he roars heavenward with all his might. The veins in his neck are clearly visible from the effort. As the camera pulls out, a swirling sea of clouds appears beneath him. Thick clouds drift slowly across the mountain range. The sky is a gloomy, leaden gray, with only a hint of pale light piercing the distant horizon, casting a cold, even, diffuse light over the entire scene. The main colors of the entire picture are cold blue, gray and black, full of epic loneliness and indomitable heroism, which fully expresses the ultimate tension and majestic momentum in martial arts aesthetics.", + "seed": 21864, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_8/index_7843_seed_21864_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11919475328177213, + "vlm_logprob": -0.11919475328177213, + "vlm_yes_no_margin": 3.3749999940395354, + "vlm_token_credit": -0.11919475328177213, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012701095081865788 + ], + "opposite_logprob_tokens": [ + -5.137701034545898 + ], + "logprob_reward": -0.012701095081865788, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.012701095081865788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32359111309051514, + 0.2829190194606781, + 0.05562667176127434, + 0.07406153529882431, + 0.058760304003953934, + 0.06570977717638016, + 0.03474905341863632, + 0.0256531722843647, + 0.026452340185642242, + 0.05247708410024643 + ], + "credit_map": [ + 0.044734254479408264, + 0.030708327889442444, + 0.042551472783088684, + 0.02128516137599945, + 0.04567636922001839, + 0.06297110766172409, + 0.028017673641443253, + 0.04764670506119728, + 0.06506748497486115, + 0.014498572796583176, + 0.027736755087971687, + 0.01593867875635624, + 0.017753571271896362, + 0.059152018278837204, + 0.015797697007656097, + 0.0669742301106453, + 0.010467718355357647, + 0.007254485040903091, + 0.007253773510456085, + 0.0045052277855575085, + 0.004185854457318783, + 0.007410909049212933, + 0.0043637859635055065, + 0.010184910148382187, + 0.009266605600714684, + 0.021260712295770645, + 0.0075386036187410355, + 0.005086958874017, + 0.007927208207547665, + 0.006362270098179579, + 0.0052852933295071125, + 0.011333882808685303, + 0.00912490300834179, + 0.007845024578273296, + 0.013174020685255527, + 0.0065619912929832935, + 0.003620991250500083, + 0.005869499407708645, + 0.0063375066965818405, + 0.006226364988833666, + 0.009493115358054638, + 0.009291582740843296, + 0.01193767599761486, + 0.007863316684961319, + 0.007237487006932497, + 0.00631334213539958, + 0.00523754907771945, + 0.008335706777870655, + 0.004619690123945475, + 0.0063515473157167435, + 0.006282667629420757, + 0.0038832537829875946, + 0.003131810575723648, + 0.0029751767870038748, + 0.0034831410739570856, + 0.004021764267235994, + 0.0037209216970950365, + 0.0027297630440443754, + 0.006267985794693232, + 0.002294068457558751, + 0.0019557317718863487, + 0.0029692500829696655, + 0.00219332124106586, + 0.0035221269354224205, + 0.004393051378428936, + 0.0017776599852368236, + 0.00799002218991518, + 0.002299401443451643, + 0.0021023042500019073, + 0.0027216291055083275, + 0.0019063970539718866, + 0.003261870937421918, + 0.006605726201087236, + 0.008149346336722374, + 0.007832823321223259, + 0.004924784880131483, + 0.007134641520678997, + 0.00390441226772964, + 0.00529141491279006, + 0.008633931167423725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.012701095081865788 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373406082391739 + ], + "opposite_logprob_tokens": [ + -3.408734083175659 + ], + "logprob_reward": -0.03373406082391739, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373406082391739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0678817480802536, + 0.4260343015193939, + 0.08781751990318298, + 0.07636944949626923, + 0.18601590394973755, + 0.05956244841217995, + 0.03334086760878563, + 0.016280097886919975, + 0.01714683137834072, + 0.029550788924098015 + ], + "credit_map": [ + 0.009597069583833218, + 0.007228156551718712, + 0.007835788652300835, + 0.0058950879611074924, + 0.011481190100312233, + 0.010949577204883099, + 0.006316601298749447, + 0.008578283712267876, + 0.0988418236374855, + 0.018399136140942574, + 0.041117772459983826, + 0.024482207372784615, + 0.026814011856913567, + 0.09408240765333176, + 0.024936199188232422, + 0.097360759973526, + 0.013223541900515556, + 0.010712861083447933, + 0.010341590270400047, + 0.009988748468458652, + 0.012504161335527897, + 0.010356605984270573, + 0.009466134011745453, + 0.011223876848816872, + 0.00863978173583746, + 0.01648491993546486, + 0.009811880998313427, + 0.006025230977684259, + 0.00774249155074358, + 0.007612248416990042, + 0.008963285945355892, + 0.01108961459249258, + 0.022637035697698593, + 0.028212517499923706, + 0.028377987444400787, + 0.02168464846909046, + 0.02091987431049347, + 0.019572699442505836, + 0.028754401952028275, + 0.01585674099624157, + 0.006474131252616644, + 0.010478210635483265, + 0.009591741487383842, + 0.005211710464209318, + 0.00844371598213911, + 0.005109049845486879, + 0.004829595796763897, + 0.009424291551113129, + 0.0038826086092740297, + 0.007712723221629858, + 0.004586426075547934, + 0.002950906055048108, + 0.00327062769792974, + 0.0024426390882581472, + 0.0038043803069740534, + 0.004690556321293116, + 0.002117454307153821, + 0.0015943312318995595, + 0.0035223732702434063, + 0.0016432426637038589, + 0.0014033023035153747, + 0.0015112151158973575, + 0.001531186979264021, + 0.0029569915495812893, + 0.0020125629380345345, + 0.002193446969613433, + 0.0035023000091314316, + 0.0019761384464800358, + 0.0016866724472492933, + 0.0015396025264635682, + 0.0016411372926086187, + 0.0025949703995138407, + 0.003385906806215644, + 0.004656009376049042, + 0.005408053752034903, + 0.003072562161833048, + 0.003950021229684353, + 0.0026123621501028538, + 0.0032052965834736824, + 0.003260578028857708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.03373406082391739 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260316655039787 + ], + "opposite_logprob_tokens": [ + -5.255260467529297 + ], + "logprob_reward": -0.005260316655039787, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260316655039787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1615230292081833, + 0.09168092161417007, + 0.07454703748226166, + 0.06489598006010056, + 0.07321654260158539, + 0.42549511790275574, + 0.026284009218215942, + 0.017639653757214546, + 0.010374026373028755, + 0.05434370040893555 + ], + "credit_map": [ + 0.0163163710385561, + 0.02157759480178356, + 0.02092639170587063, + 0.021846212446689606, + 0.01828177459537983, + 0.016918867826461792, + 0.026991743594408035, + 0.018664080649614334, + 0.012688354589045048, + 0.007716365158557892, + 0.011327564716339111, + 0.012652655132114887, + 0.00793707836419344, + 0.012440907768905163, + 0.016004784032702446, + 0.010913217440247536, + 0.008477295748889446, + 0.010439879260957241, + 0.009404350072145462, + 0.009323826059699059, + 0.011217940598726273, + 0.007830175571143627, + 0.009942608885467052, + 0.0079109538346529, + 0.008023669943213463, + 0.011355801485478878, + 0.008135874755680561, + 0.005456528160721064, + 0.006337669678032398, + 0.008014420047402382, + 0.007530470844358206, + 0.01004154421389103, + 0.012636804021894932, + 0.008538556285202503, + 0.014481477439403534, + 0.007250990252941847, + 0.007020268123596907, + 0.008184176869690418, + 0.006118122488260269, + 0.008986148983240128, + 0.06512068212032318, + 0.0423581562936306, + 0.07877976447343826, + 0.04472725838422775, + 0.05649949982762337, + 0.03805357590317726, + 0.04010062664747238, + 0.05985555052757263, + 0.0033121146261692047, + 0.004956510849297047, + 0.003568882355466485, + 0.003115699626505375, + 0.002980268793180585, + 0.0025556401815265417, + 0.003238098928704858, + 0.0025567931588739157, + 0.0018605388468131423, + 0.0030561976600438356, + 0.0026001124642789364, + 0.0024340993259102106, + 0.0019914344884455204, + 0.0014643241884186864, + 0.0021136184222996235, + 0.002119329059496522, + 0.0011720771435648203, + 0.001363007235340774, + 0.002284131245687604, + 0.0013475411105901003, + 0.0008720173500478268, + 0.0008559610578231514, + 0.0008401483646593988, + 0.001639143330976367, + 0.006433213595300913, + 0.008275910280644894, + 0.008532131090760231, + 0.005155963823199272, + 0.008005150593817234, + 0.004271307960152626, + 0.005819674115628004, + 0.007850348018109798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.005260316655039787 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001781463623047 + ], + "opposite_logprob_tokens": [ + -2.6950178146362305 + ], + "logprob_reward": -0.07001781463623047, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001781463623047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09521269798278809, + 0.43649712204933167, + 0.05340895801782608, + 0.08726879954338074, + 0.1096629872918129, + 0.1143035888671875, + 0.025611715391278267, + 0.019796079024672508, + 0.010581235401332378, + 0.04765681177377701 + ], + "credit_map": [ + 0.013643325306475163, + 0.015042505227029324, + 0.010914982296526432, + 0.008005008101463318, + 0.013086177408695221, + 0.013498307205736637, + 0.010324954986572266, + 0.010697426274418831, + 0.10251133888959885, + 0.02068852074444294, + 0.03838075324892998, + 0.025581223890185356, + 0.030251963064074516, + 0.10098743438720703, + 0.024951713159680367, + 0.09314414858818054, + 0.00967926625162363, + 0.005204084794968367, + 0.006040159612894058, + 0.006316505838185549, + 0.006287591531872749, + 0.007982411421835423, + 0.004447871819138527, + 0.007451064884662628, + 0.011004063300788403, + 0.02063662000000477, + 0.009793182834982872, + 0.007122931070625782, + 0.008136852644383907, + 0.007621156517416239, + 0.008010296151041985, + 0.014943688176572323, + 0.016624242067337036, + 0.014983654022216797, + 0.01915227621793747, + 0.01093685906380415, + 0.009486022405326366, + 0.013288469053804874, + 0.015451282262802124, + 0.009740175679326057, + 0.01879393868148327, + 0.012670201249420643, + 0.021309301257133484, + 0.01225303951650858, + 0.012481764890253544, + 0.010084401816129684, + 0.012347416952252388, + 0.014363515190780163, + 0.0029369660187512636, + 0.0048329150304198265, + 0.004229369107633829, + 0.002452435437589884, + 0.002220340073108673, + 0.002294753212481737, + 0.002058182843029499, + 0.004586751572787762, + 0.002247690921649337, + 0.0019766592886298895, + 0.00410666735842824, + 0.0023002021480351686, + 0.0015904109459370375, + 0.0015463248128071427, + 0.002048967406153679, + 0.0039791539311409, + 0.001357865403406322, + 0.0009526402573101223, + 0.0015978205483406782, + 0.0013703560689464211, + 0.0008912532939575613, + 0.0009549290407449007, + 0.00104723556432873, + 0.002409134991466999, + 0.005691864062100649, + 0.006594388280063868, + 0.007259597070515156, + 0.004129763692617416, + 0.0062752761878073215, + 0.003820891724899411, + 0.004882269073277712, + 0.009002759121358395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.07001781463623047 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set atop a precipitous peak with a swirling sea of clouds beneath?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a knight wearing a dark cyan silk and linen blend Hanfu with wide sleeves and a layered hem visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the knight's wide sleeves and layered hem lifted high by the wind like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the knight leaning back with his feet firmly planted on weathered gray-black rock while roaring heavenward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin from an upward perspective and slowly pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47426047921180725 + ], + "opposite_logprob_tokens": [ + -0.9742604494094849 + ], + "logprob_reward": -0.47426047921180725, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.47426047921180725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1072239875793457, + 0.08669071644544601, + 0.18296661972999573, + 0.04928173869848251, + 0.09028773009777069, + 0.2881438136100769, + 0.08375166356563568, + 0.06539037823677063, + 0.017915086820721626, + 0.02834824100136757 + ], + "credit_map": [ + 0.013132334686815739, + 0.01813431642949581, + 0.01072404533624649, + 0.011188716627657413, + 0.014536901377141476, + 0.014108524657785892, + 0.011668432503938675, + 0.013730727136135101, + 0.009515674784779549, + 0.007359913084656, + 0.009332193993031979, + 0.007994137704372406, + 0.012663494795560837, + 0.011530742049217224, + 0.012777949683368206, + 0.015516616404056549, + 0.024871796369552612, + 0.024010546505451202, + 0.02116541378200054, + 0.02233457937836647, + 0.02399957925081253, + 0.021173367276787758, + 0.02460034377872944, + 0.02081100270152092, + 0.007515045814216137, + 0.01123281754553318, + 0.006896662525832653, + 0.004069630987942219, + 0.003732227487489581, + 0.005962593015283346, + 0.003958340268582106, + 0.005914422683417797, + 0.01553252525627613, + 0.011473636142909527, + 0.017151860520243645, + 0.007626212202012539, + 0.00711220083758235, + 0.009912582114338875, + 0.011050444096326828, + 0.010428274050354958, + 0.04214906692504883, + 0.03379754722118378, + 0.0512351468205452, + 0.021280726417899132, + 0.04788680374622345, + 0.027158377692103386, + 0.025325585156679153, + 0.03931059315800667, + 0.011633765883743763, + 0.018064215779304504, + 0.013385864906013012, + 0.00854038167744875, + 0.008521132171154022, + 0.007578921504318714, + 0.007343820761889219, + 0.008683574385941029, + 0.007491735275834799, + 0.0106683149933815, + 0.013885186053812504, + 0.00639230664819479, + 0.007435522507876158, + 0.006309058051556349, + 0.006175223272293806, + 0.007033034693449736, + 0.0020866254344582558, + 0.002047915244475007, + 0.004272076301276684, + 0.0020868019200861454, + 0.0016073045553639531, + 0.001449735020287335, + 0.001986759016290307, + 0.002377869328483939, + 0.0037183535750955343, + 0.004033439792692661, + 0.004716965835541487, + 0.002452464308589697, + 0.0035887910053133965, + 0.002285771770402789, + 0.002728409832343459, + 0.0048240479081869125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.47426047921180725 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_9/index_tensor([6053, 6053])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_9/index_tensor([6053, 6053])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c74de5fdfe987bb2601e5dd4e487d2fcd2dc3db6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_9/index_tensor([6053, 6053])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, eye-level close-up, rendered in the vibrant Pixar animation style, focuses on a lovable, pig-like demon. This creature, with smooth, soft pink skin and a plump, rounded body, sits innocently on the surface of a large, heart-shaped leaf. Its face is dominated by large, expressive emerald green eyes, wide and filled with wonder, reflecting the surrounding light with a lustrous sheen. Just above its eyebrows, two glossy, deep purple horns begin to sprout, their dark color creating a soft contrast against its pink complexion. The massive leaf beneath it is a deep green, covered in glistening, spherical dewdrops that catch the light like tiny gemstones. When the breeze blows, the leaves ripple in a slow, rhythmic pattern, causing the demon to sway gently and deliberately back and forth, its body moving in perfect, effortless sync with the natural movements. Its face is a look of pure, captivated wonder, its mouth slightly open. The scene is illuminated by soft, warm, golden hour light that filters through the dense tree canopy, creating a shallow depth of field that blurs the lush green forest background into a beautiful, painterly bokeh of overlapping shapes and colors, creating a serene and magical atmosphere.", + "seed": 14818, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_9/index_6053_seed_14818_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.059614524245262146, + "vlm_logprob": -0.059614524245262146, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.059614524245262146, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00846086721867323 + ], + "opposite_logprob_tokens": [ + -5.008460998535156 + ], + "logprob_reward": -0.00846086721867323, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00846086721867323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26149219274520874, + 0.10777828842401505, + 0.04874854534864426, + 0.03418101370334625, + 0.02894076518714428, + 0.03786053508520126, + 0.061958976089954376, + 0.055560264736413956, + 0.3078133761882782, + 0.0556659996509552 + ], + "credit_map": [ + 0.02065005712211132, + 0.03177840635180473, + 0.03161294758319855, + 0.04087204486131668, + 0.040751148015260696, + 0.03448875620961189, + 0.031978365033864975, + 0.02936047501862049, + 0.01172273512929678, + 0.019496982917189598, + 0.014257206581532955, + 0.013707567937672138, + 0.013584977015852928, + 0.013683084398508072, + 0.009584600105881691, + 0.01174114178866148, + 0.004830158315598965, + 0.005739421118050814, + 0.003108008299022913, + 0.0057380907237529755, + 0.006959777791053057, + 0.006267457734793425, + 0.010302886366844177, + 0.00580274173989892, + 0.004995073191821575, + 0.004430299624800682, + 0.0038867415860295296, + 0.0036916034296154976, + 0.004214558284729719, + 0.003722723573446274, + 0.00513970572501421, + 0.004100307356566191, + 0.0026550281327217817, + 0.0031294580549001694, + 0.0035470949951559305, + 0.003403988666832447, + 0.004304387141019106, + 0.0029448457062244415, + 0.0035021963994950056, + 0.005453766323626041, + 0.00583336828276515, + 0.003394379513338208, + 0.0036321845836937428, + 0.0036928642075508833, + 0.004200773313641548, + 0.0051202536560595036, + 0.004126233048737049, + 0.007860478013753891, + 0.008579817600548267, + 0.00866976659744978, + 0.006586265750229359, + 0.007643420714884996, + 0.008090334013104439, + 0.008719485253095627, + 0.0057309470139443874, + 0.007938941940665245, + 0.006639074068516493, + 0.005893830209970474, + 0.006384522188454866, + 0.006755162961781025, + 0.007765425369143486, + 0.006807453464716673, + 0.005741242319345474, + 0.009573553688824177, + 0.04257047548890114, + 0.0599503368139267, + 0.04243664816021919, + 0.030315298587083817, + 0.039025548845529556, + 0.03906449303030968, + 0.02474725991487503, + 0.029703326523303986, + 0.008539466187357903, + 0.007233159150928259, + 0.0054071214981377125, + 0.007784515619277954, + 0.008925455622375011, + 0.00706056784838438, + 0.004380122292786837, + 0.0063355909660458565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.00846086721867323 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740451790392399 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740451790392399, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740451790392399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2959653437137604, + 0.05454922467470169, + 0.05020240694284439, + 0.05790762975811958, + 0.08484727889299393, + 0.1361696571111679, + 0.09800948202610016, + 0.09805801510810852, + 0.09452672302722931, + 0.02976420149207115 + ], + "credit_map": [ + 0.02086816169321537, + 0.038045622408390045, + 0.020854754373431206, + 0.033981118351221085, + 0.04796748235821724, + 0.0391048863530159, + 0.06310729682445526, + 0.03203603997826576, + 0.005122672766447067, + 0.00686843553557992, + 0.005516901146620512, + 0.004740845877677202, + 0.008019401691854, + 0.0074499910697340965, + 0.008312130346894264, + 0.008518842980265617, + 0.005145934410393238, + 0.004107051994651556, + 0.0061132777482271194, + 0.0037727425806224346, + 0.006197337992489338, + 0.007114696316421032, + 0.005381563678383827, + 0.012369805946946144, + 0.005553866736590862, + 0.006173699628561735, + 0.005901841912418604, + 0.006291945930570364, + 0.006935825105756521, + 0.007258557248860598, + 0.010309156961739063, + 0.009482735767960548, + 0.011465483345091343, + 0.008313938044011593, + 0.01970200054347515, + 0.006170176435261965, + 0.006385957822203636, + 0.008471108973026276, + 0.007399221416562796, + 0.01693938486278057, + 0.026425877586007118, + 0.011550918221473694, + 0.011442759074270725, + 0.012450802139937878, + 0.014672519639134407, + 0.017985941842198372, + 0.01220043282955885, + 0.029440414160490036, + 0.01604893058538437, + 0.00987029355019331, + 0.01083801407366991, + 0.01171445194631815, + 0.007981400936841965, + 0.013489593751728535, + 0.007134457118809223, + 0.020932340994477272, + 0.008497904054820538, + 0.01535151805728674, + 0.011622097343206406, + 0.011095640249550343, + 0.011974303983151913, + 0.013353992253541946, + 0.009820747189223766, + 0.016341811046004295, + 0.012081262655556202, + 0.019961735233664513, + 0.011438488960266113, + 0.009205954149365425, + 0.01287003606557846, + 0.012273598462343216, + 0.008421252481639385, + 0.008274387568235397, + 0.002523731905966997, + 0.003499058075249195, + 0.005782044027000666, + 0.0025245221331715584, + 0.0036971981171518564, + 0.004108343739062548, + 0.0030118131544440985, + 0.004617490340024233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.006740451790392399 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864465907216072 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.00864465907216072, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864465907216072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30695176124572754, + 0.09073331207036972, + 0.0566912405192852, + 0.07335806638002396, + 0.08374330401420593, + 0.0460907481610775, + 0.08422888815402985, + 0.1282699704170227, + 0.09272415935993195, + 0.037208568304777145 + ], + "credit_map": [ + 0.026646330952644348, + 0.04232513904571533, + 0.030858609825372696, + 0.0411221981048584, + 0.05109952762722969, + 0.04069390147924423, + 0.03389978036284447, + 0.04030625522136688, + 0.010813514702022076, + 0.016871199011802673, + 0.008121808990836143, + 0.012647652067244053, + 0.011603196151554585, + 0.011313455179333687, + 0.009918253868818283, + 0.009444225579500198, + 0.006664641201496124, + 0.007718340959399939, + 0.004949990194290876, + 0.007626782171428204, + 0.00850011222064495, + 0.007292440161108971, + 0.006752998102456331, + 0.007185928523540497, + 0.006914783734828234, + 0.012178082019090652, + 0.008457430638372898, + 0.007890894077718258, + 0.008705934509634972, + 0.009579850360751152, + 0.00974776316434145, + 0.009883319959044456, + 0.010189013555645943, + 0.009891572408378124, + 0.016697462648153305, + 0.008354932069778442, + 0.010883067734539509, + 0.007565111853182316, + 0.00862798374146223, + 0.011534148827195168, + 0.006892941892147064, + 0.006868627853691578, + 0.0049675749614834785, + 0.004547589458525181, + 0.005442800000309944, + 0.005230273120105267, + 0.00506488187238574, + 0.0070760599337518215, + 0.007725353352725506, + 0.014323447830975056, + 0.011312640272080898, + 0.010968134738504887, + 0.010385747067630291, + 0.011267018504440784, + 0.005674605257809162, + 0.012571935541927814, + 0.01112367957830429, + 0.019372105598449707, + 0.014793704263865948, + 0.015104487538337708, + 0.018481839448213577, + 0.016103964298963547, + 0.014579381793737411, + 0.01871078833937645, + 0.013628916814923286, + 0.017695637419819832, + 0.010465309955179691, + 0.010973184369504452, + 0.012047972530126572, + 0.01202528364956379, + 0.010010059922933578, + 0.005877786781638861, + 0.004533758386969566, + 0.00457849632948637, + 0.0029587997123599052, + 0.005155263002961874, + 0.005505107343196869, + 0.004830603022128344, + 0.0035588201135396957, + 0.006087714806199074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.00864465907216072 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048668064177036285 + ], + "opposite_logprob_tokens": [ + -3.048668146133423 + ], + "logprob_reward": -0.048668064177036285, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048668064177036285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2270485907793045, + 0.06991279870271683, + 0.030085494741797447, + 0.07346165925264359, + 0.10531522333621979, + 0.08323934674263, + 0.09026389569044113, + 0.24746885895729065, + 0.056361954659223557, + 0.016842136159539223 + ], + "credit_map": [ + 0.02565566822886467, + 0.02896345779299736, + 0.020005779340863228, + 0.022902747616171837, + 0.029408885166049004, + 0.02848002500832081, + 0.03729291260242462, + 0.034339115023612976, + 0.009375710971653461, + 0.010837098583579063, + 0.006960530299693346, + 0.008278879337012768, + 0.008569709025323391, + 0.008363666944205761, + 0.006868906784802675, + 0.010658295825123787, + 0.0033690023701637983, + 0.0033791831228882074, + 0.002656726399436593, + 0.003483342472463846, + 0.0034348417539149523, + 0.0042518507689237595, + 0.004035061225295067, + 0.005475486628711224, + 0.007348652463406324, + 0.010449389927089214, + 0.0074659776873886585, + 0.007594170980155468, + 0.008478337898850441, + 0.009060322307050228, + 0.012240785174071789, + 0.010824024677276611, + 0.012049547396600246, + 0.01707596518099308, + 0.01008238922804594, + 0.011670294217765331, + 0.016142170876264572, + 0.011032347567379475, + 0.011735498905181885, + 0.015527015551924706, + 0.014802508056163788, + 0.009225341491401196, + 0.007245837245136499, + 0.009570986032485962, + 0.009021369740366936, + 0.01062736101448536, + 0.008475374430418015, + 0.014270567335188389, + 0.010459243319928646, + 0.009532724507153034, + 0.013254359364509583, + 0.013421008363366127, + 0.009345147758722305, + 0.011708129197359085, + 0.00703009869903326, + 0.015513189136981964, + 0.023388328030705452, + 0.04052205756306648, + 0.02751733735203743, + 0.030803529545664787, + 0.03355769068002701, + 0.03018731065094471, + 0.025184743106365204, + 0.036307867616415024, + 0.006928037852048874, + 0.011787020601332188, + 0.007660441566258669, + 0.005570508074015379, + 0.0074227722361683846, + 0.007426787167787552, + 0.004628967493772507, + 0.0049374219961464405, + 0.0030509172938764095, + 0.0021991257090121508, + 0.0016984683461487293, + 0.0021899540442973375, + 0.002773019950836897, + 0.002034443663433194, + 0.0015020630089566112, + 0.0013941435609012842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.048668064177036285 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255585789680481 + ], + "opposite_logprob_tokens": [ + -1.6005585193634033 + ], + "logprob_reward": -0.2255585789680481, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255585789680481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10456331819295883, + 0.03384331613779068, + 0.0322638563811779, + 0.034112267196178436, + 0.04304865747690201, + 0.05763762816786766, + 0.02745680697262287, + 0.030869517475366592, + 0.6056253910064697, + 0.030579203739762306 + ], + "credit_map": [ + 0.010630731470882893, + 0.011399014852941036, + 0.008550992235541344, + 0.013791671022772789, + 0.015394979156553745, + 0.013636574149131775, + 0.019565032795071602, + 0.011594321578741074, + 0.004268667660653591, + 0.0038214647211134434, + 0.003994494676589966, + 0.003848274005576968, + 0.004124670289456844, + 0.0046973396092653275, + 0.00491868844255805, + 0.004169716499745846, + 0.0039626783691346645, + 0.0039050618652254343, + 0.0024556077551096678, + 0.003323653945699334, + 0.0037508239038288593, + 0.004146797116845846, + 0.005373960826545954, + 0.005345270968973637, + 0.004160965792834759, + 0.004194066859781742, + 0.0025770310312509537, + 0.0036731897853314877, + 0.004093783907592297, + 0.004547278396785259, + 0.005000577308237553, + 0.005865372717380524, + 0.00618409737944603, + 0.0064796279184520245, + 0.008281425572931767, + 0.004435245413333178, + 0.004605273716151714, + 0.004637255799025297, + 0.0037626931443810463, + 0.004663038067519665, + 0.010360750369727612, + 0.006337003316730261, + 0.004573191981762648, + 0.0061788964085280895, + 0.005816074553877115, + 0.007725936360657215, + 0.005549225024878979, + 0.011096545495092869, + 0.0034647539723664522, + 0.0042574163526296616, + 0.0027385097928345203, + 0.003088233759626746, + 0.0034014054108411074, + 0.0033536325208842754, + 0.002950862282887101, + 0.004201993346214294, + 0.005021980032324791, + 0.003963573835790157, + 0.003643343225121498, + 0.004066893365234137, + 0.002715583425015211, + 0.004100221674889326, + 0.0025547114200890064, + 0.004803210962563753, + 0.08288859575986862, + 0.12111317366361618, + 0.08097855746746063, + 0.06247343122959137, + 0.07827136665582657, + 0.07656398415565491, + 0.05228285491466522, + 0.051053401082754135, + 0.004565741401165724, + 0.004344090819358826, + 0.002821663860231638, + 0.00341043621301651, + 0.005011706613004208, + 0.003737921128049493, + 0.0030195889994502068, + 0.003668054472655058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.2255585789680481 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, eye-level close-up, rendered in the vibrant Pixar animation style, focuses on a lovable, pig-like demon. This creature, with smooth, soft pink skin and a plump, rounded body, sits innocently on the surface of a large, heart-shaped leaf. Its face is dominated by large, expressive emerald green eyes, wide and filled with wonder, reflecting the surrounding light with a lustrous sheen. Just above its eyebrows, two glossy, deep purple horns begin to sprout, their dark color creating a soft contrast against its pink complexion. The massive leaf beneath it is a deep green, covered in glistening, spherical dewdrops that catch the light like tiny gemstones. When the breeze blows, the leaves ripple in a slow, rhythmic pattern, causing the demon to sway gently and deliberately back and forth, its body moving in perfect, effortless sync with the natural movements. Its face is a look of pure, captivated wonder, its mouth slightly open. The scene is illuminated by soft, warm, golden hour light that filters through the dense tree canopy, creating a shallow depth of field that blurs the lush green forest background into a beautiful, painterly bokeh of overlapping shapes and colors, creating a serene and magical atmosphere.", + "seed": 14819, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_9/index_6053_seed_14819_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026802302757278084, + "vlm_logprob": -0.026802302757278084, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.026802302757278084, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004011798650026321 + ], + "opposite_logprob_tokens": [ + -5.754011631011963 + ], + "logprob_reward": -0.004011798650026321, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004011798650026321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08641277998685837, + 0.03710482269525528, + 0.021918080747127533, + 0.09491586685180664, + 0.31078818440437317, + 0.16928730905056, + 0.08891420811414719, + 0.08188663423061371, + 0.03711581230163574, + 0.07165632396936417 + ], + "credit_map": [ + 0.015563087537884712, + 0.004937965422868729, + 0.018851665779948235, + 0.008808734826743603, + 0.005851309280842543, + 0.012490469962358475, + 0.007607087958604097, + 0.0123024582862854, + 0.005607740953564644, + 0.002441064454615116, + 0.005764239467680454, + 0.005006866995245218, + 0.003003960708156228, + 0.004807468969374895, + 0.005652643274515867, + 0.0048208339139819145, + 0.002715585520491004, + 0.002197830006480217, + 0.004577831830829382, + 0.0030167829245328903, + 0.0023221357259899378, + 0.002694288967177272, + 0.002133376896381378, + 0.002260246081277728, + 0.014546768739819527, + 0.011429370380938053, + 0.009609229862689972, + 0.01246577687561512, + 0.012195296585559845, + 0.014979059807956219, + 0.009980020113289356, + 0.009710332378745079, + 0.05145440995693207, + 0.02968703769147396, + 0.03481511399149895, + 0.030517974868416786, + 0.030162304639816284, + 0.043992988765239716, + 0.03829126060009003, + 0.05186707526445389, + 0.02711673453450203, + 0.011820392683148384, + 0.025249062106013298, + 0.017477042973041534, + 0.009517711587250233, + 0.02104024402797222, + 0.020796077325940132, + 0.03627002239227295, + 0.011155745014548302, + 0.00891003105789423, + 0.013214929029345512, + 0.009904015809297562, + 0.010980014689266682, + 0.011189487762749195, + 0.009624584577977657, + 0.013935391791164875, + 0.008462384343147278, + 0.010652018710970879, + 0.008548334240913391, + 0.010586907155811787, + 0.014218995347619057, + 0.008156820200383663, + 0.01086871325969696, + 0.010392459109425545, + 0.003588826395571232, + 0.004479765426367521, + 0.0028146326076239347, + 0.003977950196713209, + 0.004000330343842506, + 0.005708392709493637, + 0.005208753049373627, + 0.007337160408496857, + 0.004026524256914854, + 0.0032086470164358616, + 0.0030635236762464046, + 0.004158797673881054, + 0.003628286998718977, + 0.025933779776096344, + 0.004585789516568184, + 0.023050962015986443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.004011798650026321 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195420140400529 + ], + "opposite_logprob_tokens": [ + -5.753195285797119 + ], + "logprob_reward": -0.003195420140400529, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003195420140400529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12248522788286209, + 0.08426728844642639, + 0.072630874812603, + 0.052091553807258606, + 0.33909475803375244, + 0.1221386268734932, + 0.05938683822751045, + 0.04993977025151253, + 0.05961304157972336, + 0.03835194930434227 + ], + "credit_map": [ + 0.01977221667766571, + 0.007557868957519531, + 0.017651040107011795, + 0.010522646829485893, + 0.010860646143555641, + 0.023175613954663277, + 0.012038621120154858, + 0.020906593650579453, + 0.016581378877162933, + 0.006353538017719984, + 0.013130669482052326, + 0.008191670291125774, + 0.009440918453037739, + 0.013307853601872921, + 0.010847941972315311, + 0.006413324270397425, + 0.01567317545413971, + 0.004438247065991163, + 0.013525419868528843, + 0.00405937759205699, + 0.009873117320239544, + 0.009157073684036732, + 0.0053778947331011295, + 0.010526581667363644, + 0.00691497977823019, + 0.007444112561643124, + 0.006757004652172327, + 0.0068840146996080875, + 0.003631822532042861, + 0.00595759367570281, + 0.007265223190188408, + 0.007236805744469166, + 0.046136245131492615, + 0.021779701113700867, + 0.049082349985837936, + 0.02561904862523079, + 0.021187322214245796, + 0.0712866485118866, + 0.03926289081573486, + 0.06474056094884872, + 0.021440593525767326, + 0.005456918850541115, + 0.025514990091323853, + 0.005861264653503895, + 0.004481492098420858, + 0.015325234271585941, + 0.011007052846252918, + 0.03305109217762947, + 0.006112160161137581, + 0.007843012921512127, + 0.007750253193080425, + 0.008647049777209759, + 0.008455573581159115, + 0.00675846915692091, + 0.007864314131438732, + 0.005956009496003389, + 0.005239483434706926, + 0.005659725051373243, + 0.005009491462260485, + 0.005434355698525906, + 0.005924368277192116, + 0.006965243257582188, + 0.007778973784297705, + 0.007928130216896534, + 0.005485835950821638, + 0.006258539389818907, + 0.0041705467738211155, + 0.005656729452311993, + 0.0063514551147818565, + 0.010741197504103184, + 0.0072721634060144424, + 0.013676578179001808, + 0.0029228338971734047, + 0.002160770818591118, + 0.0040674470365047455, + 0.002188738202676177, + 0.002671777969226241, + 0.010925586335361004, + 0.004824700299650431, + 0.008590098470449448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.003195420140400529 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028228219598531723 + ], + "opposite_logprob_tokens": [ + -5.8778228759765625 + ], + "logprob_reward": -0.0028228219598531723, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028228219598531723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22737722098827362, + 0.1592491865158081, + 0.045280348509550095, + 0.2134680449962616, + 0.1287030428647995, + 0.08564429730176926, + 0.03239329904317856, + 0.026973256841301918, + 0.060361120849847794, + 0.02055017091333866 + ], + "credit_map": [ + 0.04748603701591492, + 0.014413624070584774, + 0.03688595071434975, + 0.017988642677664757, + 0.020782625302672386, + 0.03130943700671196, + 0.023834742605686188, + 0.034676142036914825, + 0.03801288828253746, + 0.009861651808023453, + 0.015646034851670265, + 0.009937499649822712, + 0.018756074830889702, + 0.03817832097411156, + 0.015101119875907898, + 0.013755593448877335, + 0.008126690052449703, + 0.0034081360790878534, + 0.006440266501158476, + 0.003220324870198965, + 0.00823952816426754, + 0.005621508229523897, + 0.0036967599298805, + 0.0065271309576928616, + 0.025083055719733238, + 0.026012646034359932, + 0.022732103243470192, + 0.029120855033397675, + 0.028604062274098396, + 0.027479683980345726, + 0.024238772690296173, + 0.03019685670733452, + 0.018050525337457657, + 0.006537908222526312, + 0.01884276233613491, + 0.007875139825046062, + 0.0064282724633812904, + 0.03017234429717064, + 0.011793533340096474, + 0.02900255098938942, + 0.010022161528468132, + 0.007245300803333521, + 0.013193143531680107, + 0.010283824056386948, + 0.005988495424389839, + 0.009584395214915276, + 0.01554928719997406, + 0.013777670450508595, + 0.0034802216105163097, + 0.004309968091547489, + 0.005541857331991196, + 0.0036363748367875814, + 0.004356402903795242, + 0.004077033139765263, + 0.004287068732082844, + 0.0027043712325394154, + 0.0048003531992435455, + 0.002224771771579981, + 0.0027224563527852297, + 0.002733323723077774, + 0.002440738957375288, + 0.0041822707280516624, + 0.003106197575107217, + 0.004763142205774784, + 0.0052618300542235374, + 0.005896407645195723, + 0.005192198790609837, + 0.008732245303690434, + 0.005896418821066618, + 0.008465863764286041, + 0.010312040336430073, + 0.010604110546410084, + 0.001614092499949038, + 0.001050501479767263, + 0.0014400213258340955, + 0.001491346163675189, + 0.0011243976186960936, + 0.006080547347664833, + 0.0018689028220251203, + 0.005880360491573811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0028228219598531723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077453382313251 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077453382313251, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077453382313251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06044343113899231, + 0.01586354710161686, + 0.029182903468608856, + 0.06851933151483536, + 0.7026546001434326, + 0.015193645842373371, + 0.015222729183733463, + 0.030199522152543068, + 0.043951280415058136, + 0.01876896247267723 + ], + "credit_map": [ + 0.013263304717838764, + 0.003933012951165438, + 0.011333346366882324, + 0.0037540846969932318, + 0.005468844436109066, + 0.009948646649718285, + 0.005533733405172825, + 0.007208461873233318, + 0.0020591439679265022, + 0.0011276205768808722, + 0.0034872167743742466, + 0.0010533748427405953, + 0.0017448865110054612, + 0.0023801231291145086, + 0.0018413159996271133, + 0.0021698654163628817, + 0.006015202030539513, + 0.002456208225339651, + 0.00409721489995718, + 0.002628500107675791, + 0.004774139262735844, + 0.004277490079402924, + 0.0023889290168881416, + 0.002545221708714962, + 0.011958882212638855, + 0.007073538843542337, + 0.007738287094980478, + 0.007225878536701202, + 0.007549915928393602, + 0.01351591944694519, + 0.00610552029684186, + 0.007351385895162821, + 0.10015735030174255, + 0.04052315279841423, + 0.1017918661236763, + 0.0503273531794548, + 0.04120203107595444, + 0.15746641159057617, + 0.07938002794981003, + 0.1318064033985138, + 0.0024668718688189983, + 0.0017275039572268724, + 0.0025708714965730906, + 0.0017225266201421618, + 0.0016048516845330596, + 0.0015085549093782902, + 0.0017635641852393746, + 0.0018289013532921672, + 0.002653277711942792, + 0.0016168684232980013, + 0.0020965812727808952, + 0.00122393190395087, + 0.0014541196869686246, + 0.0016070844139903784, + 0.0015330931637436152, + 0.003037772374227643, + 0.004370870068669319, + 0.0027484309393912554, + 0.005698773544281721, + 0.0029297987930476665, + 0.0027672494761645794, + 0.004275024868547916, + 0.0033919448032975197, + 0.0040174308232963085, + 0.0066425492987036705, + 0.004933726508170366, + 0.004398823715746403, + 0.0035117114894092083, + 0.0037318288814276457, + 0.008875646628439426, + 0.004194214474409819, + 0.007662781048566103, + 0.0014039356028661132, + 0.0010049976408481598, + 0.001072894548997283, + 0.0011218809522688389, + 0.0009479553555138409, + 0.006288584321737289, + 0.001463185646571219, + 0.00546552799642086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.011077453382313251 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290401965379715 + ], + "opposite_logprob_tokens": [ + -2.2379040718078613 + ], + "logprob_reward": -0.11290401965379715, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290401965379715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25623494386672974, + 0.02172999456524849, + 0.10941572487354279, + 0.08818022161722183, + 0.19619224965572357, + 0.16250377893447876, + 0.040044523775577545, + 0.04405062273144722, + 0.04998262971639633, + 0.031665313988924026 + ], + "credit_map": [ + 0.05980437994003296, + 0.015714380890130997, + 0.035177748650312424, + 0.02069314941763878, + 0.020222365856170654, + 0.04158330708742142, + 0.028703659772872925, + 0.03433595225214958, + 0.0034246044233441353, + 0.0013675515074282885, + 0.0033674531150609255, + 0.0026801202911883593, + 0.0021845984738320112, + 0.0037690582685172558, + 0.002743025543168187, + 0.0021935831755399704, + 0.025876712054014206, + 0.008513846434652805, + 0.010951444506645203, + 0.008411778137087822, + 0.02459545060992241, + 0.013406548649072647, + 0.009414133615791798, + 0.008245809003710747, + 0.011919408105313778, + 0.011435605585575104, + 0.008405004628002644, + 0.012616913765668869, + 0.01318407990038395, + 0.013225204311311245, + 0.009672972373664379, + 0.007721031084656715, + 0.03364443778991699, + 0.021050095558166504, + 0.016833173111081123, + 0.021709049120545387, + 0.020688600838184357, + 0.03049628436565399, + 0.029503775760531425, + 0.022266829386353493, + 0.034740231931209564, + 0.008195680566132069, + 0.03124246560037136, + 0.006303767673671246, + 0.004759870935231447, + 0.022106030955910683, + 0.01141151413321495, + 0.04374422878026962, + 0.004398801829665899, + 0.006463091354817152, + 0.00555206136777997, + 0.004783615469932556, + 0.005599253345280886, + 0.004680100828409195, + 0.005839080549776554, + 0.0027285220567137003, + 0.006181746255606413, + 0.005516231060028076, + 0.0057561956346035, + 0.005446707829833031, + 0.005686859134584665, + 0.004919956438243389, + 0.004106043837964535, + 0.00643688440322876, + 0.0057036676444113255, + 0.003915830049663782, + 0.0055734240449965, + 0.007580931298434734, + 0.0033860167022794485, + 0.00905440654605627, + 0.00741625064983964, + 0.007352103479206562, + 0.003670282894745469, + 0.0017511106561869383, + 0.0031202768441289663, + 0.002446564380079508, + 0.0020160849671810865, + 0.008290876634418964, + 0.002666400745511055, + 0.007703719194978476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.11290401965379715 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, eye-level close-up, rendered in the vibrant Pixar animation style, focuses on a lovable, pig-like demon. This creature, with smooth, soft pink skin and a plump, rounded body, sits innocently on the surface of a large, heart-shaped leaf. Its face is dominated by large, expressive emerald green eyes, wide and filled with wonder, reflecting the surrounding light with a lustrous sheen. Just above its eyebrows, two glossy, deep purple horns begin to sprout, their dark color creating a soft contrast against its pink complexion. The massive leaf beneath it is a deep green, covered in glistening, spherical dewdrops that catch the light like tiny gemstones. When the breeze blows, the leaves ripple in a slow, rhythmic pattern, causing the demon to sway gently and deliberately back and forth, its body moving in perfect, effortless sync with the natural movements. Its face is a look of pure, captivated wonder, its mouth slightly open. The scene is illuminated by soft, warm, golden hour light that filters through the dense tree canopy, creating a shallow depth of field that blurs the lush green forest background into a beautiful, painterly bokeh of overlapping shapes and colors, creating a serene and magical atmosphere.", + "seed": 14820, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_9/index_6053_seed_14820_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.032610203186050055, + "vlm_logprob": -0.032610203186050055, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.032610203186050055, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0023669579531997442 + ], + "opposite_logprob_tokens": [ + -6.75236701965332 + ], + "logprob_reward": -0.0023669579531997442, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0023669579531997442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17503930628299713, + 0.14118626713752747, + 0.07958776503801346, + 0.09325200319290161, + 0.06835364550352097, + 0.09592416882514954, + 0.09712998569011688, + 0.07973746210336685, + 0.09157919138669968, + 0.07821017503738403 + ], + "credit_map": [ + 0.041834838688373566, + 0.017279265448451042, + 0.02314474619925022, + 0.014821707271039486, + 0.017055537551641464, + 0.025840310379862785, + 0.019719010218977928, + 0.015343889594078064, + 0.025218522176146507, + 0.016795959323644638, + 0.013730219565331936, + 0.014155508950352669, + 0.018892958760261536, + 0.016260499134659767, + 0.02352753095328808, + 0.012605068273842335, + 0.01370988693088293, + 0.007895143702626228, + 0.012145617976784706, + 0.008616819977760315, + 0.009288258850574493, + 0.008979970589280128, + 0.011451770551502705, + 0.0075002992525696754, + 0.012526070699095726, + 0.011079320684075356, + 0.008895563893020153, + 0.009923056699335575, + 0.01604592613875866, + 0.01210954412817955, + 0.014749615453183651, + 0.007922903634607792, + 0.012790551409125328, + 0.006345445290207863, + 0.007370301056653261, + 0.006945907603949308, + 0.008639966137707233, + 0.009500698186457157, + 0.00793304294347763, + 0.008827729150652885, + 0.020713668316602707, + 0.008334275335073471, + 0.014759385026991367, + 0.006870514713227749, + 0.00997780729085207, + 0.00939135905355215, + 0.010864942334592342, + 0.015012218616902828, + 0.015545807778835297, + 0.00769544392824173, + 0.012261329218745232, + 0.009450333192944527, + 0.008843306452035904, + 0.010072748176753521, + 0.012483258731663227, + 0.020777754485607147, + 0.00993935763835907, + 0.009104275144636631, + 0.006916147191077471, + 0.01535426452755928, + 0.009991740807890892, + 0.009284688159823418, + 0.010223329067230225, + 0.008923660963773727, + 0.018406102433800697, + 0.008216959424316883, + 0.010115605778992176, + 0.010454472154378891, + 0.008263002149760723, + 0.012109541334211826, + 0.01148519292473793, + 0.012528316117823124, + 0.011751839891076088, + 0.010792010463774204, + 0.009690267033874989, + 0.010012315586209297, + 0.008310873992741108, + 0.008017528802156448, + 0.010558382607996464, + 0.009076954796910286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0023669579531997442 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003199222730472684 + ], + "opposite_logprob_tokens": [ + -5.753199100494385 + ], + "logprob_reward": -0.003199222730472684, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003199222730472684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2115541249513626, + 0.09826701134443283, + 0.08267438411712646, + 0.06167629361152649, + 0.051899246871471405, + 0.08153458684682846, + 0.15674856305122375, + 0.08125165104866028, + 0.07733477652072906, + 0.09705938398838043 + ], + "credit_map": [ + 0.05026959627866745, + 0.018491245806217194, + 0.037203267216682434, + 0.01779838092625141, + 0.015244853682816029, + 0.025455020368099213, + 0.01787625439465046, + 0.029215512797236443, + 0.017109094187617302, + 0.009724040515720844, + 0.01284689549356699, + 0.009177399799227715, + 0.011303220875561237, + 0.010415613651275635, + 0.015012882649898529, + 0.012677865102887154, + 0.020606640726327896, + 0.005416396073997021, + 0.013536060228943825, + 0.007412789389491081, + 0.0060956962406635284, + 0.009097805246710777, + 0.009051324799656868, + 0.011457672342658043, + 0.00796703714877367, + 0.00805842038244009, + 0.007779700215905905, + 0.0056674485094845295, + 0.009964482858777046, + 0.007497819606214762, + 0.008752801455557346, + 0.005988586228340864, + 0.010145159438252449, + 0.004182413220405579, + 0.008426188491284847, + 0.0038930955342948437, + 0.005479397252202034, + 0.008993697352707386, + 0.005907614249736071, + 0.004871683660894632, + 0.013228398747742176, + 0.009057257324457169, + 0.010150394402444363, + 0.009943239390850067, + 0.010062040761113167, + 0.008170945569872856, + 0.009335649199783802, + 0.011586661450564861, + 0.02179248258471489, + 0.01380632072687149, + 0.02113139070570469, + 0.02023576758801937, + 0.015355197712779045, + 0.016198765486478806, + 0.0186911653727293, + 0.02953747659921646, + 0.012226114980876446, + 0.00951738003641367, + 0.008764844387769699, + 0.011366813443601131, + 0.00872620940208435, + 0.00983953382819891, + 0.011149928905069828, + 0.009660826064646244, + 0.015867499634623528, + 0.006975606549531221, + 0.007677379995584488, + 0.009818961843848228, + 0.006580458488315344, + 0.009602276608347893, + 0.01216359343379736, + 0.008649000898003578, + 0.019851619377732277, + 0.0061715468764305115, + 0.01694919355213642, + 0.009045307524502277, + 0.00687197782099247, + 0.011343336664140224, + 0.009033489972352982, + 0.017792902886867523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.003199222730472684 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036207374650985003 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036207374650985003, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036207374650985003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20553027093410492, + 0.12502160668373108, + 0.0628650039434433, + 0.06169668585062027, + 0.06256252527236938, + 0.09474495053291321, + 0.15828150510787964, + 0.10356147587299347, + 0.07850293815135956, + 0.04723294451832771 + ], + "credit_map": [ + 0.04049808159470558, + 0.024404099211096764, + 0.03173055127263069, + 0.01844620518386364, + 0.0209936685860157, + 0.02778065949678421, + 0.017098145559430122, + 0.02457888051867485, + 0.01940605789422989, + 0.012862893752753735, + 0.012320749461650848, + 0.013586915098130703, + 0.0197506882250309, + 0.013946125283837318, + 0.022147392854094505, + 0.011000802740454674, + 0.015043324790894985, + 0.00510804308578372, + 0.011559251695871353, + 0.005144469905644655, + 0.006138209719210863, + 0.0060625397600233555, + 0.005857518874108791, + 0.007951650768518448, + 0.007522255647927523, + 0.009020908735692501, + 0.006250767037272453, + 0.007126196287572384, + 0.010607883334159851, + 0.007581391837447882, + 0.009349641390144825, + 0.004237649496644735, + 0.010434598661959171, + 0.0058695110492408276, + 0.006868710741400719, + 0.007292053662240505, + 0.0090857008472085, + 0.00863048154860735, + 0.00789352972060442, + 0.0064879427663981915, + 0.017169540748000145, + 0.010785147547721863, + 0.011654204688966274, + 0.009230093099176884, + 0.012414045631885529, + 0.009986352175474167, + 0.009997835382819176, + 0.013507740572094917, + 0.0223799180239439, + 0.013526181690394878, + 0.021326035261154175, + 0.020203281193971634, + 0.016047846525907516, + 0.01649964228272438, + 0.019415760412812233, + 0.028882863000035286, + 0.01418035477399826, + 0.011301026679575443, + 0.011688660830259323, + 0.013592980802059174, + 0.012547035701572895, + 0.012631074525415897, + 0.01388297788798809, + 0.013737381435930729, + 0.014886542223393917, + 0.006802860181778669, + 0.006841764319688082, + 0.010796060785651207, + 0.00708301505073905, + 0.009196212515234947, + 0.014224694110453129, + 0.00867179874330759, + 0.009665499441325665, + 0.004616049118340015, + 0.006135646719485521, + 0.006184769328683615, + 0.005210378672927618, + 0.0050482479855418205, + 0.005421293433755636, + 0.004951064474880695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0036207374650985003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079457588493824 + ], + "opposite_logprob_tokens": [ + -4.51107931137085 + ], + "logprob_reward": -0.011079457588493824, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079457588493824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20014721155166626, + 0.11397483199834824, + 0.08689821511507034, + 0.03422567620873451, + 0.0506582073867321, + 0.12756210565567017, + 0.2076801061630249, + 0.08628886938095093, + 0.05565071105957031, + 0.03691405430436134 + ], + "credit_map": [ + 0.04254443943500519, + 0.02091073989868164, + 0.03749927878379822, + 0.01624169573187828, + 0.01661119982600212, + 0.023103633895516396, + 0.02091449685394764, + 0.022321756929159164, + 0.0194181390106678, + 0.011501322500407696, + 0.01644187793135643, + 0.01066168025135994, + 0.013960686512291431, + 0.012464056722819805, + 0.019316771999001503, + 0.010210307314991951, + 0.01917003095149994, + 0.0060116564854979515, + 0.01814587414264679, + 0.00878346897661686, + 0.005913841538131237, + 0.008325803093612194, + 0.009341451339423656, + 0.011206105351448059, + 0.005658945068717003, + 0.003265398321673274, + 0.00606111204251647, + 0.0023586645256727934, + 0.0036762740928679705, + 0.0041986131109297276, + 0.003875099355354905, + 0.005131575744599104, + 0.009185903705656528, + 0.0043002828024327755, + 0.007705263327807188, + 0.005173718091100454, + 0.005890828091651201, + 0.006507073529064655, + 0.0064960019662976265, + 0.005399142391979694, + 0.019755017012357712, + 0.015403441153466702, + 0.014757168479263783, + 0.013454570434987545, + 0.0175460297614336, + 0.014200001955032349, + 0.01353059895336628, + 0.018915288150310516, + 0.027580270543694496, + 0.019523613154888153, + 0.025366831570863724, + 0.029902542009949684, + 0.024069711565971375, + 0.022822648286819458, + 0.029174624010920525, + 0.02923988178372383, + 0.017906347289681435, + 0.008296191692352295, + 0.01031429786235094, + 0.011784904636442661, + 0.007818074896931648, + 0.009996358305215836, + 0.009062893688678741, + 0.011109811253845692, + 0.009654343128204346, + 0.003920233342796564, + 0.010513151995837688, + 0.00517526688054204, + 0.0034791443031281233, + 0.005232634488493204, + 0.00665704719722271, + 0.011018896475434303, + 0.006119174882769585, + 0.0028497532475739717, + 0.005366146098822355, + 0.003524811938405037, + 0.00396359758451581, + 0.004423889331519604, + 0.004063444212079048, + 0.006603240966796875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.011079457588493824 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278464019298553 + ], + "opposite_logprob_tokens": [ + -2.017784595489502 + ], + "logprob_reward": -0.14278464019298553, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278464019298553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30141517519950867, + 0.09565873444080353, + 0.04611165076494217, + 0.04007089138031006, + 0.0598859004676342, + 0.05364196002483368, + 0.06840314716100693, + 0.07575351744890213, + 0.12323378771543503, + 0.13582521677017212 + ], + "credit_map": [ + 0.0553436316549778, + 0.03254163637757301, + 0.05298476666212082, + 0.025181515142321587, + 0.03145249933004379, + 0.03957347944378853, + 0.03311681002378464, + 0.031220823526382446, + 0.016251638531684875, + 0.008483607321977615, + 0.015573987737298012, + 0.008955740369856358, + 0.009917402639985085, + 0.010401351377367973, + 0.014151602052152157, + 0.011923403479158878, + 0.009188356809318066, + 0.00389495468698442, + 0.00842615682631731, + 0.004733507987111807, + 0.0037836527917534113, + 0.004726976156234741, + 0.0050736963748931885, + 0.006284351926296949, + 0.007710018195211887, + 0.00352427433244884, + 0.004422437399625778, + 0.004242005292326212, + 0.005837289150804281, + 0.004995664581656456, + 0.00575977610424161, + 0.003579426556825638, + 0.012220208533108234, + 0.005346242338418961, + 0.005682683549821377, + 0.006143434904515743, + 0.007283468265086412, + 0.009522470645606518, + 0.007626665756106377, + 0.006060726474970579, + 0.012300979346036911, + 0.005580039229243994, + 0.008212701417505741, + 0.004573664162307978, + 0.0036652476992458105, + 0.007695361040532589, + 0.004668281879276037, + 0.006945683620870113, + 0.01112912967801094, + 0.006641734391450882, + 0.00800822302699089, + 0.009826895780861378, + 0.006163537967950106, + 0.007558055687695742, + 0.008481020107865334, + 0.010594552382826805, + 0.008015740662813187, + 0.010461734607815742, + 0.005428794771432877, + 0.010347920469939709, + 0.010110096074640751, + 0.009723170660436153, + 0.012310699559748173, + 0.00935535877943039, + 0.023230310529470444, + 0.011869432404637337, + 0.009422344155609608, + 0.019146529957652092, + 0.012537574395537376, + 0.01277310959994793, + 0.022722475230693817, + 0.011532027274370193, + 0.02077176608145237, + 0.011605199426412582, + 0.022162701934576035, + 0.013408014550805092, + 0.010239670984447002, + 0.01927090808749199, + 0.01343245804309845, + 0.024934493005275726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.14278464019298553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, eye-level close-up, rendered in the vibrant Pixar animation style, focuses on a lovable, pig-like demon. This creature, with smooth, soft pink skin and a plump, rounded body, sits innocently on the surface of a large, heart-shaped leaf. Its face is dominated by large, expressive emerald green eyes, wide and filled with wonder, reflecting the surrounding light with a lustrous sheen. Just above its eyebrows, two glossy, deep purple horns begin to sprout, their dark color creating a soft contrast against its pink complexion. The massive leaf beneath it is a deep green, covered in glistening, spherical dewdrops that catch the light like tiny gemstones. When the breeze blows, the leaves ripple in a slow, rhythmic pattern, causing the demon to sway gently and deliberately back and forth, its body moving in perfect, effortless sync with the natural movements. Its face is a look of pure, captivated wonder, its mouth slightly open. The scene is illuminated by soft, warm, golden hour light that filters through the dense tree canopy, creating a shallow depth of field that blurs the lush green forest background into a beautiful, painterly bokeh of overlapping shapes and colors, creating a serene and magical atmosphere.", + "seed": 14821, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000014/group_9/index_6053_seed_14821_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02889138078317046, + "vlm_logprob": -0.02889138078317046, + "vlm_yes_no_margin": 4.5749999523162845, + "vlm_token_credit": -0.02889138078317046, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003418912645429373 + ], + "opposite_logprob_tokens": [ + -6.003418922424316 + ], + "logprob_reward": -0.003418912645429373, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.003418912645429373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27574753761291504, + 0.052606724202632904, + 0.07894296199083328, + 0.029521530494093895, + 0.12019050121307373, + 0.052302535623311996, + 0.0897674560546875, + 0.1720418483018875, + 0.09009378403425217, + 0.03878524526953697 + ], + "credit_map": [ + 0.060790807008743286, + 0.012003395706415176, + 0.019374025985598564, + 0.009608548134565353, + 0.024427395313978195, + 0.10268193483352661, + 0.011582501232624054, + 0.035278867930173874, + 0.008162176236510277, + 0.00997346080839634, + 0.006937077268958092, + 0.00404561311006546, + 0.005916915368288755, + 0.006644673179835081, + 0.005377034656703472, + 0.005549764726310968, + 0.014214104041457176, + 0.006661262828856707, + 0.004002575762569904, + 0.0062460098415613174, + 0.005882193334400654, + 0.023837145417928696, + 0.008558245375752449, + 0.009541417472064495, + 0.003925558179616928, + 0.0027387100271880627, + 0.0026395765598863363, + 0.0031095410231500864, + 0.003729419317096472, + 0.0045869299210608006, + 0.00503654545173049, + 0.0037552479188889265, + 0.012638481333851814, + 0.013935104943811893, + 0.014738790690898895, + 0.00819416344165802, + 0.011697803623974323, + 0.02869419753551483, + 0.01031358540058136, + 0.019978368654847145, + 0.007495350204408169, + 0.003655480220913887, + 0.005160956177860498, + 0.0029951417818665504, + 0.0025107853580266237, + 0.013496845960617065, + 0.00414306903257966, + 0.012844902463257313, + 0.01330465730279684, + 0.007301387842744589, + 0.009731310419738293, + 0.006173545029014349, + 0.006738499738276005, + 0.020204119384288788, + 0.007447817828506231, + 0.018866103142499924, + 0.018181409686803818, + 0.041774969547986984, + 0.008823435753583908, + 0.009346681647002697, + 0.008562530390918255, + 0.06322228908538818, + 0.00768532557412982, + 0.014445193111896515, + 0.010645654983818531, + 0.012992149218916893, + 0.007622090633958578, + 0.009184550493955612, + 0.009686650708317757, + 0.013590707443654537, + 0.011931917630136013, + 0.014440052211284637, + 0.005221727769821882, + 0.005362371448427439, + 0.0038490844890475273, + 0.003847825340926647, + 0.004108776338398457, + 0.0041834041476249695, + 0.005781393498182297, + 0.006430655252188444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.003418912645429373 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003196252044290304 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.003196252044290304, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003196252044290304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12226599454879761, + 0.1506238877773285, + 0.15382149815559387, + 0.0667751356959343, + 0.11364373564720154, + 0.1239301860332489, + 0.0747838020324707, + 0.09173587709665298, + 0.07597579061985016, + 0.02644411101937294 + ], + "credit_map": [ + 0.021641168743371964, + 0.008951098658144474, + 0.012204236350953579, + 0.008481412194669247, + 0.012286363169550896, + 0.0330728217959404, + 0.013834825716912746, + 0.011794055812060833, + 0.024294888600707054, + 0.02279162034392357, + 0.02955380268394947, + 0.009595230221748352, + 0.009772101417183876, + 0.018006931990385056, + 0.011714254505932331, + 0.02489505149424076, + 0.03341469168663025, + 0.007350509520620108, + 0.00831408891826868, + 0.00787037331610918, + 0.006596088409423828, + 0.060262493789196014, + 0.0073699504137039185, + 0.02264331467449665, + 0.01129623968154192, + 0.004315635655075312, + 0.008685224689543247, + 0.004215971566736698, + 0.0042452458292245865, + 0.01319417729973793, + 0.005347305443137884, + 0.015475335530936718, + 0.01249682530760765, + 0.013609424233436584, + 0.010479596443474293, + 0.008040513843297958, + 0.010122007690370083, + 0.030094832181930542, + 0.010694103315472603, + 0.018106428906321526, + 0.020764457061886787, + 0.006188969127833843, + 0.01360662654042244, + 0.006627345457673073, + 0.005781329236924648, + 0.04042435064911842, + 0.00867769680917263, + 0.021859416738152504, + 0.011602088809013367, + 0.006128426641225815, + 0.008163167163729668, + 0.005074819549918175, + 0.004588358104228973, + 0.016754373908042908, + 0.006968855392187834, + 0.015503712929785252, + 0.017203308641910553, + 0.006618335843086243, + 0.012975594028830528, + 0.005405792035162449, + 0.0041101304814219475, + 0.02467651292681694, + 0.007493041921406984, + 0.013253162615001202, + 0.012636625207960606, + 0.004580255132168531, + 0.013187330216169357, + 0.00503711448982358, + 0.004100685939192772, + 0.015829481184482574, + 0.005688583478331566, + 0.014915719628334045, + 0.003475932637229562, + 0.0030669623520225286, + 0.004891902208328247, + 0.0029310216195881367, + 0.0024524445179849863, + 0.002471577376127243, + 0.003100213361904025, + 0.0040540555492043495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.003196252044290304 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740807089954615 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740807089954615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740807089954615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22385063767433167, + 0.06413519382476807, + 0.10295110940933228, + 0.02555507980287075, + 0.08699657022953033, + 0.10836922377347946, + 0.10081292688846588, + 0.2011675387620926, + 0.05174655467271805, + 0.03441517427563667 + ], + "credit_map": [ + 0.048556290566921234, + 0.011490670032799244, + 0.01512510422617197, + 0.008405008353292942, + 0.02111264318227768, + 0.0784342885017395, + 0.013841448351740837, + 0.026885176077485085, + 0.009866250678896904, + 0.01213810034096241, + 0.008555209264159203, + 0.0052599250338971615, + 0.0061663733795285225, + 0.007710824720561504, + 0.0070375786162912846, + 0.007400935515761375, + 0.0233107078820467, + 0.005209444090723991, + 0.005791847128421068, + 0.005397012922912836, + 0.00616154819726944, + 0.040036823600530624, + 0.007210962940007448, + 0.009832760319113731, + 0.0034067179076373577, + 0.003202359890565276, + 0.002615771722048521, + 0.0018903008894994855, + 0.004218778107315302, + 0.003121807239949703, + 0.0026395623572170734, + 0.004459782037883997, + 0.011378735303878784, + 0.0081481346860528, + 0.012775227427482605, + 0.004673297982662916, + 0.009072288870811462, + 0.016851801425218582, + 0.009471822530031204, + 0.014625267125666142, + 0.018541553989052773, + 0.006032200530171394, + 0.011872139759361744, + 0.0061960043385624886, + 0.0051769232377409935, + 0.035399191081523895, + 0.007863819599151611, + 0.017287395894527435, + 0.016228020191192627, + 0.006550881080329418, + 0.00986794475466013, + 0.0061795348301529884, + 0.005942744202911854, + 0.024600598961114883, + 0.007632296532392502, + 0.023810911923646927, + 0.04812098294496536, + 0.01393997110426426, + 0.012491022236645222, + 0.013832390308380127, + 0.008529006503522396, + 0.07328665256500244, + 0.012323025614023209, + 0.01864450052380562, + 0.00823238492012024, + 0.004125988110899925, + 0.0049437652342021465, + 0.005021146964281797, + 0.003378765657544136, + 0.013784877955913544, + 0.004720701836049557, + 0.007538923062384129, + 0.004770271480083466, + 0.0047207982279360294, + 0.0036328269634395838, + 0.003874402027577162, + 0.003981341607868671, + 0.003742049215361476, + 0.0044228630140423775, + 0.00527062127366662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.006740807089954615 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188709393143654 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.018188709393143654, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018188709393143654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17862030863761902, + 0.0970979854464531, + 0.059798043221235275, + 0.025041576474905014, + 0.20676691830158234, + 0.14620600640773773, + 0.18807224929332733, + 0.05825219303369522, + 0.022903019562363625, + 0.017241721972823143 + ], + "credit_map": [ + 0.04150472953915596, + 0.01083084661513567, + 0.012685042805969715, + 0.006579275708645582, + 0.016837215051054955, + 0.06884969025850296, + 0.007427801378071308, + 0.013905715197324753, + 0.015682952478528023, + 0.006287296302616596, + 0.020270079374313354, + 0.0057479036040604115, + 0.00524898711591959, + 0.020940685644745827, + 0.007462240289896727, + 0.015457838773727417, + 0.013782399706542492, + 0.0025242152623832226, + 0.0030000731348991394, + 0.0033974989783018827, + 0.0025677925441414118, + 0.02347649447619915, + 0.004105853848159313, + 0.006943716201931238, + 0.004062944557517767, + 0.0020466528367251158, + 0.0030364380218088627, + 0.0024226561654359102, + 0.0024958255235105753, + 0.002713022055104375, + 0.004111444111913443, + 0.004152592737227678, + 0.021236252039670944, + 0.030489658936858177, + 0.018541298806667328, + 0.013999075628817081, + 0.018216244876384735, + 0.056869372725486755, + 0.019012637436389923, + 0.028402384370565414, + 0.028029944747686386, + 0.006396793760359287, + 0.013052236288785934, + 0.008479060605168343, + 0.0054005710408091545, + 0.0546245314180851, + 0.008767819032073021, + 0.02145504578948021, + 0.031133338809013367, + 0.010524486191570759, + 0.018529461696743965, + 0.01051047071814537, + 0.009304031729698181, + 0.042861152440309525, + 0.013938912190496922, + 0.05127039551734924, + 0.008302857168018818, + 0.021153893321752548, + 0.002876399317756295, + 0.006299803964793682, + 0.003943991847336292, + 0.009741219691932201, + 0.002493562875315547, + 0.0034404704347252846, + 0.003177780658006668, + 0.0020456237252801657, + 0.002609193092212081, + 0.0020636534318327904, + 0.0017702743643894792, + 0.0031615961343050003, + 0.002583802444860339, + 0.005491096992045641, + 0.0022793051321059465, + 0.002971008885651827, + 0.0011626932537183166, + 0.0018307517748326063, + 0.002304447814822197, + 0.0016400640597566962, + 0.0026278120931237936, + 0.0024256391916424036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.018188709393143654 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pig-like demon with smooth, soft pink skin and a plump, rounded body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature sitting on a large, heart-shaped leaf that is deep green and covered in glistening spherical dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have two glossy, deep purple horns sprouting above its eyebrows and large, expressive emerald green eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft, warm, golden hour light filtering through a dense tree canopy with a shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf ripple in a slow, rhythmic pattern causing the demon to sway gently back and forth in sync with the movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291222274303436 + ], + "opposite_logprob_tokens": [ + -2.237912178039551 + ], + "logprob_reward": -0.11291222274303436, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291222274303436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09690333157777786, + 0.041523490101099014, + 0.07527893036603928, + 0.03909098729491234, + 0.05312952771782875, + 0.09207694977521896, + 0.07592341303825378, + 0.4504874348640442, + 0.050142884254455566, + 0.02544306218624115 + ], + "credit_map": [ + 0.01391581166535616, + 0.015146540477871895, + 0.011791740544140339, + 0.008840844966471195, + 0.01035118568688631, + 0.012177132070064545, + 0.014124023728072643, + 0.010556056164205074, + 0.004503635689616203, + 0.005379016045480967, + 0.005886620841920376, + 0.003283663187175989, + 0.004113762639462948, + 0.008914386853575706, + 0.003708291333168745, + 0.005734113045036793, + 0.015368153341114521, + 0.0035405042581260204, + 0.007600170094519854, + 0.0043389685451984406, + 0.003721744753420353, + 0.02361282706260681, + 0.005431622266769409, + 0.01166494283825159, + 0.006704400759190321, + 0.004422083031386137, + 0.004067962057888508, + 0.003145241644233465, + 0.005808001384139061, + 0.004954297095537186, + 0.004672135226428509, + 0.005316866561770439, + 0.006067092064768076, + 0.004483620170503855, + 0.006372579373419285, + 0.004047846421599388, + 0.0055364579893648624, + 0.011782534420490265, + 0.006614843849092722, + 0.008224553428590298, + 0.013985623605549335, + 0.0051031350158154964, + 0.006840384565293789, + 0.006292602512985468, + 0.005318383686244488, + 0.02572593465447426, + 0.007312473375350237, + 0.021498408168554306, + 0.011651215143501759, + 0.012775122188031673, + 0.004167016129940748, + 0.005687255412340164, + 0.0066612036898732185, + 0.01832715794444084, + 0.006323447450995445, + 0.010330994613468647, + 0.056943751871585846, + 0.09156404435634613, + 0.038550496101379395, + 0.023396605625748634, + 0.02190816029906273, + 0.15851493179798126, + 0.01920377090573311, + 0.040405675768852234, + 0.00720961345359683, + 0.004257921129465103, + 0.005492004565894604, + 0.004650818184018135, + 0.004477880895137787, + 0.010003257542848587, + 0.004973705392330885, + 0.00907768402248621, + 0.0032718884758651257, + 0.0032886117696762085, + 0.0027037616819143295, + 0.002830126089975238, + 0.002868868177756667, + 0.0037029418163001537, + 0.0033830543980002403, + 0.003393810475245118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.11291222274303436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_0/index_tensor([2475, 2475])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_0/index_tensor([2475, 2475])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..40c17ae63f0bd7ba7e9d5864fab28f99859cf750 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_0/index_tensor([2475, 2475])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide-angle tracking shot steadily follows a young boy, approximately ten years old, as he crosses an outdoor basketball court from left to right. His face is a mask of concentration, his brow slightly furrowed, his eyes fixed on the ball he dribbles. He has short black hair, his cheeks slightly damp with sweat. Dressed in a sleeveless royal blue mesh sweatshirt and matching baggy shorts, his athletic physique poised for action. With practiced expertise, he keeps his body low to the ground as he rhythmically and decisively dribbles a worn orange basketball with his right hand. The ball clings to the ground, its textured surface catching the light with each powerful bounce on the cracked gray asphalt. The court is bathed in the warm golden glow of the afternoon sun, and the boy's shadow stretches and twists with his fluid movements. In the background, the faded white paint of the three-point line is clearly visible, along with a tall chain-link fence, through which the distant green trees gently blur, creating an atmosphere of serene, focused solitary devotion and youthful ambition.", + "seed": 8902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_0/index_2475_seed_8902_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.22332539800554513, + "vlm_logprob": -0.22332539800554513, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.22332539800554513, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7572137117385864 + ], + "opposite_logprob_tokens": [ + -0.7572137117385864 + ], + "logprob_reward": -0.7572137117385864, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7572137117385864, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2194303274154663, + 0.09647717326879501, + 0.10836946219205856, + 0.115822434425354, + 0.06035512313246727, + 0.06677892804145813, + 0.1454140841960907, + 0.12173078209161758, + 0.032205160707235336, + 0.033416517078876495 + ], + "credit_map": [ + 0.024574337527155876, + 0.023409968242049217, + 0.051216889172792435, + 0.01732366904616356, + 0.02396273985505104, + 0.029867518693208694, + 0.025902407243847847, + 0.023172779008746147, + 0.008293543010950089, + 0.008215116336941719, + 0.010988102294504642, + 0.011565925553441048, + 0.012576531618833542, + 0.012539681047201157, + 0.018347494304180145, + 0.013950781896710396, + 0.008965331129729748, + 0.011249658651649952, + 0.012671764940023422, + 0.011947971768677235, + 0.014032086357474327, + 0.014045681804418564, + 0.021836264058947563, + 0.013620699755847454, + 0.006958009675145149, + 0.007875869981944561, + 0.006438552401959896, + 0.008225941099226475, + 0.019711006432771683, + 0.025962766259908676, + 0.023315418511629105, + 0.017334874719381332, + 0.0077277193777263165, + 0.00411721458658576, + 0.0053538912907242775, + 0.004486942198127508, + 0.005781437736004591, + 0.020812680944800377, + 0.006185557227581739, + 0.005889681167900562, + 0.007209467235952616, + 0.0061456551775336266, + 0.009038866497576237, + 0.004004407208412886, + 0.00880105048418045, + 0.011518055573105812, + 0.005951232742518187, + 0.014110195450484753, + 0.034254204481840134, + 0.009746799245476723, + 0.014722184278070927, + 0.005638971459120512, + 0.017285743728280067, + 0.03579976409673691, + 0.01081916131079197, + 0.017147259786725044, + 0.013583885505795479, + 0.016048207879066467, + 0.01798444613814354, + 0.01662784069776535, + 0.019621139392256737, + 0.010811922140419483, + 0.012375704012811184, + 0.014677630737423897, + 0.004119331482797861, + 0.002468092367053032, + 0.00491067161783576, + 0.002914057346060872, + 0.0034234831109642982, + 0.006097895558923483, + 0.004260918125510216, + 0.004010711330920458, + 0.003750510746613145, + 0.0024426113814115524, + 0.005975170526653528, + 0.0034699796233326197, + 0.003778046229854226, + 0.0058702933602035046, + 0.004065304063260555, + 0.004064601846039295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.7572137117385864 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026385555043816566 + ], + "opposite_logprob_tokens": [ + -3.651385545730591 + ], + "logprob_reward": -0.026385555043816566, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026385555043816566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09113520383834839, + 0.05431269481778145, + 0.054946962743997574, + 0.06367306411266327, + 0.0911647230386734, + 0.051602888852357864, + 0.0705280527472496, + 0.025936808437108994, + 0.1370050609111786, + 0.3596945106983185 + ], + "credit_map": [ + 0.00772863021120429, + 0.015936389565467834, + 0.008369510993361473, + 0.008383995853364468, + 0.0129737863317132, + 0.010325843468308449, + 0.012294212356209755, + 0.015122833661735058, + 0.0053849779069423676, + 0.0051500857807695866, + 0.00813254527747631, + 0.006650798954069614, + 0.00661665853112936, + 0.005078497342765331, + 0.011187044903635979, + 0.006112085655331612, + 0.006049908697605133, + 0.006502058357000351, + 0.006842874456197023, + 0.005279656033962965, + 0.005549021530896425, + 0.00751454895362258, + 0.009741238318383694, + 0.0074676573276519775, + 0.010075758211314678, + 0.010762638412415981, + 0.0040946053341031075, + 0.007794474251568317, + 0.008058713749051094, + 0.006573664024472237, + 0.00901477038860321, + 0.007298439275473356, + 0.013924294151365757, + 0.012477638199925423, + 0.009442416951060295, + 0.013644900172948837, + 0.008810024708509445, + 0.012939652428030968, + 0.011090205051004887, + 0.008835592307150364, + 0.00487859733402729, + 0.005842996761202812, + 0.00646231509745121, + 0.009872469119727612, + 0.008583001792430878, + 0.004927013069391251, + 0.005423165857791901, + 0.0056133330799639225, + 0.00671734195202589, + 0.01569734513759613, + 0.006862120237201452, + 0.007525651250034571, + 0.011317260563373566, + 0.007412477862089872, + 0.00799750629812479, + 0.006998348515480757, + 0.003216381883248687, + 0.00232098251581192, + 0.004451840650290251, + 0.003060770221054554, + 0.002691891510039568, + 0.0031290072947740555, + 0.003511029528453946, + 0.0035549055319279432, + 0.011247934773564339, + 0.018318146467208862, + 0.02207764983177185, + 0.028293561190366745, + 0.012172809801995754, + 0.013396207243204117, + 0.014851255342364311, + 0.016647500917315483, + 0.05749213695526123, + 0.03453764319419861, + 0.05466962233185768, + 0.0385323204100132, + 0.030867109075188637, + 0.06349679082632065, + 0.03877388685941696, + 0.041325006633996964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.026385555043816566 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307105019688606 + ], + "opposite_logprob_tokens": [ + -3.1680710315704346 + ], + "logprob_reward": -0.04307105019688606, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307105019688606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36492919921875, + 0.0731033906340599, + 0.11008632183074951, + 0.08743427693843842, + 0.03236708790063858, + 0.04863346740603447, + 0.09970477968454361, + 0.0787862092256546, + 0.03587882220745087, + 0.06907644122838974 + ], + "credit_map": [ + 0.032330472022295, + 0.03863251954317093, + 0.03785219416022301, + 0.02857188507914543, + 0.042758580297231674, + 0.06664545834064484, + 0.06659410148859024, + 0.05154398828744888, + 0.00515896687284112, + 0.007352448999881744, + 0.008583702147006989, + 0.010070240125060081, + 0.008060827851295471, + 0.008340016938745975, + 0.015901388600468636, + 0.009635799564421177, + 0.008677853271365166, + 0.009053760208189487, + 0.009984711185097694, + 0.010444404557347298, + 0.01154128648340702, + 0.011448914185166359, + 0.03179614990949631, + 0.017139242962002754, + 0.012134642340242863, + 0.007017495110630989, + 0.005275962874293327, + 0.008928383700549603, + 0.009293255396187305, + 0.007780194282531738, + 0.023999081924557686, + 0.013005262240767479, + 0.005666516721248627, + 0.002880252432078123, + 0.0028024485800415277, + 0.00400126026943326, + 0.004000789951533079, + 0.00598125159740448, + 0.003652585670351982, + 0.003381981048732996, + 0.004800434224307537, + 0.0051863607950508595, + 0.006088888738304377, + 0.007143666036427021, + 0.008939041756093502, + 0.005571078509092331, + 0.004639165010303259, + 0.006264837924391031, + 0.02533261850476265, + 0.0076221683993935585, + 0.010691280476748943, + 0.004425054881721735, + 0.011097400449216366, + 0.02204744704067707, + 0.008084210567176342, + 0.010404608212411404, + 0.010248011909425259, + 0.005854879505932331, + 0.013199914246797562, + 0.008611425757408142, + 0.010246256366372108, + 0.009958100505173206, + 0.011747746728360653, + 0.008919874206185341, + 0.00374045898206532, + 0.0030059630516916513, + 0.005613475106656551, + 0.003380322130396962, + 0.00425236439332366, + 0.006886734161525965, + 0.003432151861488819, + 0.005567352753132582, + 0.008602126501500607, + 0.004718838259577751, + 0.01207254733890295, + 0.006452856119722128, + 0.007201782893389463, + 0.013291735202074051, + 0.007922516204416752, + 0.008814037777483463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.04307105019688606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641703985631466 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.008641703985631466, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641703985631466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26680827140808105, + 0.04208846017718315, + 0.08149881660938263, + 0.20779336988925934, + 0.06501329690217972, + 0.1125650480389595, + 0.05303426459431648, + 0.04934701696038246, + 0.0500066764652729, + 0.07184477895498276 + ], + "credit_map": [ + 0.025305021554231644, + 0.028374159708619118, + 0.051121894270181656, + 0.01612839102745056, + 0.0235852412879467, + 0.04859228432178497, + 0.03810172900557518, + 0.035599544644355774, + 0.002471915213391185, + 0.003945144359022379, + 0.004153094254434109, + 0.0038501042872667313, + 0.0051375264301896095, + 0.009376997128129005, + 0.007443015463650227, + 0.005710665136575699, + 0.006570861674845219, + 0.004802275914698839, + 0.006123527884483337, + 0.005951533559709787, + 0.011282810010015965, + 0.01737508550286293, + 0.01648714765906334, + 0.012905570678412914, + 0.022632695734500885, + 0.012034685350954533, + 0.013989069499075413, + 0.02011854574084282, + 0.022152038291096687, + 0.02033139206469059, + 0.061774108558893204, + 0.03476082906126976, + 0.011485504917800426, + 0.0054954346269369125, + 0.005681077018380165, + 0.007633349392563105, + 0.006708452943712473, + 0.016818102449178696, + 0.00560566084459424, + 0.005585712380707264, + 0.014804133214056492, + 0.007574871182441711, + 0.01415168959647417, + 0.007448864635080099, + 0.01227600034326315, + 0.03148171305656433, + 0.008228529244661331, + 0.016599256545305252, + 0.007733567617833614, + 0.005522945895791054, + 0.006702576298266649, + 0.0035611684434115887, + 0.006720497272908688, + 0.008140451274812222, + 0.004574415273964405, + 0.01007864810526371, + 0.006179280113428831, + 0.005916051100939512, + 0.007829309441149235, + 0.0068969521671533585, + 0.0074840630404651165, + 0.003876253729686141, + 0.004730109591037035, + 0.006435000337660313, + 0.005176529288291931, + 0.006172414869070053, + 0.007386417128145695, + 0.01033132802695036, + 0.004649459384381771, + 0.006526456214487553, + 0.005047904327511787, + 0.004716166760772467, + 0.010627678595483303, + 0.0070035806857049465, + 0.009147006087005138, + 0.007385765202343464, + 0.006406254600733519, + 0.013254895806312561, + 0.009639028459787369, + 0.008380570448935032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.008641703985631466 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813149690628052 + ], + "opposite_logprob_tokens": [ + -1.4063149690628052 + ], + "logprob_reward": -0.2813149690628052, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813149690628052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28988412022590637, + 0.08961966633796692, + 0.08896932005882263, + 0.11076552420854568, + 0.052415117621421814, + 0.04833102226257324, + 0.02414422482252121, + 0.1388562172651291, + 0.07927669584751129, + 0.07773812115192413 + ], + "credit_map": [ + 0.030695633962750435, + 0.038004592061042786, + 0.05917046219110489, + 0.019708245992660522, + 0.025149399414658546, + 0.04969073086977005, + 0.03542761877179146, + 0.0320374071598053, + 0.007635991554707289, + 0.009484018199145794, + 0.01623675413429737, + 0.00907285325229168, + 0.0117656160145998, + 0.009748391807079315, + 0.015949584543704987, + 0.009726454503834248, + 0.005789048504084349, + 0.005182434339076281, + 0.005767412018030882, + 0.007626703940331936, + 0.016681009903550148, + 0.019765494391322136, + 0.015231065452098846, + 0.012926146388053894, + 0.007480804808437824, + 0.006602667272090912, + 0.006984620820730925, + 0.010845289565622807, + 0.009901526384055614, + 0.009746603667736053, + 0.039570726454257965, + 0.019633276388049126, + 0.005169520154595375, + 0.0057432809844613075, + 0.006729879882186651, + 0.005392236169427633, + 0.007168344687670469, + 0.010663582943379879, + 0.006216332782059908, + 0.005331934429705143, + 0.005273329559713602, + 0.0038942655082792044, + 0.00607314333319664, + 0.00527571514248848, + 0.007167853880673647, + 0.00870220921933651, + 0.004369172267615795, + 0.0075753298588097095, + 0.0030022230930626392, + 0.003657006425783038, + 0.003069524187594652, + 0.001805586856789887, + 0.0029345534276217222, + 0.0036353147588670254, + 0.002529327291995287, + 0.0035106854047626257, + 0.01637362502515316, + 0.010858666151762009, + 0.02278084121644497, + 0.015275097452104092, + 0.028051933273673058, + 0.011099833995103836, + 0.01598224602639675, + 0.018433965742588043, + 0.006739034783095121, + 0.007951822131872177, + 0.01862378604710102, + 0.007882548496127129, + 0.007906213402748108, + 0.008561421185731888, + 0.0059928507544100285, + 0.015619012527167797, + 0.019440198317170143, + 0.008759541437029839, + 0.008685681037604809, + 0.003962790127843618, + 0.005596264265477657, + 0.01745070144534111, + 0.005171151366084814, + 0.008671785704791546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.2813149690628052 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle tracking shot steadily follows a young boy, approximately ten years old, as he crosses an outdoor basketball court from left to right. His face is a mask of concentration, his brow slightly furrowed, his eyes fixed on the ball he dribbles. He has short black hair, his cheeks slightly damp with sweat. Dressed in a sleeveless royal blue mesh sweatshirt and matching baggy shorts, his athletic physique poised for action. With practiced expertise, he keeps his body low to the ground as he rhythmically and decisively dribbles a worn orange basketball with his right hand. The ball clings to the ground, its textured surface catching the light with each powerful bounce on the cracked gray asphalt. The court is bathed in the warm golden glow of the afternoon sun, and the boy's shadow stretches and twists with his fluid movements. In the background, the faded white paint of the three-point line is clearly visible, along with a tall chain-link fence, through which the distant green trees gently blur, creating an atmosphere of serene, focused solitary devotion and youthful ambition.", + "seed": 8903, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_0/index_2475_seed_8903_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2318698313087225, + "vlm_logprob": -0.2318698313087225, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.2318698313087225, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.7482321262359619 + ], + "opposite_logprob_tokens": [ + -0.8732321262359619 + ], + "logprob_reward": -0.7482321262359619, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.7482321262359619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2668682634830475, + 0.05691630020737648, + 0.2295357882976532, + 0.07415205240249634, + 0.04291152581572533, + 0.02918706089258194, + 0.09816659241914749, + 0.08325590193271637, + 0.0846560150384903, + 0.03435060381889343 + ], + "credit_map": [ + 0.029928645119071007, + 0.02619745023548603, + 0.06388717144727707, + 0.024333365261554718, + 0.025813115760684013, + 0.045683447271585464, + 0.023283522576093674, + 0.027741502970457077, + 0.008839071728289127, + 0.004424394574016333, + 0.015432490967214108, + 0.003770645707845688, + 0.00529966875910759, + 0.008915564976632595, + 0.004647950641810894, + 0.005586508195847273, + 0.0546807125210762, + 0.011966543272137642, + 0.0682503879070282, + 0.01130598969757557, + 0.010099989362061024, + 0.03438052162528038, + 0.014916387386620045, + 0.0239352248609066, + 0.008719095960259438, + 0.007756790611892939, + 0.01406126469373703, + 0.006715798284858465, + 0.0076027484610676765, + 0.0108054019510746, + 0.0074479226022958755, + 0.011043019592761993, + 0.004774435888975859, + 0.005110747180879116, + 0.005134150851517916, + 0.004263599868863821, + 0.004366861190646887, + 0.005706372205168009, + 0.0045862337574362755, + 0.008969123475253582, + 0.005216406658291817, + 0.0025133516173809767, + 0.004196899011731148, + 0.002802426926791668, + 0.002937976038083434, + 0.005503647029399872, + 0.0023752532433718443, + 0.0036410978063941, + 0.009578825905919075, + 0.012977337464690208, + 0.01184018887579441, + 0.012129309587180614, + 0.011492528952658176, + 0.008995842188596725, + 0.01857556588947773, + 0.01257698331028223, + 0.011443215422332287, + 0.00549697782844305, + 0.010485528036952019, + 0.006474342662841082, + 0.006903389468789101, + 0.00994222890585661, + 0.016189787536859512, + 0.016320427879691124, + 0.01191101223230362, + 0.013930469751358032, + 0.020021842792630196, + 0.00879928283393383, + 0.006345952861011028, + 0.010526599362492561, + 0.007237699814140797, + 0.005883151665329933, + 0.004357274621725082, + 0.0028246641159057617, + 0.005523452069610357, + 0.005700632464140654, + 0.0028040839824825525, + 0.005183863453567028, + 0.00331339449621737, + 0.004643233492970467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.7482321262359619 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006105035543442 + ], + "opposite_logprob_tokens": [ + -2.695060968399048 + ], + "logprob_reward": -0.07006105035543442, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006105035543442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09658392518758774, + 0.08313921093940735, + 0.15346775949001312, + 0.38122034072875977, + 0.03773277997970581, + 0.050658807158470154, + 0.023251106962561607, + 0.06372072547674179, + 0.06411656737327576, + 0.04610877484083176 + ], + "credit_map": [ + 0.012335245497524738, + 0.006391736678779125, + 0.02772190421819687, + 0.006081962026655674, + 0.006419355981051922, + 0.01981736160814762, + 0.006956708151847124, + 0.010859652422368526, + 0.010703190229833126, + 0.006279292982071638, + 0.017989709973335266, + 0.005175985861569643, + 0.006503702607005835, + 0.0215439535677433, + 0.006068234331905842, + 0.008875136263668537, + 0.01506037451326847, + 0.018472103402018547, + 0.028405187651515007, + 0.016059085726737976, + 0.021948300302028656, + 0.024968767538666725, + 0.01354181207716465, + 0.015012119896709919, + 0.06171803176403046, + 0.02345058135688305, + 0.06306611746549606, + 0.016246063634753227, + 0.04172808676958084, + 0.10500815510749817, + 0.02609799988567829, + 0.04390529915690422, + 0.004721651319414377, + 0.004952550400048494, + 0.004446176812052727, + 0.003849779488518834, + 0.006082897074520588, + 0.005965519230812788, + 0.004364455118775368, + 0.0033497519325464964, + 0.009188313968479633, + 0.004717005882412195, + 0.006867968011647463, + 0.004224900156259537, + 0.003437707433477044, + 0.011689056642353535, + 0.004054857883602381, + 0.0064789969474077225, + 0.003057958325371146, + 0.0038864072412252426, + 0.002593314042314887, + 0.00321853905916214, + 0.002934363903477788, + 0.0029521803371608257, + 0.0023791941348463297, + 0.0022291489876806736, + 0.012115729041397572, + 0.00886514037847519, + 0.0055965641513466835, + 0.004878359846770763, + 0.008825737982988358, + 0.008121301420032978, + 0.006639518775045872, + 0.008678370155394077, + 0.010930906049907207, + 0.009034929797053337, + 0.008414003998041153, + 0.012031658552587032, + 0.004785117693245411, + 0.008742718026041985, + 0.005685595329850912, + 0.004491639789193869, + 0.0051463390700519085, + 0.006046433933079243, + 0.006796424742788076, + 0.008818643167614937, + 0.005206089932471514, + 0.004972211085259914, + 0.00485375477001071, + 0.004268882796168327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.07006105035543442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866215959191322 + ], + "opposite_logprob_tokens": [ + -3.0486621856689453 + ], + "logprob_reward": -0.04866215959191322, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866215959191322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3383370339870453, + 0.2285727560520172, + 0.061105359345674515, + 0.14603528380393982, + 0.045097410678863525, + 0.03243597596883774, + 0.04387444257736206, + 0.041159067302942276, + 0.042968980967998505, + 0.020413706079125404 + ], + "credit_map": [ + 0.034988533705472946, + 0.015547986142337322, + 0.11133690923452377, + 0.01595924235880375, + 0.014261720702052116, + 0.08866579830646515, + 0.019957462325692177, + 0.03761938959360123, + 0.026039930060505867, + 0.011550644412636757, + 0.06832379847764969, + 0.00806015357375145, + 0.010049807839095592, + 0.07577023655176163, + 0.010562471114099026, + 0.018215734511613846, + 0.009848116897046566, + 0.00709878234192729, + 0.00802596565335989, + 0.004814037587493658, + 0.004442967474460602, + 0.008482546545565128, + 0.005409981589764357, + 0.012982971966266632, + 0.02461082674562931, + 0.008772931061685085, + 0.019758278504014015, + 0.00590034294873476, + 0.013971756212413311, + 0.04555682837963104, + 0.010565211996436119, + 0.016899120062589645, + 0.005809703841805458, + 0.004467331804335117, + 0.00547758350148797, + 0.002386505017057061, + 0.00758335692808032, + 0.0068292152136564255, + 0.004598571918904781, + 0.007945148274302483, + 0.004058428108692169, + 0.002351943403482437, + 0.005986360367387533, + 0.00340154767036438, + 0.004810413811355829, + 0.003309756051748991, + 0.004353816155344248, + 0.0041637117974460125, + 0.004616301041096449, + 0.0027906836476176977, + 0.0053773606196045876, + 0.003553226822987199, + 0.0060288975946605206, + 0.0045464979484677315, + 0.009863653220236301, + 0.007097827736288309, + 0.006852048449218273, + 0.003349906764924526, + 0.0034958827309310436, + 0.0029033434111624956, + 0.004664524458348751, + 0.009026547893881798, + 0.005580440163612366, + 0.005286377854645252, + 0.003725396003574133, + 0.004401524551212788, + 0.004469939507544041, + 0.005412033759057522, + 0.006212856154888868, + 0.00429714797064662, + 0.007962764240801334, + 0.00648732203990221, + 0.0021976714488118887, + 0.0014862550888210535, + 0.0039750500582158566, + 0.0016968152485787868, + 0.0016651037149131298, + 0.002770613180473447, + 0.0020147929899394512, + 0.0046074059791862965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.04866215959191322 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082993820309639 + ], + "opposite_logprob_tokens": [ + -4.511083126068115 + ], + "logprob_reward": -0.011082993820309639, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082993820309639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26538923382759094, + 0.04525930434465408, + 0.05729563534259796, + 0.1510491967201233, + 0.015997081995010376, + 0.03312050923705101, + 0.04853967949748039, + 0.1000894159078598, + 0.2581343948841095, + 0.025125540792942047 + ], + "credit_map": [ + 0.029818585142493248, + 0.012633963488042355, + 0.08882500231266022, + 0.012291494756937027, + 0.009851865470409393, + 0.07165008038282394, + 0.014558548107743263, + 0.02575972117483616, + 0.008132005110383034, + 0.003781392704695463, + 0.007853768765926361, + 0.0022827775683254004, + 0.0037389560602605343, + 0.01171119324862957, + 0.0026868439745157957, + 0.0050723678432404995, + 0.006871678400784731, + 0.004204506520181894, + 0.00887791533023119, + 0.004357113968580961, + 0.00473135057836771, + 0.00875604897737503, + 0.00547637278214097, + 0.014020653441548347, + 0.02691076323390007, + 0.008626369759440422, + 0.02228083647787571, + 0.005083281546831131, + 0.013326396234333515, + 0.050580695271492004, + 0.009337395429611206, + 0.014903469011187553, + 0.0019689269829541445, + 0.0022389166988432407, + 0.0019751745276153088, + 0.0013586300192400813, + 0.0021253072191029787, + 0.0019520489731803536, + 0.0020541627891361713, + 0.0023239150177687407, + 0.0027049481868743896, + 0.0023187417536973953, + 0.005324432160705328, + 0.0031440507154911757, + 0.00282000913284719, + 0.004754090216010809, + 0.005193660035729408, + 0.006860579829663038, + 0.006171459797769785, + 0.005204819608479738, + 0.004935915116220713, + 0.006649363320320845, + 0.005950422491878271, + 0.005468184128403664, + 0.00839719083160162, + 0.005762329325079918, + 0.014073542319238186, + 0.009814083576202393, + 0.012862840667366982, + 0.007696392480283976, + 0.00947339553385973, + 0.017567496746778488, + 0.014818676747381687, + 0.01378299668431282, + 0.03649124875664711, + 0.04090084135532379, + 0.06648006290197372, + 0.0211652759462595, + 0.020350690931081772, + 0.034872956573963165, + 0.020108893513679504, + 0.017764445394277573, + 0.003686424344778061, + 0.0022139642387628555, + 0.003917274530977011, + 0.0021490559447556734, + 0.001934521715156734, + 0.004538537934422493, + 0.002064872533082962, + 0.004620892461389303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.011082993820309639 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813108265399933 + ], + "opposite_logprob_tokens": [ + -1.406310796737671 + ], + "logprob_reward": -0.2813108265399933, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813108265399933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17747166752815247, + 0.23664812743663788, + 0.08505755662918091, + 0.10151926428079605, + 0.0677613615989685, + 0.06989452987909317, + 0.11046652495861053, + 0.06611744314432144, + 0.05253777280449867, + 0.03252570331096649 + ], + "credit_map": [ + 0.02108759805560112, + 0.010913028381764889, + 0.05184662342071533, + 0.011149219237267971, + 0.010863220319151878, + 0.04028820991516113, + 0.012743845582008362, + 0.018579915165901184, + 0.02636592462658882, + 0.012170621193945408, + 0.07239503413438797, + 0.008380015380680561, + 0.011628718115389347, + 0.07858539372682571, + 0.010073387995362282, + 0.017049025744199753, + 0.022064166143536568, + 0.004838242661207914, + 0.021167488768696785, + 0.004828230477869511, + 0.004760699346661568, + 0.01389282289892435, + 0.006579397711902857, + 0.00692651467397809, + 0.016470281407237053, + 0.006886888295412064, + 0.014778109267354012, + 0.004339567851275206, + 0.009128478355705738, + 0.030468834564089775, + 0.007788987830281258, + 0.01165811438113451, + 0.006698758341372013, + 0.005841502919793129, + 0.00841197557747364, + 0.004765142221003771, + 0.014201116748154163, + 0.00827718060463667, + 0.007080944254994392, + 0.012484736740589142, + 0.013779149390757084, + 0.006560157984495163, + 0.00933186337351799, + 0.004991965834051371, + 0.007487274240702391, + 0.014876787550747395, + 0.006414635106921196, + 0.006452692672610283, + 0.010747835971415043, + 0.019267024472355843, + 0.011347305029630661, + 0.0197502039372921, + 0.008810719475150108, + 0.015878036618232727, + 0.016124427318572998, + 0.00854097492992878, + 0.009240341372787952, + 0.005297667346894741, + 0.00904198456555605, + 0.005550796631723642, + 0.005009266547858715, + 0.007736284751445055, + 0.010886229574680328, + 0.013354870490729809, + 0.005502644460648298, + 0.006676677148789167, + 0.009558266028761864, + 0.007257882971316576, + 0.0053345924243330956, + 0.007466314360499382, + 0.005552559159696102, + 0.0051888395100831985, + 0.005222718231379986, + 0.0027118748985230923, + 0.0038600857369601727, + 0.003352999221533537, + 0.0032223144080489874, + 0.0047370088286697865, + 0.003448587842285633, + 0.005970112979412079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.2813108265399933 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle tracking shot steadily follows a young boy, approximately ten years old, as he crosses an outdoor basketball court from left to right. His face is a mask of concentration, his brow slightly furrowed, his eyes fixed on the ball he dribbles. He has short black hair, his cheeks slightly damp with sweat. Dressed in a sleeveless royal blue mesh sweatshirt and matching baggy shorts, his athletic physique poised for action. With practiced expertise, he keeps his body low to the ground as he rhythmically and decisively dribbles a worn orange basketball with his right hand. The ball clings to the ground, its textured surface catching the light with each powerful bounce on the cracked gray asphalt. The court is bathed in the warm golden glow of the afternoon sun, and the boy's shadow stretches and twists with his fluid movements. In the background, the faded white paint of the three-point line is clearly visible, along with a tall chain-link fence, through which the distant green trees gently blur, creating an atmosphere of serene, focused solitary devotion and youthful ambition.", + "seed": 8904, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_0/index_2475_seed_8904_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.17944328840821983, + "vlm_logprob": -0.17944328840821983, + "vlm_yes_no_margin": 2.425000011920929, + "vlm_token_credit": -0.17944328840821983, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5198690295219421 + ], + "opposite_logprob_tokens": [ + -1.144869089126587 + ], + "logprob_reward": -0.5198690295219421, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.5198690295219421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2464134693145752, + 0.387271523475647, + 0.042284317314624786, + 0.06909101456403732, + 0.04049122333526611, + 0.027724632993340492, + 0.0632513239979744, + 0.040534328669309616, + 0.06143489480018616, + 0.02150331810116768 + ], + "credit_map": [ + 0.01912635937333107, + 0.01951187662780285, + 0.024109138175845146, + 0.026116132736206055, + 0.04918689280748367, + 0.0305134616792202, + 0.03320387005805969, + 0.04464573413133621, + 0.013357856310904026, + 0.013574638403952122, + 0.03276686742901802, + 0.011490881443023682, + 0.037562329322099686, + 0.12051436305046082, + 0.04407747834920883, + 0.11392710357904434, + 0.004489611368626356, + 0.0038591136690229177, + 0.004893206991255283, + 0.0036777001805603504, + 0.006474937312304974, + 0.0051757944747805595, + 0.007527812849730253, + 0.006186138838529587, + 0.007239040918648243, + 0.008675731718540192, + 0.006762051954865456, + 0.006760620046406984, + 0.007546873763203621, + 0.011638668365776539, + 0.012416638433933258, + 0.008051400072872639, + 0.0047018821351230145, + 0.00380379194393754, + 0.003959865774959326, + 0.0036089117638766766, + 0.005700958892703056, + 0.009467991068959236, + 0.005015169735997915, + 0.0042326548136770725, + 0.0031239998061209917, + 0.0027458746917545795, + 0.0022031408734619617, + 0.0024354462511837482, + 0.004065767861902714, + 0.0059491111896932125, + 0.003797811223194003, + 0.0034034796990454197, + 0.0072531625628471375, + 0.006842971779406071, + 0.005166829098016024, + 0.006394568365067244, + 0.010381916537880898, + 0.0060454728081822395, + 0.01208638958632946, + 0.009080015122890472, + 0.0050209700129926205, + 0.0051793488673865795, + 0.004116696771234274, + 0.00305184512399137, + 0.006633815821260214, + 0.006487770937383175, + 0.004843881819397211, + 0.005199998617172241, + 0.007171198725700378, + 0.005591332446783781, + 0.006667401175945997, + 0.008461261168122292, + 0.008565571159124374, + 0.009994570165872574, + 0.008396359160542488, + 0.006587201729416847, + 0.002189062535762787, + 0.0018440786516293883, + 0.0016719436971470714, + 0.002650684444233775, + 0.0026222635060548782, + 0.003596871392801404, + 0.0040892199613153934, + 0.002839192748069763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.5198690295219421 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372737392783165 + ], + "opposite_logprob_tokens": [ + -3.4087274074554443 + ], + "logprob_reward": -0.03372737392783165, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372737392783165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15960347652435303, + 0.20235854387283325, + 0.0399458110332489, + 0.05586623772978783, + 0.03194817528128624, + 0.05014081299304962, + 0.12582021951675415, + 0.10944842547178268, + 0.11760204285383224, + 0.10726618766784668 + ], + "credit_map": [ + 0.012349172495305538, + 0.012365595437586308, + 0.01981492154300213, + 0.015736078843474388, + 0.025456182658672333, + 0.03725125268101692, + 0.01570981740951538, + 0.020920492708683014, + 0.011866504326462746, + 0.02217252366244793, + 0.018209613859653473, + 0.013751612044870853, + 0.024157989770174026, + 0.04845469817519188, + 0.0218091681599617, + 0.04193644970655441, + 0.006473882589489222, + 0.0034830986987799406, + 0.0049655684269964695, + 0.0034412506502121687, + 0.004809967707842588, + 0.003554336726665497, + 0.007773854304105043, + 0.005443855654448271, + 0.005377127788960934, + 0.006983527448028326, + 0.004941996186971664, + 0.004306466784328222, + 0.0060118879191577435, + 0.0108064915984869, + 0.007627060171216726, + 0.00981168169528246, + 0.0027979458682239056, + 0.002655482618138194, + 0.0029886423144489527, + 0.002847201656550169, + 0.004703742917627096, + 0.00601955596357584, + 0.004936904646456242, + 0.0049987006932497025, + 0.004913933575153351, + 0.007878059521317482, + 0.00494266115128994, + 0.007085874676704407, + 0.007466226816177368, + 0.005213942378759384, + 0.0068769920617341995, + 0.0057631260715425014, + 0.021763663738965988, + 0.01155939046293497, + 0.009733600541949272, + 0.00958574190735817, + 0.01390659250319004, + 0.016094662249088287, + 0.02090502716600895, + 0.022271549329161644, + 0.014471098780632019, + 0.021368904039263725, + 0.010379639454185963, + 0.00835401751101017, + 0.013124346733093262, + 0.015902014449238777, + 0.012691470794379711, + 0.01315693836659193, + 0.011701648123562336, + 0.009781023487448692, + 0.012936560437083244, + 0.01690671779215336, + 0.01854781061410904, + 0.014954447746276855, + 0.01851804368197918, + 0.014255814254283905, + 0.010620607063174248, + 0.01248607411980629, + 0.014794294722378254, + 0.01552649401128292, + 0.01010088063776493, + 0.013837170787155628, + 0.016531674191355705, + 0.013369005173444748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.03372737392783165 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899120450019836 + ], + "opposite_logprob_tokens": [ + -2.578991174697876 + ], + "logprob_reward": -0.07899120450019836, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899120450019836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35867786407470703, + 0.21876779198646545, + 0.11917279660701752, + 0.028504690155386925, + 0.015197494067251682, + 0.02036205306649208, + 0.11691126972436905, + 0.03913303092122078, + 0.058807484805583954, + 0.024465488269925117 + ], + "credit_map": [ + 0.020760200917720795, + 0.02253401279449463, + 0.05306743457913399, + 0.03002815879881382, + 0.05099552497267723, + 0.10686849057674408, + 0.035812195390462875, + 0.038611847907304764, + 0.009923849254846573, + 0.008720600046217442, + 0.0174434594810009, + 0.010213992558419704, + 0.02126447670161724, + 0.06427399814128876, + 0.02553112432360649, + 0.0613962784409523, + 0.009957513771951199, + 0.01106614526361227, + 0.017918100580573082, + 0.016785049811005592, + 0.01634395122528076, + 0.010364007204771042, + 0.018848523497581482, + 0.017889510840177536, + 0.0030827675946056843, + 0.002769400365650654, + 0.0030988205689936876, + 0.0031880356837064028, + 0.003252618247643113, + 0.004286321345716715, + 0.004930268041789532, + 0.0038964576087892056, + 0.0015473986277356744, + 0.0014129697810858488, + 0.0013068909756839275, + 0.0024379417300224304, + 0.0019157422939315438, + 0.0022233964409679174, + 0.0023470891173928976, + 0.002006066031754017, + 0.0017811174038797617, + 0.002235998399555683, + 0.001697865896858275, + 0.0023722287733107805, + 0.003327574348077178, + 0.0035537765361368656, + 0.0029602162539958954, + 0.0024332753382623196, + 0.0094909043982625, + 0.013122264295816422, + 0.00968543253839016, + 0.01112828217446804, + 0.024146568030118942, + 0.009454255923628807, + 0.023320676758885384, + 0.016562890261411667, + 0.004148864187300205, + 0.0047058165073394775, + 0.005984875839203596, + 0.0037379260174930096, + 0.0043355138041079044, + 0.005668691359460354, + 0.006030692253261805, + 0.004520650487393141, + 0.004821671638637781, + 0.0054026017896831036, + 0.005702519323676825, + 0.010107152163982391, + 0.006250341422855854, + 0.007023693528026342, + 0.011849398724734783, + 0.007650105748325586, + 0.002269087126478553, + 0.001881253207102418, + 0.0026199095882475376, + 0.003744665300473571, + 0.0024673936422914267, + 0.0034119784832000732, + 0.004802102223038673, + 0.0032690977677702904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07899120450019836 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543476186692715 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543476186692715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543476186692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1283990740776062, + 0.1690254658460617, + 0.026861630380153656, + 0.06155023351311684, + 0.151236429810524, + 0.08253685384988785, + 0.11278761178255081, + 0.08599197864532471, + 0.09734494239091873, + 0.08426576852798462 + ], + "credit_map": [ + 0.007896972820162773, + 0.006624688394367695, + 0.02091875486075878, + 0.009148103184998035, + 0.016766609624028206, + 0.04245058074593544, + 0.008322177454829216, + 0.016271190717816353, + 0.006541026756167412, + 0.007838049903512001, + 0.01442374475300312, + 0.006407618056982756, + 0.016887392848730087, + 0.052223313599824905, + 0.018204564228653908, + 0.04649975150823593, + 0.0029364570509642363, + 0.0019019179744645953, + 0.0028111860156059265, + 0.0029178906697779894, + 0.0038884300738573074, + 0.0029446433763951063, + 0.004209420178085566, + 0.005251685157418251, + 0.004957924131304026, + 0.006111887749284506, + 0.005779622122645378, + 0.005874205846339464, + 0.006533184554427862, + 0.010439787991344929, + 0.012730534188449383, + 0.00912308506667614, + 0.012049269862473011, + 0.009556535631418228, + 0.012091996148228645, + 0.017868733033537865, + 0.018937962129712105, + 0.023232916370034218, + 0.033599305897951126, + 0.023899713531136513, + 0.009374977089464664, + 0.007337593007832766, + 0.006152487359941006, + 0.005991285666823387, + 0.010798518545925617, + 0.019633889198303223, + 0.012109753675758839, + 0.011138346046209335, + 0.012015963904559612, + 0.011282929219305515, + 0.007809977978467941, + 0.010198081843554974, + 0.01416507363319397, + 0.011688142083585262, + 0.026436708867549896, + 0.01919073797762394, + 0.008133082650601864, + 0.008899541571736336, + 0.0066165076568722725, + 0.007100991904735565, + 0.014109854586422443, + 0.016933387145400047, + 0.013149253092706203, + 0.011049359105527401, + 0.009796042926609516, + 0.008297314867377281, + 0.009263155050575733, + 0.013625433668494225, + 0.014944447204470634, + 0.014115473255515099, + 0.015594652853906155, + 0.011708417907357216, + 0.008628287352621555, + 0.006686572451144457, + 0.005490254610776901, + 0.006718686781823635, + 0.008475304581224918, + 0.011641331017017365, + 0.022606581449508667, + 0.01401874516159296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.012543476186692715 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520853579044342 + ], + "opposite_logprob_tokens": [ + -1.5020853281021118 + ], + "logprob_reward": -0.2520853579044342, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520853579044342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12053464353084564, + 0.13537286221981049, + 0.0694514587521553, + 0.1484711766242981, + 0.05254725366830826, + 0.03040224500000477, + 0.19949586689472198, + 0.0918092131614685, + 0.10120659321546555, + 0.05070873349905014 + ], + "credit_map": [ + 0.012593850493431091, + 0.01060996949672699, + 0.018998073413968086, + 0.010251188650727272, + 0.016615813598036766, + 0.02565000206232071, + 0.012710364535450935, + 0.013105375692248344, + 0.01014585793018341, + 0.018142657354474068, + 0.01016993448138237, + 0.00791341345757246, + 0.01648801378905773, + 0.029042089357972145, + 0.015054729767143726, + 0.028416166082024574, + 0.009864955209195614, + 0.007431352976709604, + 0.01089426502585411, + 0.008353224955499172, + 0.008104842156171799, + 0.006117872428148985, + 0.010806393809616566, + 0.007878552190959454, + 0.01704772189259529, + 0.019237466156482697, + 0.012942259199917316, + 0.011344688944518566, + 0.01627400889992714, + 0.028499282896518707, + 0.02353339083492756, + 0.019592346623539925, + 0.0053144521079957485, + 0.004820383619517088, + 0.00370018626563251, + 0.0037085451185703278, + 0.009477186016738415, + 0.013266947120428085, + 0.005726357456296682, + 0.006533198524266481, + 0.0036800969392061234, + 0.0036641147453337908, + 0.00199224054813385, + 0.002520171692594886, + 0.0037805307656526566, + 0.0062532019801437855, + 0.004319885279983282, + 0.00419200211763382, + 0.016184547916054726, + 0.020604899153113365, + 0.013679333962500095, + 0.01831311173737049, + 0.029672985896468163, + 0.021200232207775116, + 0.04731819033622742, + 0.03252257779240608, + 0.008752601221203804, + 0.008563468232750893, + 0.009393766522407532, + 0.007865440100431442, + 0.018586650490760803, + 0.013069070875644684, + 0.013713505119085312, + 0.011864714324474335, + 0.010229489766061306, + 0.008059603162109852, + 0.010652091354131699, + 0.015259146690368652, + 0.01276327669620514, + 0.01552638504654169, + 0.01663411781191826, + 0.012082492001354694, + 0.005066512152552605, + 0.006388155277818441, + 0.0035274436231702566, + 0.004562122281640768, + 0.006558503024280071, + 0.011950359679758549, + 0.006584801711142063, + 0.006070834118872881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.2520853579044342 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle tracking shot steadily follows a young boy, approximately ten years old, as he crosses an outdoor basketball court from left to right. His face is a mask of concentration, his brow slightly furrowed, his eyes fixed on the ball he dribbles. He has short black hair, his cheeks slightly damp with sweat. Dressed in a sleeveless royal blue mesh sweatshirt and matching baggy shorts, his athletic physique poised for action. With practiced expertise, he keeps his body low to the ground as he rhythmically and decisively dribbles a worn orange basketball with his right hand. The ball clings to the ground, its textured surface catching the light with each powerful bounce on the cracked gray asphalt. The court is bathed in the warm golden glow of the afternoon sun, and the boy's shadow stretches and twists with his fluid movements. In the background, the faded white paint of the three-point line is clearly visible, along with a tall chain-link fence, through which the distant green trees gently blur, creating an atmosphere of serene, focused solitary devotion and youthful ambition.", + "seed": 8905, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_0/index_2475_seed_8905_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1621967814862728, + "vlm_logprob": -0.1621967814862728, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.1621967814862728, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.46779322624206543 + ], + "opposite_logprob_tokens": [ + -1.2177932262420654 + ], + "logprob_reward": -0.46779322624206543, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.46779322624206543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18460966646671295, + 0.03261743485927582, + 0.06584886461496353, + 0.015077860094606876, + 0.03551076725125313, + 0.040168117731809616, + 0.05852775648236275, + 0.04119519144296646, + 0.5155425071716309, + 0.010901893489062786 + ], + "credit_map": [ + 0.023504143580794334, + 0.012481323443353176, + 0.0500144325196743, + 0.015322537161409855, + 0.01258879341185093, + 0.011147204786539078, + 0.0257789995521307, + 0.033772218972444534, + 0.0032466358970850706, + 0.002531481208279729, + 0.008443242870271206, + 0.004474467132240534, + 0.0025820177979767323, + 0.0036558550782501698, + 0.003868835512548685, + 0.0038148972671478987, + 0.0068906317465007305, + 0.005360886454582214, + 0.008319661021232605, + 0.008946728892624378, + 0.0077630504965782166, + 0.006047722417861223, + 0.01573866792023182, + 0.0067815096117556095, + 0.0016665506409481168, + 0.0012944957707077265, + 0.002232834231108427, + 0.0022808839567005634, + 0.0017822975059971213, + 0.0024331086315214634, + 0.0013345994520932436, + 0.0020530875772237778, + 0.005185701418668032, + 0.0022556211333721876, + 0.005063998978585005, + 0.004359247628599405, + 0.004259780514985323, + 0.0061952946707606316, + 0.0041014282032847404, + 0.004089687019586563, + 0.0034064953215420246, + 0.002472477965056896, + 0.005661476403474808, + 0.0024496440310031176, + 0.004246214404702187, + 0.004505428485572338, + 0.008809627033770084, + 0.008616750128567219, + 0.004441127181053162, + 0.002454668516293168, + 0.013559317216277122, + 0.003562065539881587, + 0.0022330288775265217, + 0.003426601877436042, + 0.008697936311364174, + 0.02015300840139389, + 0.0035140926484018564, + 0.0045171775855124, + 0.0031937856692820787, + 0.004347329027950764, + 0.0047899349592626095, + 0.004233673680573702, + 0.00994989275932312, + 0.006649302784353495, + 0.02214691787958145, + 0.015334023162722588, + 0.01868721842765808, + 0.010470096953213215, + 0.13967907428741455, + 0.06606946140527725, + 0.16845457255840302, + 0.07470107823610306, + 0.0011179669527336955, + 0.0014958747196942568, + 0.0011951547348871827, + 0.0009746946743689477, + 0.0022651085164397955, + 0.0013993794564157724, + 0.0012476731790229678, + 0.0012060408480465412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.46779322624206543 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637278474867344 + ], + "opposite_logprob_tokens": [ + -3.6513726711273193 + ], + "logprob_reward": -0.02637278474867344, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637278474867344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2033240646123886, + 0.07823266834020615, + 0.025260016322135925, + 0.14843516051769257, + 0.020630883052945137, + 0.03520834445953369, + 0.02990393340587616, + 0.03872622549533844, + 0.40486499667167664, + 0.015413695946335793 + ], + "credit_map": [ + 0.02539636939764023, + 0.01926565170288086, + 0.03768739476799965, + 0.023068644106388092, + 0.021868862211704254, + 0.013034899719059467, + 0.03076527826488018, + 0.03223695978522301, + 0.007235844153910875, + 0.009834994561970234, + 0.01777533069252968, + 0.008508944883942604, + 0.00769464485347271, + 0.00557149900123477, + 0.009321280755102634, + 0.01229013316333294, + 0.002812253078445792, + 0.0031780540011823177, + 0.002970351604744792, + 0.004356897436082363, + 0.002742919372394681, + 0.0024353654589504004, + 0.004252573940902948, + 0.002511601196601987, + 0.014225711114704609, + 0.019085727632045746, + 0.01835421845316887, + 0.03352491930127144, + 0.013684221543371677, + 0.017406228929758072, + 0.017353791743516922, + 0.01480034552514553, + 0.0021051065996289253, + 0.001601831172592938, + 0.002738840878009796, + 0.002634697826579213, + 0.001580080483108759, + 0.003101701382547617, + 0.003924086224287748, + 0.002944537904113531, + 0.0032767232041805983, + 0.002963482169434428, + 0.004684293642640114, + 0.0034554318990558386, + 0.003922377247363329, + 0.003970419056713581, + 0.0062516117468476295, + 0.006684002000838518, + 0.001948509132489562, + 0.0036227963864803314, + 0.0031040608882904053, + 0.004261089954525232, + 0.003565635299310088, + 0.002924516564235091, + 0.005209595430642366, + 0.005267728585749865, + 0.0032640984281897545, + 0.006722721271216869, + 0.0030284449458122253, + 0.003317031543701887, + 0.0040903654880821705, + 0.004548027645796537, + 0.008286512456834316, + 0.0054690223187208176, + 0.017456023022532463, + 0.012366793118417263, + 0.015115000307559967, + 0.009135622531175613, + 0.11019907146692276, + 0.05308212339878082, + 0.12824122607707977, + 0.05926913022994995, + 0.0012338300002738833, + 0.002517960499972105, + 0.0017823103116825223, + 0.0013755876570940018, + 0.002076616045087576, + 0.0016083483351394534, + 0.0022070433478802443, + 0.002612000098451972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.02637278474867344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496106669306755 + ], + "opposite_logprob_tokens": [ + -2.9299609661102295 + ], + "logprob_reward": -0.05496106669306755, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496106669306755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15497393906116486, + 0.13477982580661774, + 0.10613539069890976, + 0.08876228332519531, + 0.034842416644096375, + 0.057492081075906754, + 0.05288596823811531, + 0.13086295127868652, + 0.21445876359939575, + 0.02480633743107319 + ], + "credit_map": [ + 0.020080551505088806, + 0.010236025787889957, + 0.043547432869672775, + 0.01379892136901617, + 0.009639573283493519, + 0.014911802485585213, + 0.01786046475172043, + 0.024899186566472054, + 0.011397643014788628, + 0.006024044007062912, + 0.04230382665991783, + 0.01864776946604252, + 0.01020970568060875, + 0.009688321501016617, + 0.016995998099446297, + 0.019512537866830826, + 0.011599432677030563, + 0.007752481382340193, + 0.015100694261491299, + 0.009547196328639984, + 0.012083441019058228, + 0.010892769321799278, + 0.024505624547600746, + 0.014653765596449375, + 0.011081756092607975, + 0.008924574591219425, + 0.012256978079676628, + 0.01376124657690525, + 0.012458776123821735, + 0.011477687396109104, + 0.010347709991037846, + 0.008453569374978542, + 0.0034799135755747557, + 0.0019401541212573647, + 0.00581889133900404, + 0.0034043327905237675, + 0.0033297704067081213, + 0.0046057249419391155, + 0.005276398733258247, + 0.006987233646214008, + 0.005798792000859976, + 0.005962430965155363, + 0.00694732740521431, + 0.0057010166347026825, + 0.007162011694163084, + 0.006543203257024288, + 0.011156666092574596, + 0.008220640011131763, + 0.004433103837072849, + 0.00573606975376606, + 0.00846056453883648, + 0.003511179471388459, + 0.0039045000448822975, + 0.004203075543045998, + 0.00834636203944683, + 0.014291118830442429, + 0.011041213758289814, + 0.010889590717852116, + 0.007521744817495346, + 0.0123305544257164, + 0.013768957927823067, + 0.014345334842801094, + 0.037494707852602005, + 0.023470863699913025, + 0.009555421769618988, + 0.007444176357239485, + 0.008238996379077435, + 0.004637077450752258, + 0.05658585578203201, + 0.027898050844669342, + 0.06839458644390106, + 0.0317046195268631, + 0.0015753131592646241, + 0.002606659196317196, + 0.0018226273823529482, + 0.002430496970191598, + 0.00502533745020628, + 0.0026708273217082024, + 0.004773112945258617, + 0.003901965683326125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.05496106669306755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785657748579979 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.009785657748579979, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785657748579979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023059172555804253, + 0.0074043963104486465, + 0.004550596233457327, + 0.007111939135938883, + 0.00328031531535089, + 0.0030685015954077244, + 0.015678515657782555, + 0.009307808242738247, + 0.917695939540863, + 0.008842833340168 + ], + "credit_map": [ + 0.002449613530188799, + 0.0015980664175003767, + 0.004003110807389021, + 0.0018755339551717043, + 0.0021460263524204493, + 0.0015887811314314604, + 0.00555866165086627, + 0.0038393796421587467, + 0.0009463187889195979, + 0.0006310370517894626, + 0.0009145815274678171, + 0.0006360486149787903, + 0.000826281204354018, + 0.001391784637235105, + 0.000967918720562011, + 0.0010904263472184539, + 0.0005796042387373745, + 0.00041445039096288383, + 0.0006256008637137711, + 0.0005146243493072689, + 0.0005314484005793929, + 0.0005504820728674531, + 0.0008623609319329262, + 0.00047202559653669596, + 0.0008830837323330343, + 0.0007852176786400378, + 0.0008966011228039861, + 0.0005580183351412416, + 0.0014947798335924745, + 0.0008530121413059533, + 0.0007002025959081948, + 0.0009410243364982307, + 0.0003062231407966465, + 0.00020765990484505892, + 0.0004831019905395806, + 0.00045763698290102184, + 0.0003772226336877793, + 0.0004124724364373833, + 0.0005709080141969025, + 0.0004650905029848218, + 0.0002963892475236207, + 0.0002917927340604365, + 0.00039874407229945064, + 0.0002740852069109678, + 0.0003442155139055103, + 0.000308266025967896, + 0.0006522354669868946, + 0.0005027734441682696, + 0.0012693241005763412, + 0.001260389108210802, + 0.002049630042165518, + 0.0010634611826390028, + 0.0016007836675271392, + 0.0020183215383440256, + 0.0028204673435539007, + 0.003596140304580331, + 0.0008418461657129228, + 0.0012742690742015839, + 0.0006478072609752417, + 0.0008283227798528969, + 0.0009319651289843023, + 0.0009297829819843173, + 0.0023348238319158554, + 0.0015189918922260404, + 0.03925924003124237, + 0.02615503780543804, + 0.03359593451023102, + 0.018290268257260323, + 0.251799076795578, + 0.1198706179857254, + 0.2970965802669525, + 0.13162921369075775, + 0.0006123470957390964, + 0.0010006636148318648, + 0.0004912059521302581, + 0.0005671524559147656, + 0.0012549256207421422, + 0.0012903730385005474, + 0.0019801175221800804, + 0.0016460487386211753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.009785657748579979 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy wearing a sleeveless royal blue mesh sweatshirt and matching baggy shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy dribbling a worn orange basketball with his right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on an outdoor basketball court with cracked gray asphalt and faded white three-point line paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track steadily from left to right following the boy's movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25207117199897766 + ], + "opposite_logprob_tokens": [ + -1.5020711421966553 + ], + "logprob_reward": -0.25207117199897766, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25207117199897766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028356041759252548, + 0.028147181496024132, + 0.016776839271187782, + 0.02442461997270584, + 0.017562249675393105, + 0.015455423854291439, + 0.055693838745355606, + 0.00976218469440937, + 0.7909806966781616, + 0.01284082978963852 + ], + "credit_map": [ + 0.0028201076202094555, + 0.0025511670392006636, + 0.0051401290111243725, + 0.0024548147339373827, + 0.0025793034583330154, + 0.0018932305974885821, + 0.006069356109946966, + 0.004847935400903225, + 0.002741187112405896, + 0.0016541068907827139, + 0.007631998974829912, + 0.003132656216621399, + 0.002196766436100006, + 0.0030294887255877256, + 0.0034836975391954184, + 0.00427728146314621, + 0.002749714767560363, + 0.0012324642157182097, + 0.002276258310303092, + 0.0020938559900969267, + 0.002521529560908675, + 0.0020103086717426777, + 0.002460353309288621, + 0.0014323561917990446, + 0.0023642517626285553, + 0.0022398647852241993, + 0.0032893288880586624, + 0.005288540851324797, + 0.002608392620459199, + 0.0021209637634456158, + 0.0036437714006751776, + 0.0028695084620267153, + 0.0018944829935207963, + 0.0009468342759646475, + 0.004892236553132534, + 0.0018121710745617747, + 0.0017529570031911135, + 0.0017791998106986284, + 0.002220981055870652, + 0.0022633871994912624, + 0.0015636460157111287, + 0.0017561757704243064, + 0.0018774181371554732, + 0.0020185380708426237, + 0.001824220409616828, + 0.0011179097928106785, + 0.003094046376645565, + 0.0022034707944840193, + 0.0038295784033834934, + 0.006639673840254545, + 0.011448286473751068, + 0.003991779405623674, + 0.004056009463965893, + 0.005601515527814627, + 0.007015878800302744, + 0.013111123815178871, + 0.0009922001045197248, + 0.0020713130943477154, + 0.0007890097913332283, + 0.001187409390695393, + 0.0009319593082182109, + 0.0009685580153018236, + 0.001654021325521171, + 0.0011677149450406432, + 0.0337313637137413, + 0.023098532110452652, + 0.028702978044748306, + 0.015603250823915005, + 0.21701638400554657, + 0.10334063321352005, + 0.256159245967865, + 0.11332838982343674, + 0.0011135105742141604, + 0.0023244288749992847, + 0.001454165205359459, + 0.0011763229267671704, + 0.0021784380078315735, + 0.0013558105565607548, + 0.0017839584033936262, + 0.001454196055419743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.25207117199897766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_1/index_tensor([5147, 5147])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_1/index_tensor([5147, 5147])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d1eed103d33ee7ecbdea0620b948a0b2a0740c48 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_1/index_tensor([5147, 5147])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The shot begins with a medium close-up, the camera smoothly circling a female athlete suspended in mid-air, her toned body captured in extreme slow motion. She wears a stylish, dark blue one-piece swimsuit that accentuates her powerful physique, scattered water droplets clinging to her skin like tiny suspended jewels. Her dark hair is pulled into a tight, functional bun, her eyes closed in concentration, her face a mask of calm focus as she executes a graceful, twisting dive. Her body twirls at a deliberate, almost balletic pace, one arm extended overhead, the other tucked in, creating a dynamic spiral as the muscles in her back and shoulders subtly ripple with controlled movement. As the camera continues its steady arc and simultaneously pulls back, the background gradually comes into focus, revealing the true immensity of the moss-covered Leshan Giant Buddha. The ancient stone statue, carved directly into the cliffside, fills the frame, its serene, weathered face overlooking the misty river below, dotted with vibrant green moss and dark, damp streaks from centuries of rainfall, depicting patterns across its massive form. Soft, diffused morning light filters through the thin mist, casting an ethereal glow over the scene and creating a muted palette of earthy greens, grays, and a sporting deep blue, establishing an atmosphere of profound serenity and awe—a breathtaking juxtaposition of fleeting moments of human perfection with the timeless, immortal grandeur of the ancient world.", + "seed": 7577, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_1/index_5147_seed_7577_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.020138928294181825, + "vlm_logprob": -0.020138928294181825, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.020138928294181825, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030726343393325806 + ], + "opposite_logprob_tokens": [ + -3.530726432800293 + ], + "logprob_reward": -0.030726343393325806, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.030726343393325806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08372005075216293, + 0.34451019763946533, + 0.0586128793656826, + 0.1383984088897705, + 0.035321373492479324, + 0.08073937147855759, + 0.0775720402598381, + 0.08570846915245056, + 0.059246715158224106, + 0.036170583218336105 + ], + "credit_map": [ + 0.010611874051392078, + 0.00809450726956129, + 0.00937559548765421, + 0.012054000049829483, + 0.009844995103776455, + 0.008221030235290527, + 0.01168570015579462, + 0.013832335360348225, + 0.03580551967024803, + 0.041354671120643616, + 0.06293409317731857, + 0.03572101145982742, + 0.04568571597337723, + 0.04925808310508728, + 0.04162686690688133, + 0.03212418779730797, + 0.00848261546343565, + 0.008037266321480274, + 0.006246344186365604, + 0.00725531717762351, + 0.00771078048273921, + 0.0069898818619549274, + 0.005527676548808813, + 0.008362989872694016, + 0.008289160206913948, + 0.026597831398248672, + 0.007062106393277645, + 0.010549728758633137, + 0.018577363342046738, + 0.030878618359565735, + 0.01613817922770977, + 0.02030540071427822, + 0.003273783950135112, + 0.002470879815518856, + 0.009858513250946999, + 0.006898312363773584, + 0.0032554573845118284, + 0.0032737255096435547, + 0.0036538499407470226, + 0.0026368480175733566, + 0.010823213495314121, + 0.00821607280522585, + 0.011095279827713966, + 0.01037587784230709, + 0.00900997780263424, + 0.005883355624973774, + 0.009928733110427856, + 0.01540684700012207, + 0.00814601220190525, + 0.010769574902951717, + 0.012812216766178608, + 0.008268372155725956, + 0.009461037814617157, + 0.007929148152470589, + 0.007789214141666889, + 0.01239645667374134, + 0.012072063982486725, + 0.01077182311564684, + 0.006274736486375332, + 0.010196026414632797, + 0.012709363363683224, + 0.007716395426541567, + 0.010796076618134975, + 0.015171974897384644, + 0.005462282802909613, + 0.009535697288811207, + 0.00713846692815423, + 0.005886514205485582, + 0.0071612512692809105, + 0.009794453158974648, + 0.006390315480530262, + 0.007877730764448643, + 0.003882385091856122, + 0.0040811882354319096, + 0.005321661476045847, + 0.004861227236688137, + 0.004564177244901657, + 0.004590657074004412, + 0.00484338216483593, + 0.004025899805128574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.030726343393325806 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430726483464241 + ], + "opposite_logprob_tokens": [ + -3.1680727005004883 + ], + "logprob_reward": -0.0430726483464241, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430726483464241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13718192279338837, + 0.2654978036880493, + 0.11335387080907822, + 0.09674686938524246, + 0.08655478060245514, + 0.05539075657725334, + 0.06378627568483353, + 0.04888952523469925, + 0.07389013469219208, + 0.05870804935693741 + ], + "credit_map": [ + 0.017652522772550583, + 0.012321330606937408, + 0.01983620598912239, + 0.015715856105089188, + 0.009578187018632889, + 0.027557533234357834, + 0.013182904571294785, + 0.021337386220693588, + 0.037137433886528015, + 0.060850221663713455, + 0.029227223247289658, + 0.019650936126708984, + 0.04178297147154808, + 0.037418581545352936, + 0.01920161582529545, + 0.020228829234838486, + 0.021094124764204025, + 0.017729466781020164, + 0.008804048411548138, + 0.009643912315368652, + 0.012657249346375465, + 0.013038651086390018, + 0.009118195623159409, + 0.02126823551952839, + 0.010222475044429302, + 0.0188397616147995, + 0.007447383366525173, + 0.010479379445314407, + 0.008187768049538136, + 0.022053776308894157, + 0.008045420050621033, + 0.01147091668099165, + 0.010355734266340733, + 0.006055936217308044, + 0.02408076822757721, + 0.017951451241970062, + 0.006428711581975222, + 0.007564104627817869, + 0.007329992018640041, + 0.006788085680454969, + 0.0068153138272464275, + 0.005933218635618687, + 0.006890552584081888, + 0.007116115652024746, + 0.007148104254156351, + 0.004757437855005264, + 0.006652496289461851, + 0.010077522136271, + 0.0057897488586604595, + 0.009566101245582104, + 0.008040258660912514, + 0.007573226932436228, + 0.007158696185797453, + 0.0074394759722054005, + 0.009230644442141056, + 0.008988128043711185, + 0.007349864579737186, + 0.005643699783831835, + 0.005283385049551725, + 0.004547645337879658, + 0.007331795524805784, + 0.0059776087291538715, + 0.0069076260551810265, + 0.0058479029685258865, + 0.005320671945810318, + 0.009546940214931965, + 0.014514664188027382, + 0.007088965270668268, + 0.008416673168540001, + 0.013865044340491295, + 0.00830160640180111, + 0.006835571490228176, + 0.006325351539999247, + 0.007052634842693806, + 0.009891689755022526, + 0.005001840181648731, + 0.009067157283425331, + 0.0074233608320355415, + 0.006587195210158825, + 0.007358821574598551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0430726483464241 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732873618602753 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006732873618602753, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732873618602753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09168445318937302, + 0.22946713864803314, + 0.042258940637111664, + 0.06828297674655914, + 0.1217912882566452, + 0.04247824475169182, + 0.0990457758307457, + 0.10132963955402374, + 0.07553227990865707, + 0.12812922894954681 + ], + "credit_map": [ + 0.011127902194857597, + 0.009056271985173225, + 0.010887940414249897, + 0.012136041186749935, + 0.011617075651884079, + 0.01040844526141882, + 0.014940804801881313, + 0.011509976349771023, + 0.02809797041118145, + 0.03920094668865204, + 0.032739993184804916, + 0.021900378167629242, + 0.0357060469686985, + 0.034559302031993866, + 0.017372528091073036, + 0.01988997496664524, + 0.006601407658308744, + 0.00495531689375639, + 0.0058869654312729836, + 0.004989296197891235, + 0.0043642716482281685, + 0.0037062771152704954, + 0.004190245643258095, + 0.007565160281956196, + 0.008675393648445606, + 0.01082936953753233, + 0.00863850861787796, + 0.009193254634737968, + 0.00488349748775363, + 0.007099914364516735, + 0.0059339129365980625, + 0.013029123656451702, + 0.021707182750105858, + 0.0073687853291630745, + 0.02866417169570923, + 0.021897895261645317, + 0.007429992314428091, + 0.011191456578671932, + 0.008252347819507122, + 0.015279453247785568, + 0.005069417413324118, + 0.005621050018817186, + 0.0067393602803349495, + 0.006396790500730276, + 0.002912309719249606, + 0.002623532433062792, + 0.004180360585451126, + 0.008935421705245972, + 0.011812112294137478, + 0.016003204509615898, + 0.01644844561815262, + 0.013578076846897602, + 0.006781233008950949, + 0.007330798543989658, + 0.013160033151507378, + 0.013931862078607082, + 0.012735147960484028, + 0.010167591273784637, + 0.009091283194720745, + 0.015475680120289326, + 0.015400273725390434, + 0.011678280308842659, + 0.013614836148917675, + 0.013166545890271664, + 0.006510536652058363, + 0.009090429171919823, + 0.01145312748849392, + 0.007409144192934036, + 0.008913996629416943, + 0.016486303880810738, + 0.007865388877689838, + 0.0078033581376075745, + 0.01003517396748066, + 0.013574821874499321, + 0.01625186577439308, + 0.01310036238282919, + 0.028277097269892693, + 0.015093917958438396, + 0.016405431553721428, + 0.015390564687550068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.006732873618602753 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253299880772829 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.01253299880772829, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253299880772829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1220564991235733, + 0.2475777417421341, + 0.08753939718008041, + 0.09404701739549637, + 0.10147465765476227, + 0.07755003869533539, + 0.08446194976568222, + 0.04103948548436165, + 0.06108352169394493, + 0.08316964656114578 + ], + "credit_map": [ + 0.014754748903214931, + 0.011170009151101112, + 0.015042047016322613, + 0.017975909635424614, + 0.013446342200040817, + 0.01860935427248478, + 0.016991110518574715, + 0.014066970907151699, + 0.02835324965417385, + 0.0349031537771225, + 0.041258491575717926, + 0.024030381813645363, + 0.03579163923859596, + 0.04119311645627022, + 0.0199908260256052, + 0.022056883201003075, + 0.012343755923211575, + 0.009178687818348408, + 0.008320951834321022, + 0.010760952718555927, + 0.007656110916286707, + 0.009536596946418285, + 0.008919259533286095, + 0.020823081955313683, + 0.011549757793545723, + 0.016113942489027977, + 0.0111547214910388, + 0.011318582110106945, + 0.006059195846319199, + 0.008896393701434135, + 0.0063502127304673195, + 0.022604214027523994, + 0.012964515946805477, + 0.005458187311887741, + 0.025187965482473373, + 0.019886912778019905, + 0.007153461221605539, + 0.009340481832623482, + 0.010090084746479988, + 0.011393052525818348, + 0.01067627128213644, + 0.012065083719789982, + 0.01540522649884224, + 0.010338996537029743, + 0.0056421575136482716, + 0.0046333856880664825, + 0.006785516161471605, + 0.012003403156995773, + 0.010035588406026363, + 0.012727782130241394, + 0.013688566163182259, + 0.012158014811575413, + 0.006602108012884855, + 0.006696781143546104, + 0.009495547972619534, + 0.013057558797299862, + 0.00582395913079381, + 0.005797605030238628, + 0.004404745530337095, + 0.005516643635928631, + 0.006074892356991768, + 0.003696166444569826, + 0.004875169601291418, + 0.0048503028228878975, + 0.005811349023133516, + 0.007766701281070709, + 0.00903162732720375, + 0.005033199675381184, + 0.00689650047570467, + 0.011580023914575577, + 0.006985497195273638, + 0.00797862745821476, + 0.007557555101811886, + 0.008671341463923454, + 0.012924056500196457, + 0.00933122355490923, + 0.01670829765498638, + 0.00819367729127407, + 0.009691031649708748, + 0.010092461481690407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.01253299880772829 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629777304828167 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007629777304828167, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629777304828167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1878071129322052, + 0.13870836794376373, + 0.03536500036716461, + 0.07498880475759506, + 0.16941715776920319, + 0.07754679024219513, + 0.09662533551454544, + 0.06670621782541275, + 0.09876648336648941, + 0.05406877025961876 + ], + "credit_map": [ + 0.023567719385027885, + 0.01713678427040577, + 0.03193407505750656, + 0.01724095828831196, + 0.015251708216965199, + 0.038922347128391266, + 0.01722228340804577, + 0.026531217619776726, + 0.01679684780538082, + 0.023837875574827194, + 0.019030937924981117, + 0.013736953027546406, + 0.02132475934922695, + 0.020588116720318794, + 0.010430843569338322, + 0.012962029315531254, + 0.005593339446932077, + 0.006462732795625925, + 0.00328571954742074, + 0.0043882448226213455, + 0.004515184089541435, + 0.004038598854094744, + 0.0033744967076927423, + 0.003706682240590453, + 0.008529694750905037, + 0.012255578301846981, + 0.010270926170051098, + 0.007960453629493713, + 0.005198642145842314, + 0.009193001314997673, + 0.005736363586038351, + 0.01584414206445217, + 0.028737260028719902, + 0.01041773147881031, + 0.04143695533275604, + 0.029637452214956284, + 0.011397712863981724, + 0.01638190448284149, + 0.009483572095632553, + 0.02192455157637596, + 0.007421258836984634, + 0.007765338756144047, + 0.00894468929618597, + 0.01330933254212141, + 0.007050937972962856, + 0.005388947669416666, + 0.014061514288187027, + 0.013604764826595783, + 0.009183945134282112, + 0.014703330583870411, + 0.013538418337702751, + 0.011794066987931728, + 0.013004121370613575, + 0.012216742150485516, + 0.008732079528272152, + 0.013452628627419472, + 0.007340582553297281, + 0.006396407261490822, + 0.005007932893931866, + 0.008918918669223785, + 0.011989856138825417, + 0.008138807490468025, + 0.009332455694675446, + 0.009581249207258224, + 0.009034158661961555, + 0.012230686843395233, + 0.010665972717106342, + 0.011284305714070797, + 0.0109686478972435, + 0.021241039037704468, + 0.01121024414896965, + 0.012131419964134693, + 0.0042121801525354385, + 0.00528763560578227, + 0.007310378365218639, + 0.005108247976750135, + 0.008930427022278309, + 0.007704805117100477, + 0.008427467197179794, + 0.007087625563144684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.007629777304828167 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with a medium close-up, the camera smoothly circling a female athlete suspended in mid-air, her toned body captured in extreme slow motion. She wears a stylish, dark blue one-piece swimsuit that accentuates her powerful physique, scattered water droplets clinging to her skin like tiny suspended jewels. Her dark hair is pulled into a tight, functional bun, her eyes closed in concentration, her face a mask of calm focus as she executes a graceful, twisting dive. Her body twirls at a deliberate, almost balletic pace, one arm extended overhead, the other tucked in, creating a dynamic spiral as the muscles in her back and shoulders subtly ripple with controlled movement. As the camera continues its steady arc and simultaneously pulls back, the background gradually comes into focus, revealing the true immensity of the moss-covered Leshan Giant Buddha. The ancient stone statue, carved directly into the cliffside, fills the frame, its serene, weathered face overlooking the misty river below, dotted with vibrant green moss and dark, damp streaks from centuries of rainfall, depicting patterns across its massive form. Soft, diffused morning light filters through the thin mist, casting an ethereal glow over the scene and creating a muted palette of earthy greens, grays, and a sporting deep blue, establishing an atmosphere of profound serenity and awe—a breathtaking juxtaposition of fleeting moments of human perfection with the timeless, immortal grandeur of the ancient world.", + "seed": 7578, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_1/index_5147_seed_7578_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.021501361951231958, + "vlm_logprob": -0.021501361951231958, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.021501361951231958, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027540598064661026 + ], + "opposite_logprob_tokens": [ + -3.652540683746338 + ], + "logprob_reward": -0.027540598064661026, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.027540598064661026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16974897682666779, + 0.5853915214538574, + 0.055284108966588974, + 0.026557980105280876, + 0.03113735280930996, + 0.05351007729768753, + 0.01666591316461563, + 0.020182443782687187, + 0.01764829084277153, + 0.02387329377233982 + ], + "credit_map": [ + 0.011672298423945904, + 0.03218097239732742, + 0.018143130466341972, + 0.023440362885594368, + 0.012225531041622162, + 0.04095173999667168, + 0.011718695983290672, + 0.019416267052292824, + 0.021231822669506073, + 0.04742972552776337, + 0.03402393311262131, + 0.06276488304138184, + 0.013162407092750072, + 0.17057543992996216, + 0.01228779274970293, + 0.22391556203365326, + 0.006248402874916792, + 0.006427046377211809, + 0.006911849603056908, + 0.008056504651904106, + 0.007690397556871176, + 0.00510034617036581, + 0.0069118705578148365, + 0.00793769396841526, + 0.0025822215247899294, + 0.003395443083718419, + 0.002061823382973671, + 0.0022410238161683083, + 0.004981659352779388, + 0.002643284387886524, + 0.003820389974862337, + 0.004832135513424873, + 0.003950470592826605, + 0.00330617418512702, + 0.0034607695415616035, + 0.00495902681723237, + 0.003581432858482003, + 0.002855825237929821, + 0.003677006810903549, + 0.005346646532416344, + 0.0063499510288238525, + 0.005110339727252722, + 0.011378432624042034, + 0.008029861375689507, + 0.00868308637291193, + 0.004938177298754454, + 0.00560581823810935, + 0.00341441435739398, + 0.0017909661401063204, + 0.0013213705969974399, + 0.0020493105985224247, + 0.0007972539751790464, + 0.005023787263780832, + 0.0027898037806153297, + 0.0009843562729656696, + 0.001909065991640091, + 0.0019111699657514691, + 0.0029795875307172537, + 0.0028533099684864283, + 0.0021699238568544388, + 0.002767035970464349, + 0.0029357378371059895, + 0.00227653537876904, + 0.0022891468834131956, + 0.0034402934834361076, + 0.0010330864461138844, + 0.002256192034110427, + 0.0014349357225000858, + 0.001603990443982184, + 0.004286210518330336, + 0.0011719550238922238, + 0.002421629149466753, + 0.0035291658714413643, + 0.0021054104436188936, + 0.005245723761618137, + 0.0026593636721372604, + 0.002876743208616972, + 0.0027336934581398964, + 0.002035119105130434, + 0.0026880777440965176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.027540598064661026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496377497911453 + ], + "opposite_logprob_tokens": [ + -2.9299638271331787 + ], + "logprob_reward": -0.05496377497911453, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496377497911453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2601723074913025, + 0.12190614640712738, + 0.07675772905349731, + 0.08129683136940002, + 0.1455468237400055, + 0.09999863803386688, + 0.0716715082526207, + 0.09544926881790161, + 0.02433793619275093, + 0.022862792015075684 + ], + "credit_map": [ + 0.01984674297273159, + 0.05336654931306839, + 0.02867949567735195, + 0.027084022760391235, + 0.017752178013324738, + 0.06669271737337112, + 0.017981812357902527, + 0.028768787160515785, + 0.007029603235423565, + 0.010762891732156277, + 0.0070321885868906975, + 0.013675075955688953, + 0.010740884579718113, + 0.028761541470885277, + 0.006179275922477245, + 0.03772468864917755, + 0.009511822834610939, + 0.008006730116903782, + 0.013512877747416496, + 0.012321067042648792, + 0.006614312529563904, + 0.007822243496775627, + 0.006441235542297363, + 0.01252743974328041, + 0.007493085227906704, + 0.009298237040638924, + 0.007927688769996166, + 0.009838337078690529, + 0.011251956224441528, + 0.00916278176009655, + 0.009545731358230114, + 0.016779012978076935, + 0.015349549241364002, + 0.017370600253343582, + 0.014726357534527779, + 0.01946929097175598, + 0.023419763892889023, + 0.014056358486413956, + 0.020088672637939453, + 0.02106623537838459, + 0.008924919180572033, + 0.008868751116096973, + 0.011409739963710308, + 0.015581673942506313, + 0.020296134054660797, + 0.011714053340256214, + 0.010344342328608036, + 0.012859027832746506, + 0.007973103784024715, + 0.006223377306014299, + 0.009495771490037441, + 0.005521140061318874, + 0.022345298901200294, + 0.010327814146876335, + 0.005220496561378241, + 0.004564505070447922, + 0.009759996086359024, + 0.019378576427698135, + 0.010702735744416714, + 0.009713668376207352, + 0.012557410635054111, + 0.01412416435778141, + 0.010096707381308079, + 0.009116008877754211, + 0.0032674812246114016, + 0.0020525138825178146, + 0.0031841848976910114, + 0.002759091090410948, + 0.00355441658757627, + 0.0037013720721006393, + 0.002583878580480814, + 0.0032349973917007446, + 0.0031196316704154015, + 0.002051793970167637, + 0.003654919331893325, + 0.0026370889972895384, + 0.002894860925152898, + 0.0024276156909763813, + 0.002064607571810484, + 0.004012274090200663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.05496377497911453 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622679229825735 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622679229825735, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622679229825735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07946262508630753, + 0.10131501406431198, + 0.10923053324222565, + 0.022810040041804314, + 0.053233273327350616, + 0.48914384841918945, + 0.03366531804203987, + 0.04032015800476074, + 0.03478175774216652, + 0.036037418991327286 + ], + "credit_map": [ + 0.00995909608900547, + 0.014064406976103783, + 0.007506007328629494, + 0.009530585259199142, + 0.008799838833510876, + 0.010510885156691074, + 0.0076361047104001045, + 0.011455696076154709, + 0.008831211365759373, + 0.018041111528873444, + 0.008129913359880447, + 0.011972296983003616, + 0.004011878278106451, + 0.018368156626820564, + 0.003949203994125128, + 0.02801123447716236, + 0.015340179204940796, + 0.0131986103951931, + 0.026258662343025208, + 0.0220810417085886, + 0.006614301819354296, + 0.008527351543307304, + 0.006200598552823067, + 0.011009789071977139, + 0.003365175100043416, + 0.002029271563515067, + 0.003949861507862806, + 0.003042872529476881, + 0.002183605683967471, + 0.0022509524133056402, + 0.00199917727150023, + 0.003989123739302158, + 0.006618248298764229, + 0.00573340430855751, + 0.006785770412534475, + 0.010069966316223145, + 0.0057885292917490005, + 0.004777414724230766, + 0.007094092201441526, + 0.0063658407889306545, + 0.06967633962631226, + 0.044872477650642395, + 0.10380979627370834, + 0.08377300947904587, + 0.0515010692179203, + 0.0446922592818737, + 0.059603992849588394, + 0.031214900314807892, + 0.0031160495709627867, + 0.0037548767868429422, + 0.005251848138868809, + 0.0034389542415738106, + 0.007008777465671301, + 0.003928594756871462, + 0.003225435269996524, + 0.003940785303711891, + 0.0042634685523808, + 0.0025768005289137363, + 0.006612709257751703, + 0.002449438441544771, + 0.0031840335577726364, + 0.010643528774380684, + 0.0042425598949193954, + 0.006347618065774441, + 0.005711550824344158, + 0.002976934192702174, + 0.0033970174845308065, + 0.0029052679892629385, + 0.005421141628175974, + 0.007118908688426018, + 0.0041627902537584305, + 0.0030881469137966633, + 0.004733598791062832, + 0.00298517569899559, + 0.007639897521585226, + 0.0037684799171984196, + 0.004242593888193369, + 0.0037795202806591988, + 0.004167939070612192, + 0.004720213357359171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.007622679229825735 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418730802834034 + ], + "opposite_logprob_tokens": [ + -4.264187335968018 + ], + "logprob_reward": -0.01418730802834034, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418730802834034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1278102993965149, + 0.22363685071468353, + 0.09735459089279175, + 0.055716197937726974, + 0.09965894371271133, + 0.13756075501441956, + 0.0632750391960144, + 0.07689081132411957, + 0.06531815230846405, + 0.05277831479907036 + ], + "credit_map": [ + 0.013064509257674217, + 0.021481582894921303, + 0.014351755380630493, + 0.01680697873234749, + 0.011951198801398277, + 0.0194893479347229, + 0.01246928982436657, + 0.018195651471614838, + 0.00955420732498169, + 0.018858110532164574, + 0.011649123392999172, + 0.02338736318051815, + 0.01136439386755228, + 0.06207878887653351, + 0.007555685471743345, + 0.07918919622898102, + 0.009976997040212154, + 0.013508607633411884, + 0.00911970715969801, + 0.012572829611599445, + 0.014590849168598652, + 0.01123275887221098, + 0.011966794729232788, + 0.014386049471795559, + 0.0075835296884179115, + 0.009394281543791294, + 0.00716695562005043, + 0.00731489947065711, + 0.007960916496813297, + 0.005759370978921652, + 0.005609309300780296, + 0.0049269357696175575, + 0.011693363077938557, + 0.014079478569328785, + 0.011871034279465675, + 0.020306626334786415, + 0.011604947969317436, + 0.009879473596811295, + 0.010128897614777088, + 0.010095130652189255, + 0.01919882744550705, + 0.012541444040834904, + 0.02861485257744789, + 0.021869175136089325, + 0.018204720690846443, + 0.013058154843747616, + 0.01347041130065918, + 0.010603171773254871, + 0.006555106956511736, + 0.00850133690983057, + 0.00974245648831129, + 0.006846834439784288, + 0.00780608132481575, + 0.009777226485311985, + 0.0058123026974499226, + 0.008233697153627872, + 0.007185489404946566, + 0.01524577010422945, + 0.010130503214895725, + 0.008182590827345848, + 0.012441769242286682, + 0.009165883995592594, + 0.007343398407101631, + 0.007195411249995232, + 0.011059505864977837, + 0.004676837008446455, + 0.006774937734007835, + 0.007492288015782833, + 0.007912245579063892, + 0.014603004790842533, + 0.0067644380033016205, + 0.0060349018312990665, + 0.009084498509764671, + 0.004610800184309483, + 0.01103246957063675, + 0.004610542673617601, + 0.007152763660997152, + 0.005977151449769735, + 0.004757077433168888, + 0.005553014110773802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.01418730802834034 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003192449454218149 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.003192449454218149, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003192449454218149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21324245631694794, + 0.32204747200012207, + 0.050104204565286636, + 0.0206584669649601, + 0.05666755139827728, + 0.1435507982969284, + 0.04295005276799202, + 0.03780402988195419, + 0.04442166909575462, + 0.06855328381061554 + ], + "credit_map": [ + 0.023153841495513916, + 0.04842277243733406, + 0.02641606330871582, + 0.015501005575060844, + 0.011703758500516415, + 0.05732223391532898, + 0.00845238845795393, + 0.022270401939749718, + 0.029565146192908287, + 0.06994594633579254, + 0.03392878547310829, + 0.03765535727143288, + 0.015809306874871254, + 0.05587933585047722, + 0.013536716811358929, + 0.0657268688082695, + 0.005696296226233244, + 0.0062477379105985165, + 0.006815649103373289, + 0.0065594925545156, + 0.006629536394029856, + 0.0044425236992537975, + 0.007877298630774021, + 0.005835673306137323, + 0.003156473161652684, + 0.002770242979750037, + 0.0024334650952368975, + 0.0024782675318419933, + 0.002336459467187524, + 0.002268025418743491, + 0.0019967462867498398, + 0.0032187867909669876, + 0.005924840923398733, + 0.00791353639215231, + 0.006681391038000584, + 0.007219590712338686, + 0.006897648796439171, + 0.005296131130307913, + 0.009690077975392342, + 0.007044332567602396, + 0.017120057716965675, + 0.012957226485013962, + 0.0276847742497921, + 0.02711775340139866, + 0.017820240929722786, + 0.01278550736606121, + 0.017088426277041435, + 0.010976805351674557, + 0.004474292043596506, + 0.002810237230733037, + 0.004608034621924162, + 0.003075367072597146, + 0.015466571785509586, + 0.0056528025306761265, + 0.002469440456479788, + 0.004393307957798243, + 0.005054288078099489, + 0.005762482061982155, + 0.004806358367204666, + 0.002804921008646488, + 0.003967964556068182, + 0.006243600510060787, + 0.003877905895933509, + 0.005286507308483124, + 0.006162829697132111, + 0.0036686821840703487, + 0.007589590270072222, + 0.004911895841360092, + 0.007259288337081671, + 0.005558354314416647, + 0.0047143748961389065, + 0.004556655418127775, + 0.01135945599526167, + 0.004680837970227003, + 0.011511019431054592, + 0.006877045147120953, + 0.007947608828544617, + 0.010090652853250504, + 0.00766321737319231, + 0.008423453196883202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.003192449454218149 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with a medium close-up, the camera smoothly circling a female athlete suspended in mid-air, her toned body captured in extreme slow motion. She wears a stylish, dark blue one-piece swimsuit that accentuates her powerful physique, scattered water droplets clinging to her skin like tiny suspended jewels. Her dark hair is pulled into a tight, functional bun, her eyes closed in concentration, her face a mask of calm focus as she executes a graceful, twisting dive. Her body twirls at a deliberate, almost balletic pace, one arm extended overhead, the other tucked in, creating a dynamic spiral as the muscles in her back and shoulders subtly ripple with controlled movement. As the camera continues its steady arc and simultaneously pulls back, the background gradually comes into focus, revealing the true immensity of the moss-covered Leshan Giant Buddha. The ancient stone statue, carved directly into the cliffside, fills the frame, its serene, weathered face overlooking the misty river below, dotted with vibrant green moss and dark, damp streaks from centuries of rainfall, depicting patterns across its massive form. Soft, diffused morning light filters through the thin mist, casting an ethereal glow over the scene and creating a muted palette of earthy greens, grays, and a sporting deep blue, establishing an atmosphere of profound serenity and awe—a breathtaking juxtaposition of fleeting moments of human perfection with the timeless, immortal grandeur of the ancient world.", + "seed": 7579, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_1/index_5147_seed_7579_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017996006086468695, + "vlm_logprob": -0.017996006086468695, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.017996006086468695, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03473788872361183 + ], + "opposite_logprob_tokens": [ + -3.4097378253936768 + ], + "logprob_reward": -0.03473788872361183, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03473788872361183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18755657970905304, + 0.04903389513492584, + 0.034712053835392, + 0.06119227409362793, + 0.05893538147211075, + 0.1510154753923416, + 0.10564770549535751, + 0.2564127743244171, + 0.06698036938905716, + 0.028513476252555847 + ], + "credit_map": [ + 0.033165160566568375, + 0.018961597234010696, + 0.027518564835190773, + 0.02417154051363468, + 0.02669394016265869, + 0.018141012638807297, + 0.025204358622431755, + 0.013700410723686218, + 0.009236769750714302, + 0.009781815111637115, + 0.005743104033172131, + 0.005831943359225988, + 0.005932120606303215, + 0.00389049481600523, + 0.004411752335727215, + 0.004205894190818071, + 0.004027020186185837, + 0.005843084771186113, + 0.006266082637012005, + 0.00367750390432775, + 0.004545531701296568, + 0.003918128088116646, + 0.003586391918361187, + 0.0028483106289058924, + 0.006490272469818592, + 0.016735922545194626, + 0.007307829800993204, + 0.007784631103277206, + 0.006033156532794237, + 0.006253001280128956, + 0.004367572255432606, + 0.006219884846359491, + 0.006756295915693045, + 0.011694634333252907, + 0.006502561736851931, + 0.006528441794216633, + 0.005734102800488472, + 0.008993717841804028, + 0.0037124494556337595, + 0.009013177827000618, + 0.041064076125621796, + 0.013285818509757519, + 0.05674266070127487, + 0.008769613690674305, + 0.00679257744923234, + 0.0077273850329220295, + 0.004269262310117483, + 0.012364081107079983, + 0.014549104496836662, + 0.012733237817883492, + 0.015820113942027092, + 0.009616090916097164, + 0.009999212808907032, + 0.014360593631863594, + 0.009829255752265453, + 0.0187400970607996, + 0.020941441878676414, + 0.02238195575773716, + 0.023490309715270996, + 0.06238362193107605, + 0.026345161721110344, + 0.04140012338757515, + 0.023952627554535866, + 0.03551753610372543, + 0.005345177371054888, + 0.008285865187644958, + 0.009874248877167702, + 0.006299786735326052, + 0.007926431484520435, + 0.00945767480880022, + 0.008396612480282784, + 0.011394571512937546, + 0.003499750280752778, + 0.002424949314445257, + 0.00538795767351985, + 0.0062020886689424515, + 0.0026352484710514545, + 0.0033122433815151453, + 0.0022612554021179676, + 0.0027899814303964376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.03473788872361183 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029813364148139954 + ], + "opposite_logprob_tokens": [ + -3.529813289642334 + ], + "logprob_reward": -0.029813364148139954, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029813364148139954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25976234674453735, + 0.20742100477218628, + 0.046603478491306305, + 0.040583811700344086, + 0.09727703779935837, + 0.06704786419868469, + 0.05168211832642555, + 0.11859636753797531, + 0.021485494449734688, + 0.0895405039191246 + ], + "credit_map": [ + 0.031725432723760605, + 0.02924172393977642, + 0.04729346185922623, + 0.026368172839283943, + 0.04497763141989708, + 0.034774936735630035, + 0.02378951944410801, + 0.021591484546661377, + 0.02478405274450779, + 0.02587377279996872, + 0.024106962606310844, + 0.011000323109328747, + 0.052314937114715576, + 0.04543173685669899, + 0.010602966882288456, + 0.013306254521012306, + 0.005619470961391926, + 0.008399510756134987, + 0.004544039722532034, + 0.0047769672237336636, + 0.006901290267705917, + 0.007485207170248032, + 0.005031547974795103, + 0.003845450934022665, + 0.004148960113525391, + 0.011088253930211067, + 0.004836120642721653, + 0.005309443920850754, + 0.005950849037617445, + 0.0030640834011137486, + 0.0029997010715305805, + 0.00318640423938632, + 0.010505725629627705, + 0.0233633853495121, + 0.010732060298323631, + 0.010715353302657604, + 0.010333634912967682, + 0.012615714222192764, + 0.008816394954919815, + 0.01019477192312479, + 0.007495625410228968, + 0.01351344957947731, + 0.010492585599422455, + 0.005435226485133171, + 0.008946727029979229, + 0.010414369404315948, + 0.003362240269780159, + 0.007387647405266762, + 0.009633688256144524, + 0.006299976725131273, + 0.007220607250928879, + 0.00542468624189496, + 0.004802418872714043, + 0.007333915214985609, + 0.004404938779771328, + 0.006561890710145235, + 0.01279203500598669, + 0.010468765161931515, + 0.013636087998747826, + 0.02419537492096424, + 0.013212582096457481, + 0.01707865484058857, + 0.010160754434764385, + 0.01705210842192173, + 0.002643981482833624, + 0.0022813454270362854, + 0.0034416248090565205, + 0.0029413786251097918, + 0.0018159199971705675, + 0.0024380385875701904, + 0.0021463739685714245, + 0.0037768324837088585, + 0.011019651778042316, + 0.007236326113343239, + 0.016267597675323486, + 0.016492309048771858, + 0.009529433213174343, + 0.017049161717295647, + 0.005351456347852945, + 0.0065945833921432495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.029813364148139954 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069436557590961 + ], + "opposite_logprob_tokens": [ + -4.511069297790527 + ], + "logprob_reward": -0.011069436557590961, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069436557590961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19949889183044434, + 0.23247499763965607, + 0.04394957423210144, + 0.10510337352752686, + 0.06659076362848282, + 0.13823918998241425, + 0.04773317277431488, + 0.052689988166093826, + 0.079507976770401, + 0.0342121347784996 + ], + "credit_map": [ + 0.01921696960926056, + 0.012425120919942856, + 0.044589847326278687, + 0.010187455452978611, + 0.04388590157032013, + 0.034752920269966125, + 0.013678930699825287, + 0.02076173759996891, + 0.027669567614793777, + 0.032733120024204254, + 0.03642936050891876, + 0.01373140886425972, + 0.04266359284520149, + 0.04848343878984451, + 0.013039425015449524, + 0.017725089564919472, + 0.005043085664510727, + 0.007025821600109339, + 0.007868938148021698, + 0.00401429645717144, + 0.006782464683055878, + 0.006534750573337078, + 0.003183783497661352, + 0.0034964338410645723, + 0.012326478026807308, + 0.028507187962532043, + 0.012562989257276058, + 0.012586207129061222, + 0.00883210264146328, + 0.012031815946102142, + 0.007190993521362543, + 0.01106560043990612, + 0.007037758361548185, + 0.015151528641581535, + 0.008533770218491554, + 0.007998150773346424, + 0.0073897140100598335, + 0.008094615302979946, + 0.004744576755911112, + 0.007640657015144825, + 0.03476959094405174, + 0.014101705513894558, + 0.048634760081768036, + 0.007911575958132744, + 0.009408029727637768, + 0.007411214057356119, + 0.0046218279749155045, + 0.011380486190319061, + 0.007859647274017334, + 0.006048498209565878, + 0.006664480082690716, + 0.00528974924236536, + 0.0037975304294377565, + 0.00603496702387929, + 0.003625523066148162, + 0.00841277465224266, + 0.006847666576504707, + 0.008722084574401379, + 0.008045829832553864, + 0.006810431834310293, + 0.004612595308572054, + 0.008161124773323536, + 0.00469967583194375, + 0.004790578968822956, + 0.0122429383918643, + 0.010785930790007114, + 0.010229009203612804, + 0.0062028635293245316, + 0.00614623399451375, + 0.017611142247915268, + 0.006935684010386467, + 0.009354175999760628, + 0.004544463939964771, + 0.0025863852351903915, + 0.004492805339396, + 0.004564703442156315, + 0.003877881681546569, + 0.005979356821626425, + 0.003369067097082734, + 0.004797470755875111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.011069436557590961 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624099031090736 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.007624099031090736, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624099031090736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20217856764793396, + 0.07955879718065262, + 0.06966681778430939, + 0.06516004353761673, + 0.07015765458345413, + 0.17394167184829712, + 0.06563564389944077, + 0.1724131852388382, + 0.052603986114263535, + 0.04868369549512863 + ], + "credit_map": [ + 0.021481629461050034, + 0.015143063850700855, + 0.03704017400741577, + 0.017802303656935692, + 0.03665309026837349, + 0.03535635024309158, + 0.021351780742406845, + 0.01735016144812107, + 0.009769674390554428, + 0.009928225539624691, + 0.012093410827219486, + 0.007852648384869099, + 0.015805356204509735, + 0.0077421534806489944, + 0.009183560498058796, + 0.007183759473264217, + 0.00652735261246562, + 0.008529814891517162, + 0.012503793463110924, + 0.009218058548867702, + 0.008832241408526897, + 0.009638363495469093, + 0.008997458033263683, + 0.005419730208814144, + 0.006747203413397074, + 0.012800300493836403, + 0.008038925938308239, + 0.007462882902473211, + 0.007814218290150166, + 0.0066468059085309505, + 0.007653302047401667, + 0.007996399886906147, + 0.009947256185114384, + 0.010787316597998142, + 0.012184116058051586, + 0.007479118648916483, + 0.0063108340837061405, + 0.008030707016587257, + 0.0070706382393836975, + 0.008347664959728718, + 0.048531848937273026, + 0.009697649627923965, + 0.06815205514431, + 0.010337363928556442, + 0.006745418068021536, + 0.010599127039313316, + 0.005229394417256117, + 0.014648810029029846, + 0.008241542614996433, + 0.010099113918840885, + 0.0092537309974432, + 0.00996929407119751, + 0.005794258322566748, + 0.009867002256214619, + 0.0052381521090865135, + 0.0071725426241755486, + 0.025529269129037857, + 0.022019008174538612, + 0.01638808287680149, + 0.02649875357747078, + 0.016322828829288483, + 0.028057817369699478, + 0.014851106330752373, + 0.022746317088603973, + 0.006971368566155434, + 0.004765479359775782, + 0.005975647363811731, + 0.00927021261304617, + 0.00541892321780324, + 0.0064353435300290585, + 0.006850235629826784, + 0.006916771177202463, + 0.0061156610026955605, + 0.004799978341907263, + 0.008585079573094845, + 0.008234485983848572, + 0.006850553210824728, + 0.005190418567508459, + 0.004214179236441851, + 0.00469333678483963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.007624099031090736 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673524197191 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.00673524197191, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673524197191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1643642634153366, + 0.16040123999118805, + 0.1693536937236786, + 0.08071760088205338, + 0.07146702706813812, + 0.03365715220570564, + 0.04281765967607498, + 0.08142577856779099, + 0.14126110076904297, + 0.05453448370099068 + ], + "credit_map": [ + 0.0373174287378788, + 0.016957826912403107, + 0.02675631456077099, + 0.015405627898871899, + 0.01979273557662964, + 0.01490670908242464, + 0.021253138780593872, + 0.011974473483860493, + 0.01794506050646305, + 0.01992565020918846, + 0.021648267284035683, + 0.008520271629095078, + 0.029572637751698494, + 0.040688466280698776, + 0.008925697766244411, + 0.013175180181860924, + 0.020806562155485153, + 0.03979744762182236, + 0.025379469618201256, + 0.01741226390004158, + 0.026538211852312088, + 0.01994990184903145, + 0.010557305067777634, + 0.008912511169910431, + 0.0103005925193429, + 0.025125982239842415, + 0.0093034403398633, + 0.008605034090578556, + 0.006791902240365744, + 0.007423526607453823, + 0.005558767821639776, + 0.007608351763337851, + 0.00683684041723609, + 0.014756584540009499, + 0.008790071122348309, + 0.009853427298367023, + 0.008906097151339054, + 0.008336890488862991, + 0.005924680270254612, + 0.008062432520091534, + 0.005444040056318045, + 0.003542554099112749, + 0.0077611361630260944, + 0.0027975672855973244, + 0.0031904256902635098, + 0.0040137507021427155, + 0.002140168333426118, + 0.0047675082460045815, + 0.004765205085277557, + 0.003405868774279952, + 0.006504375021904707, + 0.006845679599791765, + 0.005159575492143631, + 0.007429226767271757, + 0.0037328226026147604, + 0.0049749030731618404, + 0.010231485590338707, + 0.008158833719789982, + 0.009056654758751392, + 0.01716962270438671, + 0.007861380465328693, + 0.013443825766444206, + 0.005307044368237257, + 0.010196927934885025, + 0.016077667474746704, + 0.024140894412994385, + 0.01466933824121952, + 0.01739349775016308, + 0.01627378538250923, + 0.016832998022437096, + 0.019771486520767212, + 0.01610143482685089, + 0.00750090554356575, + 0.0031420195009559393, + 0.010872671380639076, + 0.005668368190526962, + 0.0031165408436208963, + 0.015857871621847153, + 0.004360329359769821, + 0.004015774000436068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.00673524197191 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with a medium close-up, the camera smoothly circling a female athlete suspended in mid-air, her toned body captured in extreme slow motion. She wears a stylish, dark blue one-piece swimsuit that accentuates her powerful physique, scattered water droplets clinging to her skin like tiny suspended jewels. Her dark hair is pulled into a tight, functional bun, her eyes closed in concentration, her face a mask of calm focus as she executes a graceful, twisting dive. Her body twirls at a deliberate, almost balletic pace, one arm extended overhead, the other tucked in, creating a dynamic spiral as the muscles in her back and shoulders subtly ripple with controlled movement. As the camera continues its steady arc and simultaneously pulls back, the background gradually comes into focus, revealing the true immensity of the moss-covered Leshan Giant Buddha. The ancient stone statue, carved directly into the cliffside, fills the frame, its serene, weathered face overlooking the misty river below, dotted with vibrant green moss and dark, damp streaks from centuries of rainfall, depicting patterns across its massive form. Soft, diffused morning light filters through the thin mist, casting an ethereal glow over the scene and creating a muted palette of earthy greens, grays, and a sporting deep blue, establishing an atmosphere of profound serenity and awe—a breathtaking juxtaposition of fleeting moments of human perfection with the timeless, immortal grandeur of the ancient world.", + "seed": 7580, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_1/index_5147_seed_7580_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02429750617593527, + "vlm_logprob": -0.02429750617593527, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02429750617593527, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04957035183906555 + ], + "opposite_logprob_tokens": [ + -3.049570322036743 + ], + "logprob_reward": -0.04957035183906555, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04957035183906555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15131177008152008, + 0.12593691051006317, + 0.04829540103673935, + 0.043026380240917206, + 0.28546762466430664, + 0.024037757888436317, + 0.10031111538410187, + 0.05217568576335907, + 0.07112863659858704, + 0.09830863028764725 + ], + "credit_map": [ + 0.020286530256271362, + 0.01924889348447323, + 0.017794432118535042, + 0.017863783985376358, + 0.027385259047150612, + 0.014747892506420612, + 0.021556347608566284, + 0.0124286450445652, + 0.012031091377139091, + 0.017153412103652954, + 0.012625071220099926, + 0.024023504927754402, + 0.015618668869137764, + 0.011601765640079975, + 0.018459808081388474, + 0.014423608779907227, + 0.010663928464055061, + 0.007430416066199541, + 0.008667071349918842, + 0.006470930762588978, + 0.003899041097611189, + 0.0037850874941796064, + 0.003356378758326173, + 0.004022554494440556, + 0.008733179420232773, + 0.0070822411216795444, + 0.004058001097291708, + 0.002617946360260248, + 0.0034679477103054523, + 0.007460226304829121, + 0.003097556997090578, + 0.006509284023195505, + 0.05379040539264679, + 0.023541174829006195, + 0.03017577715218067, + 0.03278729319572449, + 0.023733140900731087, + 0.0776352733373642, + 0.0157468318939209, + 0.028057772666215897, + 0.003349826205521822, + 0.004014776088297367, + 0.002907783491536975, + 0.002128986408933997, + 0.002252621343359351, + 0.004537852946668863, + 0.002207873621955514, + 0.002638040343299508, + 0.013895263895392418, + 0.009746596217155457, + 0.021482685580849648, + 0.011352453380823135, + 0.010838648304343224, + 0.010780400596559048, + 0.011654408648610115, + 0.010560662485659122, + 0.00929521769285202, + 0.0044784233905375, + 0.00652272580191493, + 0.005143255926668644, + 0.006763709709048271, + 0.00542265921831131, + 0.010031639598309994, + 0.004518060013651848, + 0.011010090820491314, + 0.004214917309582233, + 0.03004325181245804, + 0.003957177512347698, + 0.0022008258383721113, + 0.00574556365609169, + 0.002801705151796341, + 0.011155111715197563, + 0.012173131108283997, + 0.01616627722978592, + 0.010465853847563267, + 0.01534921582788229, + 0.011964226141571999, + 0.01353084109723568, + 0.009672337211668491, + 0.0089867627248168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.04957035183906555 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048669084906578064 + ], + "opposite_logprob_tokens": [ + -3.0486690998077393 + ], + "logprob_reward": -0.048669084906578064, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048669084906578064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29458147287368774, + 0.1534491926431656, + 0.06694278120994568, + 0.05174478888511658, + 0.09142225235700607, + 0.051395487040281296, + 0.07430492341518402, + 0.0502387210726738, + 0.04445110633969307, + 0.12146931141614914 + ], + "credit_map": [ + 0.037792615592479706, + 0.045375075191259384, + 0.035572439432144165, + 0.036222610622644424, + 0.04723881930112839, + 0.025040626525878906, + 0.03285694122314453, + 0.034482333809137344, + 0.033112093806266785, + 0.02383090928196907, + 0.0254980456084013, + 0.01117277704179287, + 0.02489612065255642, + 0.01064713392406702, + 0.010345587506890297, + 0.013946523889899254, + 0.015601500868797302, + 0.007973696105182171, + 0.008482780307531357, + 0.009163322858512402, + 0.007173983845859766, + 0.005996663589030504, + 0.005210035480558872, + 0.00734079722315073, + 0.009531580843031406, + 0.007987265475094318, + 0.0050089312717318535, + 0.004449172411113977, + 0.004906775895506144, + 0.007882692851126194, + 0.005039036739617586, + 0.006939335260540247, + 0.01519432757049799, + 0.008080330677330494, + 0.00973465759307146, + 0.011939094401896, + 0.009289892390370369, + 0.022003956139087677, + 0.005572800058871508, + 0.00960718933492899, + 0.007210555486381054, + 0.0059460606426000595, + 0.004909282550215721, + 0.004638800397515297, + 0.006871479097753763, + 0.011860696598887444, + 0.004082127939909697, + 0.005876484327018261, + 0.012509224005043507, + 0.007016488816589117, + 0.01760573498904705, + 0.008811084553599358, + 0.006061855237931013, + 0.005709927994757891, + 0.008740099146962166, + 0.007850509136915207, + 0.005786341615021229, + 0.007167789153754711, + 0.00619084108620882, + 0.0053132763132452965, + 0.009298369288444519, + 0.005539508070796728, + 0.007613496854901314, + 0.0033291007857769728, + 0.00689622201025486, + 0.003613047767430544, + 0.01595369167625904, + 0.0044553508050739765, + 0.0017583499429747462, + 0.003042507916688919, + 0.0018857745453715324, + 0.006846162490546703, + 0.014357835054397583, + 0.020499082282185555, + 0.016447773203253746, + 0.018743697553873062, + 0.010526576079428196, + 0.019672054797410965, + 0.008925488218665123, + 0.012296808883547783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.048669084906578064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731689441949129 + ], + "opposite_logprob_tokens": [ + -5.0067315101623535 + ], + "logprob_reward": -0.006731689441949129, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731689441949129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10548796504735947, + 0.07124148309230804, + 0.04011964797973633, + 0.057313915342092514, + 0.11548110097646713, + 0.27080681920051575, + 0.10343849658966064, + 0.08725741505622864, + 0.05240200459957123, + 0.09645113348960876 + ], + "credit_map": [ + 0.01640470325946808, + 0.010844459757208824, + 0.01620580069720745, + 0.012047049589455128, + 0.011704636737704277, + 0.007427738979458809, + 0.013846236281096935, + 0.017007343471050262, + 0.01876470074057579, + 0.013696257956326008, + 0.00810970552265644, + 0.005318389739841223, + 0.008512581698596478, + 0.004039505030959845, + 0.0031999661587178707, + 0.009600379504263401, + 0.008977334946393967, + 0.007142841350287199, + 0.0035783322528004646, + 0.004137580748647451, + 0.005357401445508003, + 0.003494205651804805, + 0.002530958503484726, + 0.004900995641946793, + 0.011887342669069767, + 0.008610407821834087, + 0.00899345614016056, + 0.005477301310747862, + 0.006562807131558657, + 0.0056770965456962585, + 0.006009387318044901, + 0.004096117336302996, + 0.019560638815164566, + 0.010741750709712505, + 0.012293309904634953, + 0.0156223950907588, + 0.010020282119512558, + 0.028454536572098732, + 0.007155398838222027, + 0.01163280289620161, + 0.04148077592253685, + 0.03531879186630249, + 0.029614554718136787, + 0.02488085627555847, + 0.030873728916049004, + 0.05831769108772278, + 0.02016492187976837, + 0.03015550971031189, + 0.010296326130628586, + 0.019121814519166946, + 0.01485742349177599, + 0.013430389575660229, + 0.011515442281961441, + 0.010393256321549416, + 0.011568889953196049, + 0.01225496269762516, + 0.015487764962017536, + 0.01517891138792038, + 0.01167834922671318, + 0.010275700129568577, + 0.008273573592305183, + 0.010761282406747341, + 0.009066597558557987, + 0.0065352399833500385, + 0.0069406600669026375, + 0.0036315165925771, + 0.020941270515322685, + 0.0034939537290483713, + 0.0017175964312627912, + 0.0038633025251328945, + 0.002320933388546109, + 0.009492775425314903, + 0.010607492178678513, + 0.014297026209533215, + 0.010318232700228691, + 0.016305383294820786, + 0.011556887067854404, + 0.013696173205971718, + 0.009940197691321373, + 0.009729747660458088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.006731689441949129 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732518319040537 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.006732518319040537, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732518319040537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10093925893306732, + 0.05521904304623604, + 0.06337720900774002, + 0.1621318906545639, + 0.25265204906463623, + 0.06165667623281479, + 0.15586821734905243, + 0.04309842362999916, + 0.0693587064743042, + 0.03569855913519859 + ], + "credit_map": [ + 0.012863864190876484, + 0.016793346032500267, + 0.010642696171998978, + 0.014169209636747837, + 0.012403026223182678, + 0.008901799097657204, + 0.016239719465374947, + 0.008925597183406353, + 0.010945259593427181, + 0.007601889781653881, + 0.009974835440516472, + 0.0041442024521529675, + 0.008000288158655167, + 0.0043102833442389965, + 0.004320233128964901, + 0.005922050215303898, + 0.012142714112997055, + 0.007604110985994339, + 0.005925521720200777, + 0.007523354142904282, + 0.008588465861976147, + 0.005987280979752541, + 0.0055800252594053745, + 0.01002573687583208, + 0.03358960896730423, + 0.02698368765413761, + 0.017914224416017532, + 0.013154317624866962, + 0.016890158876776695, + 0.02318456768989563, + 0.01497062761336565, + 0.015444701537489891, + 0.045064959675073624, + 0.02251449041068554, + 0.02725774049758911, + 0.029939336702227592, + 0.02035502903163433, + 0.06243586167693138, + 0.014114459045231342, + 0.03097015619277954, + 0.008288072422146797, + 0.008911213837563992, + 0.006424381397664547, + 0.004864586517214775, + 0.00863553211092949, + 0.013580414466559887, + 0.004893053788691759, + 0.0060594212263822556, + 0.01898888126015663, + 0.01930214650928974, + 0.02839813567698002, + 0.01747797802090645, + 0.02219817042350769, + 0.02161886729300022, + 0.013744884170591831, + 0.01413915678858757, + 0.006512568797916174, + 0.004447293933480978, + 0.005869592539966106, + 0.004464987199753523, + 0.00599333643913269, + 0.00521969236433506, + 0.0055954172275960445, + 0.004995533265173435, + 0.009652191773056984, + 0.004265066236257553, + 0.024432482197880745, + 0.0070882332511246204, + 0.003971879370510578, + 0.0068207369185984135, + 0.0031974157318472862, + 0.009930695407092571, + 0.004758444614708424, + 0.004237181507050991, + 0.005289103835821152, + 0.0035678476560860872, + 0.004376154392957687, + 0.004713874310255051, + 0.004491621628403664, + 0.004264330957084894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.006732518319040537 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female athlete wearing a dark blue one-piece swimsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete suspended in mid-air executing a twisting dive with her eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background revealed to be the moss-covered Leshan Giant Buddha carved into a cliffside?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium close-up and smoothly circle the athlete while pulling back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, diffused morning light creating an ethereal glow over the misty river?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978388637304306 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.00978388637304306, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978388637304306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2336294800043106, + 0.11418745666742325, + 0.11996520310640335, + 0.05290715768933296, + 0.05010192096233368, + 0.06215590611100197, + 0.09689696878194809, + 0.05952738970518112, + 0.09621021896600723, + 0.11441832780838013 + ], + "credit_map": [ + 0.03131486847996712, + 0.039960961788892746, + 0.025930136442184448, + 0.030543116852641106, + 0.026698684319853783, + 0.02280390076339245, + 0.028977835550904274, + 0.027399972081184387, + 0.029543420299887657, + 0.01906876266002655, + 0.014981789514422417, + 0.007447042502462864, + 0.01670045219361782, + 0.007281262893229723, + 0.006579182110726833, + 0.01258554682135582, + 0.02289178967475891, + 0.017881929874420166, + 0.010643198154866695, + 0.01434769295156002, + 0.012950033880770206, + 0.014681180007755756, + 0.010344171896576881, + 0.01622520573437214, + 0.010685701854526997, + 0.007769392337650061, + 0.007013728842139244, + 0.006519061513245106, + 0.006141828838735819, + 0.0058778938837349415, + 0.004426321014761925, + 0.004473229870200157, + 0.006553892511874437, + 0.004880157299339771, + 0.006236162036657333, + 0.007519701961427927, + 0.0055052656680345535, + 0.009614204987883568, + 0.003224276239052415, + 0.0065682572312653065, + 0.011065996252000332, + 0.009501657448709011, + 0.007021529600024223, + 0.005657856818288565, + 0.005256993696093559, + 0.012796907685697079, + 0.003855752991512418, + 0.006999214645475149, + 0.011926534585654736, + 0.014889356680214405, + 0.02033969759941101, + 0.012688552029430866, + 0.009973588399589062, + 0.008242566138505936, + 0.008588675409555435, + 0.010248000733554363, + 0.009440065361559391, + 0.006150338798761368, + 0.007586432620882988, + 0.007744999136775732, + 0.008809875696897507, + 0.00650807237252593, + 0.0068153077736496925, + 0.006472299341112375, + 0.009593617171049118, + 0.01085684821009636, + 0.03446182236075401, + 0.007275156676769257, + 0.005266357213258743, + 0.007426341529935598, + 0.0038669451605528593, + 0.017463132739067078, + 0.013288396410644054, + 0.019372500479221344, + 0.014675945043563843, + 0.019302209839224815, + 0.009933467023074627, + 0.017579510807991028, + 0.007953782565891743, + 0.012312517501413822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.00978388637304306 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_10/index_tensor([8945, 8945])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_10/index_tensor([8945, 8945])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..783dcdd160ae695985cd8009c3bcab2b57d5ed52 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_10/index_tensor([8945, 8945])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic wide-angle shot slowly tilts right to reveal a breathtaking alien sunset on a distant exoplanet. Two distinct suns sink majestically below the horizon: a massive crimson giant, its lower half already submerged, and a smaller, incredibly bright sapphire companion, hanging high above the skyline. They cast a surreal two-toned light across a thick, deep purple atmosphere filled with slowly drifting, iridescent clouds. Beneath them lies a vast ocean of liquid methane, its surface undulating not like water but in a slow, viscous, and oily manner, reflecting the twin suns' light as thin, shimmering streaks of red and blue. As the camera continues its steady roll, the dark surface ruptures, not with a violent splash but with a graceful, silent surge, revealing enormous manta-ray-like creatures. Their smooth, dark skin is adorned with an intricate, pulsating network of bioluminescent patterns in electric green, pink, and gold. They flew deliberately through the thick, hazy air, their massive wings catching the dying light before plunging back into the methane depths with slow, syrupy movements, leaving behind fading streaks of light that momentarily illuminated the ocean from within, creating an atmosphere of depth, silence, and awe-inspiring cosmic beauty.", + "seed": 12693, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_10/index_8945_seed_12693_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10783640900626779, + "vlm_logprob": -0.10783640900626779, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.10783640900626779, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06296131014823914 + ], + "opposite_logprob_tokens": [ + -2.8129613399505615 + ], + "logprob_reward": -0.06296131014823914, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06296131014823914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1103772446513176, + 0.030660999938845634, + 0.03116840124130249, + 0.1470845639705658, + 0.037590499967336655, + 0.04328266903758049, + 0.04227539151906967, + 0.3017573952674866, + 0.2074568271636963, + 0.04834606498479843 + ], + "credit_map": [ + 0.00887340772897005, + 0.013384580612182617, + 0.012110087089240551, + 0.012136323377490044, + 0.01957082375884056, + 0.012073196470737457, + 0.018244337290525436, + 0.01398447435349226, + 0.0029664025641977787, + 0.0026824488304555416, + 0.00441709952428937, + 0.0033500671852380037, + 0.0040547591634094715, + 0.004102063365280628, + 0.005554935894906521, + 0.003533220151439309, + 0.0029842485673725605, + 0.0030751305166631937, + 0.0045923213474452496, + 0.003997935447841883, + 0.003905151505023241, + 0.0033192133996635675, + 0.004656186792999506, + 0.0046382080763578415, + 0.011227740906178951, + 0.0205001812428236, + 0.015515661798417568, + 0.010739274322986603, + 0.038504790514707565, + 0.02154647745192051, + 0.010856233537197113, + 0.018194198608398438, + 0.0027680431958287954, + 0.004138469696044922, + 0.002521415241062641, + 0.007130050100386143, + 0.008013536222279072, + 0.006842257920652628, + 0.0031135310418903828, + 0.003063193056732416, + 0.0027817587833851576, + 0.005950118415057659, + 0.002785258926451206, + 0.008558092638850212, + 0.006896952632814646, + 0.005751227959990501, + 0.004185299389064312, + 0.006373957730829716, + 0.0022387499921023846, + 0.003014600370079279, + 0.005777840968221426, + 0.004703345708549023, + 0.009782922454178333, + 0.009499549865722656, + 0.003136435290798545, + 0.004121940117329359, + 0.022254277020692825, + 0.07295113056898117, + 0.017523031681776047, + 0.023879172280430794, + 0.06829334050416946, + 0.06749463826417923, + 0.013729463331401348, + 0.015632320195436478, + 0.006020892411470413, + 0.007685606833547354, + 0.00430306838825345, + 0.005888692103326321, + 0.08672880381345749, + 0.08138120919466019, + 0.009023820050060749, + 0.006424725987017155, + 0.0041739786975085735, + 0.009249648079276085, + 0.004811923950910568, + 0.005881190299987793, + 0.006638128310441971, + 0.005448903888463974, + 0.0061056665144860744, + 0.0060366177931427956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.06296131014823914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741635967046022 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741635967046022, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741635967046022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09071533381938934, + 0.04359373822808266, + 0.0345628559589386, + 0.06288070976734161, + 0.031044112518429756, + 0.05388876423239708, + 0.08040817081928253, + 0.10154067724943161, + 0.4212948679924011, + 0.08007080852985382 + ], + "credit_map": [ + 0.007102962117642164, + 0.011864634230732918, + 0.00809241458773613, + 0.010201874189078808, + 0.015969615429639816, + 0.014191710390150547, + 0.015309772454202175, + 0.007982336916029453, + 0.0035277476999908686, + 0.005364762619137764, + 0.004677645396441221, + 0.0037133435253053904, + 0.00638794107362628, + 0.010684981942176819, + 0.005241375882178545, + 0.003995937295258045, + 0.0027278524357825518, + 0.0032040050718933344, + 0.005606791935861111, + 0.004903699271380901, + 0.004884212743490934, + 0.004483177326619625, + 0.005133075173944235, + 0.0036200406029820442, + 0.0052054449915885925, + 0.008324995636940002, + 0.0073116496205329895, + 0.006599904969334602, + 0.015374431386590004, + 0.008396808989346027, + 0.005925162695348263, + 0.005742310546338558, + 0.001929388614371419, + 0.0034243506379425526, + 0.0021852583158761263, + 0.0040482040494680405, + 0.0057769534178078175, + 0.004843010101467371, + 0.004583518486469984, + 0.004253427963703871, + 0.003340486902743578, + 0.005209007766097784, + 0.004475124180316925, + 0.017490405589342117, + 0.006214268039911985, + 0.007270162459462881, + 0.004394888877868652, + 0.0054944176226854324, + 0.0046148644760251045, + 0.005202078726142645, + 0.00827016495168209, + 0.0070563629269599915, + 0.023615827783942223, + 0.0237533338367939, + 0.0033920982386916876, + 0.004503434523940086, + 0.012200643308460712, + 0.02792711928486824, + 0.008459736593067646, + 0.011033177375793457, + 0.014352812431752682, + 0.013290523551404476, + 0.005677138455212116, + 0.008599516935646534, + 0.010915309190750122, + 0.017683064565062523, + 0.0073655471205711365, + 0.009507790207862854, + 0.1830822229385376, + 0.16777077317237854, + 0.013611476868391037, + 0.011358664371073246, + 0.0075883702374994755, + 0.017679845914244652, + 0.005457297433167696, + 0.011451142840087414, + 0.011173443868756294, + 0.010208571329712868, + 0.009900965727865696, + 0.0066111707128584385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.006741635967046022 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298207193613052 + ], + "opposite_logprob_tokens": [ + -2.2379820346832275 + ], + "logprob_reward": -0.11298207193613052, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298207193613052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15522268414497375, + 0.06559392809867859, + 0.09308978915214539, + 0.03565656766295433, + 0.06666543334722519, + 0.07380108535289764, + 0.05453014746308327, + 0.03960046172142029, + 0.3716699182987213, + 0.04416995868086815 + ], + "credit_map": [ + 0.01283496618270874, + 0.0097457030788064, + 0.015172968618571758, + 0.0211825929582119, + 0.018489522859454155, + 0.023241518065333366, + 0.0313107855618, + 0.023244639858603477, + 0.005140839610248804, + 0.0061148530803620815, + 0.007702297996729612, + 0.010666935704648495, + 0.006378321908414364, + 0.011640569195151329, + 0.011540384031832218, + 0.006409726105630398, + 0.006881528068333864, + 0.008120039477944374, + 0.015109691768884659, + 0.010019064880907536, + 0.012051873840391636, + 0.014838332310318947, + 0.014061682857573032, + 0.012007584795355797, + 0.002056492492556572, + 0.0027817506343126297, + 0.003593399887904525, + 0.0032419597264379263, + 0.006210421212017536, + 0.004809477832168341, + 0.0046090236864984035, + 0.008354045450687408, + 0.0039239744655787945, + 0.005299542099237442, + 0.00956240389496088, + 0.007993378676474094, + 0.006704655941575766, + 0.009626413695514202, + 0.011986452154815197, + 0.011568615213036537, + 0.005936275236308575, + 0.006241529248654842, + 0.008736114017665386, + 0.02056409791111946, + 0.008170561864972115, + 0.00940758641809225, + 0.005809903144836426, + 0.008935024030506611, + 0.0031311940401792526, + 0.00684555247426033, + 0.0035235979594290257, + 0.003198070917278528, + 0.015980955213308334, + 0.015263646841049194, + 0.0026723863556981087, + 0.003914748318493366, + 0.005349225830286741, + 0.0064538163132965565, + 0.002738015493378043, + 0.0028369128704071045, + 0.008228966034948826, + 0.009423463605344296, + 0.0023071335162967443, + 0.0022629310842603445, + 0.008896064013242722, + 0.013777199201285839, + 0.006335016805678606, + 0.008571472950279713, + 0.16147582232952118, + 0.14990264177322388, + 0.014646963216364384, + 0.00806472823023796, + 0.004049580544233322, + 0.006156753748655319, + 0.0031839606817811728, + 0.005064085125923157, + 0.006284488830715418, + 0.006428327411413193, + 0.006401518825441599, + 0.006601243279874325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.11298207193613052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430665984749794 + ], + "opposite_logprob_tokens": [ + -3.1680665016174316 + ], + "logprob_reward": -0.0430665984749794, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430665984749794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05011415854096413, + 0.017118364572525024, + 0.017831914126873016, + 0.04508698731660843, + 0.01513917837291956, + 0.031792301684617996, + 0.0453142449259758, + 0.07541012763977051, + 0.5147300362586975, + 0.1874627023935318 + ], + "credit_map": [ + 0.003245252650231123, + 0.005166866350919008, + 0.004004961811006069, + 0.00534268282353878, + 0.005720134824514389, + 0.009981179609894753, + 0.011909425258636475, + 0.0047436547465622425, + 0.001456894213333726, + 0.0018567389342933893, + 0.0017653685063123703, + 0.002088246401399374, + 0.002139453310519457, + 0.0031646746210753918, + 0.0028846568893641233, + 0.0017623314633965492, + 0.0014913772465661168, + 0.0014683186309412122, + 0.0029750496614724398, + 0.0022990053985267878, + 0.002535364357754588, + 0.002370270201936364, + 0.002913817297667265, + 0.0017787113320082426, + 0.003846102161332965, + 0.006057280115783215, + 0.0055266376584768295, + 0.004144435282796621, + 0.011636528186500072, + 0.005607448518276215, + 0.0036776079796254635, + 0.004590947180986404, + 0.001319398172199726, + 0.0012862138682976365, + 0.0018896369729191065, + 0.0012228768318891525, + 0.002385210245847702, + 0.0030774343758821487, + 0.0015843978617340326, + 0.0023740099277347326, + 0.002574528567492962, + 0.005016855429857969, + 0.002997763454914093, + 0.0065069845877587795, + 0.005812847521156073, + 0.004143908154219389, + 0.0021500741131603718, + 0.0025893431156873703, + 0.0035237229894846678, + 0.0023771512787789106, + 0.005773283541202545, + 0.0048798187635838985, + 0.010994086973369122, + 0.011707663536071777, + 0.003325560363009572, + 0.0027329600416123867, + 0.007122371811419725, + 0.018995704129338264, + 0.006416434422135353, + 0.0069491127505898476, + 0.011282664723694324, + 0.011225658468902111, + 0.006908669602125883, + 0.006509509868919849, + 0.009936331771314144, + 0.02333029918372631, + 0.013673892244696617, + 0.013826364651322365, + 0.2145831286907196, + 0.20143568515777588, + 0.01893264427781105, + 0.019011683762073517, + 0.014689155854284763, + 0.0361998975276947, + 0.012219242751598358, + 0.023178838193416595, + 0.02561485581099987, + 0.024718649685382843, + 0.03283233195543289, + 0.018009725958108902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.0430665984749794 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31343042850494385 + ], + "opposite_logprob_tokens": [ + -1.3134304285049438 + ], + "logprob_reward": -0.31343042850494385, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31343042850494385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029894333332777023, + 0.020438460633158684, + 0.014087064191699028, + 0.021593572571873665, + 0.013246127404272556, + 0.022398239001631737, + 0.04690444841980934, + 0.06920231133699417, + 0.724868893623352, + 0.03736651688814163 + ], + "credit_map": [ + 0.0025123876985162497, + 0.0029050293378531933, + 0.003689347766339779, + 0.0030057330150157213, + 0.003442540531978011, + 0.0059703197330236435, + 0.0035797578748315573, + 0.004789217375218868, + 0.001531523186713457, + 0.00158959929831326, + 0.0024741441011428833, + 0.0035766789223998785, + 0.002521719317883253, + 0.0022981632500886917, + 0.004615940619260073, + 0.0018306936835870147, + 0.0014269778039306402, + 0.001316065201535821, + 0.0019323318265378475, + 0.0018996093422174454, + 0.0019864188507199287, + 0.0015456811524927616, + 0.002146640792489052, + 0.0018333382904529572, + 0.0015530608361586928, + 0.0021893575321882963, + 0.002158593386411667, + 0.0021026553586125374, + 0.003650146536529064, + 0.0029180110432207584, + 0.002644742140546441, + 0.0043770079500973225, + 0.0010031407000496984, + 0.0010532928863540292, + 0.0017798901535570621, + 0.0018050270155072212, + 0.0014739729231223464, + 0.0017132931388914585, + 0.002347334986552596, + 0.0020701750181615353, + 0.0014981612330302596, + 0.0020442854147404432, + 0.0016671348130330443, + 0.004975182935595512, + 0.0024300850927829742, + 0.0021773921325802803, + 0.002654682844877243, + 0.004951313138008118, + 0.003591318614780903, + 0.004880637861788273, + 0.004051300697028637, + 0.004621128086000681, + 0.007350385654717684, + 0.007894056849181652, + 0.005893773399293423, + 0.00862184539437294, + 0.006262041628360748, + 0.01538768783211708, + 0.005450475960969925, + 0.007586597464978695, + 0.008518888615071774, + 0.009687913581728935, + 0.0064740777015686035, + 0.009834629483520985, + 0.015881279483437538, + 0.028561150655150414, + 0.009417207911610603, + 0.013773450627923012, + 0.32173097133636475, + 0.29889318346977234, + 0.02077515237033367, + 0.015836484730243683, + 0.0036158799193799496, + 0.006943806074559689, + 0.003061810275539756, + 0.004426805302500725, + 0.0053141070529818535, + 0.005535175558179617, + 0.004721403121948242, + 0.003747527254745364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.31343042850494385 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide-angle shot slowly tilts right to reveal a breathtaking alien sunset on a distant exoplanet. Two distinct suns sink majestically below the horizon: a massive crimson giant, its lower half already submerged, and a smaller, incredibly bright sapphire companion, hanging high above the skyline. They cast a surreal two-toned light across a thick, deep purple atmosphere filled with slowly drifting, iridescent clouds. Beneath them lies a vast ocean of liquid methane, its surface undulating not like water but in a slow, viscous, and oily manner, reflecting the twin suns' light as thin, shimmering streaks of red and blue. As the camera continues its steady roll, the dark surface ruptures, not with a violent splash but with a graceful, silent surge, revealing enormous manta-ray-like creatures. Their smooth, dark skin is adorned with an intricate, pulsating network of bioluminescent patterns in electric green, pink, and gold. They flew deliberately through the thick, hazy air, their massive wings catching the dying light before plunging back into the methane depths with slow, syrupy movements, leaving behind fading streaks of light that momentarily illuminated the ocean from within, creating an atmosphere of depth, silence, and awe-inspiring cosmic beauty.", + "seed": 12694, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_10/index_8945_seed_12694_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08774936692789197, + "vlm_logprob": -0.08774936692789197, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.08774936692789197, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006775855552405119 + ], + "opposite_logprob_tokens": [ + -5.256775856018066 + ], + "logprob_reward": -0.006775855552405119, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006775855552405119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1543002724647522, + 0.177919402718544, + 0.0647941380739212, + 0.04783858731389046, + 0.04695415496826172, + 0.0471310093998909, + 0.05193660408258438, + 0.03669501841068268, + 0.32919543981552124, + 0.043235454708337784 + ], + "credit_map": [ + 0.013409937731921673, + 0.012546100653707981, + 0.013490492478013039, + 0.014960428699851036, + 0.02047877013683319, + 0.03657706081867218, + 0.01786009967327118, + 0.024977365508675575, + 0.0325993150472641, + 0.011771398596465588, + 0.00979529693722725, + 0.0075443945825099945, + 0.06433172523975372, + 0.028454262763261795, + 0.011425692588090897, + 0.011997313238680363, + 0.005659378133714199, + 0.005170939024537802, + 0.006798796355724335, + 0.007132498547434807, + 0.00864659994840622, + 0.012349379248917103, + 0.008531860075891018, + 0.010504676960408688, + 0.005010015796869993, + 0.006365800742059946, + 0.007739913649857044, + 0.004777788184583187, + 0.006421762518584728, + 0.006399229168891907, + 0.0044750915840268135, + 0.006648976821452379, + 0.006647262256592512, + 0.006552595179527998, + 0.00572771904990077, + 0.009430191479623318, + 0.005121409427374601, + 0.005625046323984861, + 0.004478744231164455, + 0.0033711851574480534, + 0.006294111255556345, + 0.011322385631501675, + 0.0037378091365098953, + 0.004185034427791834, + 0.004926424939185381, + 0.00769493542611599, + 0.004572662524878979, + 0.004397640470415354, + 0.0048470390029251575, + 0.00405273400247097, + 0.009042790159583092, + 0.006052418611943722, + 0.006940680090337992, + 0.006357922218739986, + 0.007624003104865551, + 0.007019012235105038, + 0.0029242492746561766, + 0.0053002187050879, + 0.004466090351343155, + 0.004605494439601898, + 0.0029620169661939144, + 0.003826702944934368, + 0.008028817363083363, + 0.004581425804644823, + 0.06475607305765152, + 0.0981469675898552, + 0.027573060244321823, + 0.0183599591255188, + 0.02475658990442753, + 0.06548570096492767, + 0.014223676174879074, + 0.015893392264842987, + 0.0032370248809456825, + 0.009666811674833298, + 0.003450057003647089, + 0.0056754956021904945, + 0.003210862400010228, + 0.008893366903066635, + 0.004840669222176075, + 0.004261160735040903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.006775855552405119 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260554142296314 + ], + "opposite_logprob_tokens": [ + -5.255260467529297 + ], + "logprob_reward": -0.005260554142296314, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260554142296314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0624190978705883, + 0.18666982650756836, + 0.014970536343753338, + 0.05051014572381973, + 0.09567945450544357, + 0.08081591874361038, + 0.08606185764074326, + 0.04303407296538353, + 0.2835768461227417, + 0.09626226872205734 + ], + "credit_map": [ + 0.006702199578285217, + 0.009058385156095028, + 0.006659490987658501, + 0.006715462543070316, + 0.007848769426345825, + 0.01135962549597025, + 0.006347183603793383, + 0.00772797642275691, + 0.035014115273952484, + 0.012500237673521042, + 0.010877293534576893, + 0.009160849265754223, + 0.06514425575733185, + 0.029033267870545387, + 0.012540332973003387, + 0.012399476021528244, + 0.001363683957606554, + 0.0017010241281241179, + 0.0014723935164511204, + 0.001809363136999309, + 0.0023540896363556385, + 0.0027489056810736656, + 0.001551518333144486, + 0.0019695586524903774, + 0.007620189338922501, + 0.006982280407100916, + 0.007904903031885624, + 0.005405096337199211, + 0.005102635361254215, + 0.0058430456556379795, + 0.005169702228158712, + 0.006482291966676712, + 0.007695490960031748, + 0.016201147809624672, + 0.011453285813331604, + 0.019353846088051796, + 0.00791320949792862, + 0.013385413214564323, + 0.00840363372117281, + 0.011273430660367012, + 0.006360440514981747, + 0.029282009229063988, + 0.004068925976753235, + 0.004301948007196188, + 0.009796110913157463, + 0.01865546405315399, + 0.0040662530809640884, + 0.004284770227968693, + 0.007300690282136202, + 0.00516426982358098, + 0.018749099224805832, + 0.012303289026021957, + 0.009670781902968884, + 0.009132114239037037, + 0.011968513019382954, + 0.01177309826016426, + 0.003288824576884508, + 0.006699653342366219, + 0.005703412927687168, + 0.0043328371830284595, + 0.005365455523133278, + 0.0069679319858551025, + 0.005867497995495796, + 0.00480846269056201, + 0.05452442169189453, + 0.08663313835859299, + 0.021847117692232132, + 0.015623994171619415, + 0.020937463268637657, + 0.05763939023017883, + 0.011826956644654274, + 0.014544365927577019, + 0.007362998556345701, + 0.022769829258322716, + 0.00674626138061285, + 0.009507851675152779, + 0.011137610301375389, + 0.018728716298937798, + 0.00963711179792881, + 0.010371891781687737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.005260554142296314 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16044354438781738 + ], + "opposite_logprob_tokens": [ + -1.9104435443878174 + ], + "logprob_reward": -0.16044354438781738, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16044354438781738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06091223284602165, + 0.2718464136123657, + 0.059808190912008286, + 0.0506172813475132, + 0.1997784972190857, + 0.01862996257841587, + 0.1482401341199875, + 0.060792937874794006, + 0.08142311125993729, + 0.04795127734541893 + ], + "credit_map": [ + 0.0076850587502121925, + 0.006966200657188892, + 0.005327554885298014, + 0.004831156227737665, + 0.012579556554555893, + 0.008643344976007938, + 0.007383616175502539, + 0.007495739962905645, + 0.04966340214014053, + 0.0177959855645895, + 0.015593679621815681, + 0.012562480755150318, + 0.0982595682144165, + 0.04319661855697632, + 0.017489051446318626, + 0.017285605892539024, + 0.0067129419185221195, + 0.004728435073047876, + 0.0042989482171833515, + 0.0061288997530937195, + 0.0072550238110125065, + 0.019809942692518234, + 0.004318796563893557, + 0.00655519962310791, + 0.007302796933799982, + 0.00849855039268732, + 0.0060074166394770145, + 0.006198251154273748, + 0.005854208488017321, + 0.006480108015239239, + 0.004146937746554613, + 0.006129013374447823, + 0.020412832498550415, + 0.02060806006193161, + 0.022521991282701492, + 0.04334040731191635, + 0.011866137385368347, + 0.026466066017746925, + 0.02698265202343464, + 0.02758035808801651, + 0.0020138814579695463, + 0.004601439461112022, + 0.0015927512431517243, + 0.0015459965215995908, + 0.0017448875587433577, + 0.003628648119047284, + 0.001563594676554203, + 0.0019387613283470273, + 0.009332549758255482, + 0.009168309159576893, + 0.01344943605363369, + 0.01069819089025259, + 0.04811122640967369, + 0.033214718103408813, + 0.01441606879234314, + 0.00984962284564972, + 0.004333158489316702, + 0.004586904775351286, + 0.014468357898294926, + 0.007754129823297262, + 0.005394198000431061, + 0.007738217711448669, + 0.00782585609704256, + 0.008692109026014805, + 0.015675092115998268, + 0.023854484781622887, + 0.0066556683741509914, + 0.005373698193579912, + 0.006121814250946045, + 0.015570539981126785, + 0.004068336915224791, + 0.004103471525013447, + 0.0032985026482492685, + 0.005074647255241871, + 0.00627535954117775, + 0.008776894770562649, + 0.005994180683046579, + 0.006561187095940113, + 0.005552676506340504, + 0.006417826283723116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 6 + ], + "selected_reward": -0.16044354438781738 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195887371897697 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195887371897697, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195887371897697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054890722036361694, + 0.06874506920576096, + 0.011462307535111904, + 0.022521136328577995, + 0.04522868990898132, + 0.11013367027044296, + 0.234236478805542, + 0.09376494586467743, + 0.2335580438375473, + 0.1254589557647705 + ], + "credit_map": [ + 0.0062742168083786964, + 0.00742854829877615, + 0.005692408420145512, + 0.006145088467746973, + 0.00608932925388217, + 0.0067023057490587234, + 0.0056423950009047985, + 0.010916423983871937, + 0.011403455398976803, + 0.004736935719847679, + 0.0033825812861323357, + 0.0026500336825847626, + 0.025842836126685143, + 0.011675993911921978, + 0.00431910390034318, + 0.004734118469059467, + 0.0008946562884375453, + 0.0009921896271407604, + 0.0011955185327678919, + 0.0013649824541062117, + 0.0013791072415187955, + 0.002810433739796281, + 0.00120560172945261, + 0.0016198171069845557, + 0.002916374010965228, + 0.004739284049719572, + 0.002453930675983429, + 0.0023234765976667404, + 0.002038429258391261, + 0.0028858736623078585, + 0.0021914581302553415, + 0.002972307847812772, + 0.007421508431434631, + 0.006733666639775038, + 0.00413340562954545, + 0.006159675773233175, + 0.004296920262277126, + 0.006602645386010408, + 0.005265255458652973, + 0.0046156104654073715, + 0.007819221355021, + 0.037096213549375534, + 0.007456615101546049, + 0.00808162335306406, + 0.010992120020091534, + 0.022774241864681244, + 0.008541724644601345, + 0.007371897343546152, + 0.016409307718276978, + 0.013282321393489838, + 0.03705452382564545, + 0.0271916463971138, + 0.04074787348508835, + 0.033828143030405045, + 0.03971225395798683, + 0.026010382920503616, + 0.006000361405313015, + 0.008461050689220428, + 0.015417291782796383, + 0.014753706753253937, + 0.00833781249821186, + 0.013857459649443626, + 0.012592392973601818, + 0.014344857074320316, + 0.045504190027713776, + 0.06986905634403229, + 0.01937238499522209, + 0.01219321321696043, + 0.017723774537444115, + 0.04671217501163483, + 0.010331088677048683, + 0.011852136813104153, + 0.009154066443443298, + 0.022466830909252167, + 0.010457690805196762, + 0.014006407931447029, + 0.02029593288898468, + 0.01857936568558216, + 0.016189564019441605, + 0.014309086836874485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.014195887371897697 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25207099318504333 + ], + "opposite_logprob_tokens": [ + -1.5020710229873657 + ], + "logprob_reward": -0.25207099318504333, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25207099318504333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030536869540810585, + 0.022472983226180077, + 0.03922257199883461, + 0.04423867538571358, + 0.04923252761363983, + 0.08103106170892715, + 0.05557280778884888, + 0.0314759835600853, + 0.5540733337402344, + 0.09214317053556442 + ], + "credit_map": [ + 0.0028228070586919785, + 0.002907834481447935, + 0.0024150218814611435, + 0.0022123854141682386, + 0.006452270783483982, + 0.003321603871881962, + 0.004291175864636898, + 0.006113763898611069, + 0.0027277374174445868, + 0.003593453438952565, + 0.003014473244547844, + 0.0036313303280621767, + 0.0025687336456030607, + 0.002412074711173773, + 0.0017170433420687914, + 0.002808135701343417, + 0.003402682952582836, + 0.0032956108916550875, + 0.0030317509081214666, + 0.003201036248356104, + 0.005564746912568808, + 0.011079368181526661, + 0.004211870953440666, + 0.005435500759631395, + 0.0038003819063305855, + 0.01140808966010809, + 0.005727943032979965, + 0.0035179255064576864, + 0.005453627556562424, + 0.006156937684863806, + 0.003702891757711768, + 0.004470874089747667, + 0.004808939062058926, + 0.009753516875207424, + 0.0036525570321828127, + 0.0059691304340958595, + 0.005470645613968372, + 0.008189127780497074, + 0.005975343286991119, + 0.00541326217353344, + 0.007501622661948204, + 0.028345884755253792, + 0.004420794080942869, + 0.005248285364359617, + 0.007817682810127735, + 0.017929010093212128, + 0.005053811706602573, + 0.00471396092325449, + 0.00473710335791111, + 0.0053518605418503284, + 0.008401475846767426, + 0.0070264809764921665, + 0.009670902043581009, + 0.0053343214094638824, + 0.009300635196268559, + 0.005750022828578949, + 0.004112479276955128, + 0.003609495935961604, + 0.005048614460974932, + 0.0026820756029337645, + 0.004550406709313393, + 0.00455073406919837, + 0.0034906521905213594, + 0.00343152042478323, + 0.10766618698835373, + 0.165645033121109, + 0.046667151153087616, + 0.03155393898487091, + 0.04141446202993393, + 0.11043208092451096, + 0.023698309436440468, + 0.026996126398444176, + 0.0061500160954892635, + 0.01541605219244957, + 0.009206852875649929, + 0.008875209838151932, + 0.013211138546466827, + 0.01458617765456438, + 0.01500058826059103, + 0.00969712808728218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.25207099318504333 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide-angle shot slowly tilts right to reveal a breathtaking alien sunset on a distant exoplanet. Two distinct suns sink majestically below the horizon: a massive crimson giant, its lower half already submerged, and a smaller, incredibly bright sapphire companion, hanging high above the skyline. They cast a surreal two-toned light across a thick, deep purple atmosphere filled with slowly drifting, iridescent clouds. Beneath them lies a vast ocean of liquid methane, its surface undulating not like water but in a slow, viscous, and oily manner, reflecting the twin suns' light as thin, shimmering streaks of red and blue. As the camera continues its steady roll, the dark surface ruptures, not with a violent splash but with a graceful, silent surge, revealing enormous manta-ray-like creatures. Their smooth, dark skin is adorned with an intricate, pulsating network of bioluminescent patterns in electric green, pink, and gold. They flew deliberately through the thick, hazy air, their massive wings catching the dying light before plunging back into the methane depths with slow, syrupy movements, leaving behind fading streaks of light that momentarily illuminated the ocean from within, creating an atmosphere of depth, silence, and awe-inspiring cosmic beauty.", + "seed": 12695, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_10/index_8945_seed_12695_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12679991805925966, + "vlm_logprob": -0.12679991805925966, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.12679991805925966, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2289263755083084 + ], + "opposite_logprob_tokens": [ + -1.603926420211792 + ], + "logprob_reward": -0.2289263755083084, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2289263755083084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1728031486272812, + 0.11252816021442413, + 0.07465493679046631, + 0.03441745415329933, + 0.21177875995635986, + 0.05216963589191437, + 0.03233412280678749, + 0.25564584136009216, + 0.03109687939286232, + 0.0225710216909647 + ], + "credit_map": [ + 0.011841816827654839, + 0.02436668798327446, + 0.02617916278541088, + 0.02358931675553322, + 0.021425412967801094, + 0.022226348519325256, + 0.017931152135133743, + 0.0252432469278574, + 0.007300134282559156, + 0.016950685530900955, + 0.012538468465209007, + 0.020607994869351387, + 0.011748402379453182, + 0.013327538035809994, + 0.013740409165620804, + 0.01631452888250351, + 0.005491539835929871, + 0.012669619172811508, + 0.007211291231215, + 0.013580868020653725, + 0.01098872721195221, + 0.006182339508086443, + 0.010447206906974316, + 0.008083348162472248, + 0.002958954544737935, + 0.006163680925965309, + 0.003968991804867983, + 0.0037522949278354645, + 0.0076112449169158936, + 0.0028527595568448305, + 0.0034001388121396303, + 0.00370938703417778, + 0.014351954683661461, + 0.0732109546661377, + 0.006144710350781679, + 0.012423817999660969, + 0.07469359785318375, + 0.007083853241056204, + 0.01625349186360836, + 0.007616384420543909, + 0.005126042291522026, + 0.015266582369804382, + 0.004568609409034252, + 0.0060049621388316154, + 0.01041592936962843, + 0.0029938211664557457, + 0.004536339547485113, + 0.0032573509961366653, + 0.003375031752511859, + 0.0038026422262191772, + 0.003866439452394843, + 0.005471385084092617, + 0.0031211983878165483, + 0.003850650042295456, + 0.004820220172405243, + 0.0040265521965920925, + 0.04902998358011246, + 0.06745094060897827, + 0.018739821389317513, + 0.020843153819441795, + 0.02676348388195038, + 0.02880827710032463, + 0.0217780452221632, + 0.022232141345739365, + 0.0025789502542465925, + 0.0035101447720080614, + 0.0029331878758966923, + 0.003625164506956935, + 0.004851543344557285, + 0.004313991405069828, + 0.004134499467909336, + 0.005149397533386946, + 0.0015662694349884987, + 0.001805718638934195, + 0.003645266406238079, + 0.003071550978347659, + 0.0030196106527000666, + 0.0040611750446259975, + 0.0027761836536228657, + 0.0026252467650920153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.2289263755083084 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764421047642827 + ], + "opposite_logprob_tokens": [ + -4.882644176483154 + ], + "logprob_reward": -0.00764421047642827, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764421047642827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12544861435890198, + 0.06428512930870056, + 0.023975089192390442, + 0.031118864193558693, + 0.22047287225723267, + 0.050269756466150284, + 0.03143348917365074, + 0.3350420296192169, + 0.08335869759321213, + 0.034595511853694916 + ], + "credit_map": [ + 0.008185687474906445, + 0.023738093674182892, + 0.012136907316744328, + 0.020815744996070862, + 0.018229890614748, + 0.013577784411609173, + 0.01541273482143879, + 0.01335175521671772, + 0.003722055349498987, + 0.00844422634691, + 0.0063421162776649, + 0.01593725197017193, + 0.008681504055857658, + 0.0070114475674927235, + 0.006506351754069328, + 0.007640172727406025, + 0.0021572010591626167, + 0.003938996698707342, + 0.0020100357942283154, + 0.003699330147355795, + 0.003844412276521325, + 0.0026516488287597895, + 0.002825475065037608, + 0.002847986528649926, + 0.002906681504100561, + 0.004654895979911089, + 0.00458204559981823, + 0.002695466624572873, + 0.00548955425620079, + 0.004409509245306253, + 0.003155599581077695, + 0.0032251100055873394, + 0.009741180576384068, + 0.08391547948122025, + 0.004642305430024862, + 0.00899666827172041, + 0.08538297563791275, + 0.007723797112703323, + 0.014063885435461998, + 0.006006564013659954, + 0.004974361043423414, + 0.008881757035851479, + 0.0046393065713346004, + 0.0064710583537817, + 0.00708865886554122, + 0.005033655557781458, + 0.007152725476771593, + 0.006028227508068085, + 0.004394328687340021, + 0.0032683841418474913, + 0.0034117389004677534, + 0.0035529115702956915, + 0.0030504523310810328, + 0.004104487597942352, + 0.003234146861359477, + 0.00641703512519598, + 0.06620641052722931, + 0.08733931928873062, + 0.024883432313799858, + 0.027271592989563942, + 0.033919330686330795, + 0.038695015013217926, + 0.02606276422739029, + 0.030664116144180298, + 0.007255519740283489, + 0.014673305675387383, + 0.009304706007242203, + 0.008644712157547474, + 0.00836983509361744, + 0.022519078105688095, + 0.006479692179709673, + 0.006111843977123499, + 0.0034463887568563223, + 0.0035267018247395754, + 0.0070424387231469154, + 0.004444400314241648, + 0.0026344414800405502, + 0.004774684552103281, + 0.0038909565191715956, + 0.004835498984903097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.00764421047642827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521876394748688 + ], + "opposite_logprob_tokens": [ + -1.5021876096725464 + ], + "logprob_reward": -0.2521876394748688, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521876394748688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07331600785255432, + 0.0651402622461319, + 0.03650006651878357, + 0.03881280496716499, + 0.20501075685024261, + 0.14928139746189117, + 0.02121894806623459, + 0.29684674739837646, + 0.023745596408843994, + 0.0901273861527443 + ], + "credit_map": [ + 0.004568461328744888, + 0.007441756781190634, + 0.010838055983185768, + 0.014944748021662235, + 0.010126636363565922, + 0.007222992833703756, + 0.011328326538205147, + 0.0068450444377958775, + 0.003738671774044633, + 0.006279090419411659, + 0.006727179978042841, + 0.016575351357460022, + 0.009148607961833477, + 0.005939464084804058, + 0.006522669456899166, + 0.010209226049482822, + 0.002293539000675082, + 0.00402660621330142, + 0.004926184192299843, + 0.006724098697304726, + 0.0038486190605908632, + 0.002430013846606016, + 0.006971240509301424, + 0.005279766861349344, + 0.0024818507954478264, + 0.009102691896259785, + 0.002708540065214038, + 0.0034594018943607807, + 0.010564117692410946, + 0.0029349566902965307, + 0.004586642142385244, + 0.0029746098443865776, + 0.007832624949514866, + 0.07821482419967651, + 0.0038043244276195765, + 0.0077707841992378235, + 0.08429417759180069, + 0.009602714329957962, + 0.00956166721880436, + 0.003929659724235535, + 0.009005637839436531, + 0.05286604166030884, + 0.008024980314075947, + 0.008852878585457802, + 0.05068092420697212, + 0.005761443171650171, + 0.0071121701039373875, + 0.0069773332215845585, + 0.002613042015582323, + 0.0027999726589769125, + 0.0021394945215433836, + 0.0035189823247492313, + 0.0019581003580242395, + 0.0022455304861068726, + 0.0029524138662964106, + 0.002991414861753583, + 0.05703918635845184, + 0.07713549584150314, + 0.024136900901794434, + 0.024593325331807137, + 0.031114183366298676, + 0.03428354486823082, + 0.023556368425488472, + 0.02498774603009224, + 0.002544354647397995, + 0.0035348511300981045, + 0.00293545750901103, + 0.003597433213144541, + 0.002827212680131197, + 0.003979112487286329, + 0.001997297862544656, + 0.002329880138859153, + 0.008149784058332443, + 0.006461302749812603, + 0.017764413729310036, + 0.015021713450551033, + 0.006566829979419708, + 0.011291459202766418, + 0.012862677685916424, + 0.0120092099532485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.2521876394748688 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191635608673096 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.018191635608673096, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018191635608673096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019849613308906555, + 0.015250195749104023, + 0.012151982635259628, + 0.01220884919166565, + 0.029678314924240112, + 0.029068846255540848, + 0.010000507347285748, + 0.7803331017494202, + 0.055712856352329254, + 0.03574566915631294 + ], + "credit_map": [ + 0.0014096980448812246, + 0.0026063539553433657, + 0.0028062460478395224, + 0.0030552572570741177, + 0.00281476229429245, + 0.00267849606461823, + 0.002140970202162862, + 0.0023378287442028522, + 0.0009573986753821373, + 0.0018378483364358544, + 0.0018977598519995809, + 0.0026351686101406813, + 0.0022259699180722237, + 0.0016847917577251792, + 0.0016105297254398465, + 0.0024007291067391634, + 0.0009429048513993621, + 0.0016788063803687692, + 0.001309718587435782, + 0.0020412395242601633, + 0.0016047192038968205, + 0.000956017232965678, + 0.0018486477201804519, + 0.0017699288437142968, + 0.001017001923173666, + 0.001817676704376936, + 0.001709947595372796, + 0.0012346248840913177, + 0.0018713847966864705, + 0.001203434425406158, + 0.0015520640881732106, + 0.0018027144251391292, + 0.0034387994091957808, + 0.007988875731825829, + 0.0014489158056676388, + 0.0018508537905290723, + 0.009299892000854015, + 0.0018522707978263497, + 0.002152962377294898, + 0.0016457440797239542, + 0.002138221636414528, + 0.009091789834201336, + 0.001885940320789814, + 0.0028144700918346643, + 0.008431735448539257, + 0.0011588196502998471, + 0.0019746851176023483, + 0.0015731840394437313, + 0.0011141553986817598, + 0.001037104520946741, + 0.001258918084204197, + 0.00169393012765795, + 0.0011363413650542498, + 0.001493140240199864, + 0.001132334116846323, + 0.0011345837265253067, + 0.14997802674770355, + 0.20245669782161713, + 0.06190834939479828, + 0.060630105435848236, + 0.08155139535665512, + 0.0905994102358818, + 0.0674646720290184, + 0.06574443727731705, + 0.003606368787586689, + 0.007384374272078276, + 0.007094688713550568, + 0.005839440040290356, + 0.006069970782846212, + 0.011426796205341816, + 0.007227394729852676, + 0.007063822355121374, + 0.003207681467756629, + 0.0022663529962301254, + 0.007008349988609552, + 0.005488375201821327, + 0.002969671506434679, + 0.00474688271060586, + 0.004890430718660355, + 0.005167924799025059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.018191635608673096 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704972922801971 + ], + "opposite_logprob_tokens": [ + -2.127049684524536 + ], + "logprob_reward": -0.12704972922801971, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704972922801971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016591597348451614, + 0.014976946637034416, + 0.009660976007580757, + 0.014402867294847965, + 0.0974307730793953, + 0.055377837270498276, + 0.012213361449539661, + 0.6843286156654358, + 0.0649978518486023, + 0.030019113793969154 + ], + "credit_map": [ + 0.0009797853417694569, + 0.0019215678330510855, + 0.0016647846205160022, + 0.003392290323972702, + 0.0030064343009144068, + 0.0020660357549786568, + 0.001745732268318534, + 0.0018149683019146323, + 0.0008860359084792435, + 0.0016347782220691442, + 0.0020893446635454893, + 0.003347605001181364, + 0.0015976133290678263, + 0.0014249348314478993, + 0.0015573544660583138, + 0.0024392809718847275, + 0.0007062371005304158, + 0.0016714413650333881, + 0.0010673736687749624, + 0.001247021253220737, + 0.0015846521127969027, + 0.0009573830757290125, + 0.0012617280008271337, + 0.001165140769444406, + 0.0007991068996489048, + 0.0032970402389764786, + 0.0007729710778221488, + 0.0013979411451146007, + 0.004291167017072439, + 0.0012239349307492375, + 0.0014995773090049624, + 0.0011211299570277333, + 0.004756818059831858, + 0.03781051188707352, + 0.001951362588442862, + 0.003041761927306652, + 0.039390336722135544, + 0.00419531250372529, + 0.004148822743445635, + 0.0021358567755669355, + 0.004066296387463808, + 0.018920060247182846, + 0.0031867800280451775, + 0.004320754203945398, + 0.01687941700220108, + 0.0018839600961655378, + 0.0034918105229735374, + 0.002628762973472476, + 0.0015193939907476306, + 0.0012331099715083838, + 0.0014978237450122833, + 0.0023771089036017656, + 0.0010654355864971876, + 0.0012174306903034449, + 0.0014581413706764579, + 0.0018449181225150824, + 0.13221272826194763, + 0.17788870632648468, + 0.053382374346256256, + 0.05287332832813263, + 0.07077106088399887, + 0.07854069769382477, + 0.05815224349498749, + 0.06050749495625496, + 0.006099238060414791, + 0.01156926155090332, + 0.006979375146329403, + 0.004282311070710421, + 0.006364044733345509, + 0.01946372538805008, + 0.005789882969111204, + 0.0044500152580440044, + 0.0027040629647672176, + 0.002568525495007634, + 0.004313559737056494, + 0.004548852331936359, + 0.003647290403023362, + 0.005246391519904137, + 0.003420719411224127, + 0.0035697161220014095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.12704972922801971 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide-angle shot slowly tilts right to reveal a breathtaking alien sunset on a distant exoplanet. Two distinct suns sink majestically below the horizon: a massive crimson giant, its lower half already submerged, and a smaller, incredibly bright sapphire companion, hanging high above the skyline. They cast a surreal two-toned light across a thick, deep purple atmosphere filled with slowly drifting, iridescent clouds. Beneath them lies a vast ocean of liquid methane, its surface undulating not like water but in a slow, viscous, and oily manner, reflecting the twin suns' light as thin, shimmering streaks of red and blue. As the camera continues its steady roll, the dark surface ruptures, not with a violent splash but with a graceful, silent surge, revealing enormous manta-ray-like creatures. Their smooth, dark skin is adorned with an intricate, pulsating network of bioluminescent patterns in electric green, pink, and gold. They flew deliberately through the thick, hazy air, their massive wings catching the dying light before plunging back into the methane depths with slow, syrupy movements, leaving behind fading streaks of light that momentarily illuminated the ocean from within, creating an atmosphere of depth, silence, and awe-inspiring cosmic beauty.", + "seed": 12696, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_10/index_8945_seed_12696_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10391466598957777, + "vlm_logprob": -0.10391466598957777, + "vlm_yes_no_margin": 2.924999976158142, + "vlm_token_credit": -0.10391466598957777, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.049835436046123505 + ], + "opposite_logprob_tokens": [ + -3.049835443496704 + ], + "logprob_reward": -0.049835436046123505, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.049835436046123505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1797860562801361, + 0.22716112434864044, + 0.1537037491798401, + 0.11738083511590958, + 0.14092852175235748, + 0.05704517662525177, + 0.03664422780275345, + 0.032938793301582336, + 0.026274263858795166, + 0.028137175366282463 + ], + "credit_map": [ + 0.026730496436357498, + 0.01359821017831564, + 0.011094117537140846, + 0.008580128662288189, + 0.02739480324089527, + 0.0707816556096077, + 0.007405686657875776, + 0.01420096680521965, + 0.027038637548685074, + 0.020723439753055573, + 0.014049060642719269, + 0.018740784376859665, + 0.08834853768348694, + 0.025760022923350334, + 0.01878104917705059, + 0.013719606213271618, + 0.012003251351416111, + 0.01822025515139103, + 0.009369177743792534, + 0.01088725309818983, + 0.04250463843345642, + 0.03468179330229759, + 0.015666073188185692, + 0.010371311567723751, + 0.008280669339001179, + 0.01792408712208271, + 0.008397867903113365, + 0.010352878831326962, + 0.013578126206994057, + 0.012737943790853024, + 0.02370852790772915, + 0.0224007461220026, + 0.0153400469571352, + 0.045390814542770386, + 0.013451575301587582, + 0.009120446629822254, + 0.021600762382149696, + 0.018696311861276627, + 0.009090696461498737, + 0.00823788158595562, + 0.0041022892110049725, + 0.007669134996831417, + 0.003341481788083911, + 0.002068209694698453, + 0.021663324907422066, + 0.013279497623443604, + 0.002551593352109194, + 0.002369648776948452, + 0.009122942574322224, + 0.004649389535188675, + 0.002476438879966736, + 0.002355876611545682, + 0.004575212486088276, + 0.0036022283602505922, + 0.005701984744518995, + 0.004160161130130291, + 0.003280418924987316, + 0.006730298511683941, + 0.0028656525537371635, + 0.0033627694938331842, + 0.004327729810029268, + 0.004524705931544304, + 0.004531996790319681, + 0.0033152287360280752, + 0.0021644385997205973, + 0.0027698231860995293, + 0.002913019387051463, + 0.003186182351782918, + 0.0039358483627438545, + 0.0022818073630332947, + 0.003527300665155053, + 0.0054958476684987545, + 0.0026267634239047766, + 0.0035498966462910175, + 0.0029961122199892998, + 0.004465645179152489, + 0.004216948989778757, + 0.003551230765879154, + 0.003294559894129634, + 0.0034360208082944155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.049835436046123505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864903163164854 + ], + "opposite_logprob_tokens": [ + -4.758648872375488 + ], + "logprob_reward": -0.00864903163164854, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864903163164854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17883838713169098, + 0.0822819173336029, + 0.037241131067276, + 0.19603808224201202, + 0.07236999273300171, + 0.03687158599495888, + 0.04008173570036888, + 0.1521940380334854, + 0.0669274851679802, + 0.1371556520462036 + ], + "credit_map": [ + 0.03886387124657631, + 0.009721505455672741, + 0.00850033201277256, + 0.00706880446523428, + 0.026628240942955017, + 0.07024680823087692, + 0.006543563213199377, + 0.01126524806022644, + 0.008530927821993828, + 0.007715329062193632, + 0.007389084901660681, + 0.007831612601876259, + 0.026261968538165092, + 0.008239087648689747, + 0.008551647886633873, + 0.007762257009744644, + 0.0027826016303151846, + 0.003439840395003557, + 0.002996661001816392, + 0.004011665936559439, + 0.007146642543375492, + 0.005969835910946131, + 0.005119289271533489, + 0.005774593446403742, + 0.01574654132127762, + 0.01558875571936369, + 0.01682818867266178, + 0.01899203471839428, + 0.032172054052352905, + 0.0214389618486166, + 0.040098994970321655, + 0.03517255187034607, + 0.006469222716987133, + 0.018869588151574135, + 0.006407576147466898, + 0.005481151398271322, + 0.013790170662105083, + 0.012977620586752892, + 0.004901306238025427, + 0.003473354270681739, + 0.004160855896770954, + 0.005508327856659889, + 0.00283607910387218, + 0.002557294676080346, + 0.00759196188300848, + 0.006384854204952717, + 0.004191149491816759, + 0.0036410631146281958, + 0.008216318674385548, + 0.0038418665062636137, + 0.003189789829775691, + 0.0033132589887827635, + 0.005184750538319349, + 0.005329374689608812, + 0.005747695919126272, + 0.005258679389953613, + 0.01796775870025158, + 0.05085209757089615, + 0.00787932425737381, + 0.008004419505596161, + 0.016844362020492554, + 0.035029854625463486, + 0.008600946515798569, + 0.007015282288193703, + 0.007984913885593414, + 0.009482229128479958, + 0.008088720962405205, + 0.007829884067177773, + 0.008600015193223953, + 0.010262997820973396, + 0.006089380476623774, + 0.008589345030486584, + 0.01287122257053852, + 0.01503431610763073, + 0.014806953258812428, + 0.01861470565199852, + 0.020274324342608452, + 0.017400765791535378, + 0.020773816853761673, + 0.01737953908741474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.00864903163164854 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270969808101654 + ], + "opposite_logprob_tokens": [ + -2.1270968914031982 + ], + "logprob_reward": -0.1270969808101654, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270969808101654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39267435669898987, + 0.06049829721450806, + 0.09333427995443344, + 0.05554986372590065, + 0.032470542937517166, + 0.07588958740234375, + 0.029974093660712242, + 0.03731321170926094, + 0.042778562754392624, + 0.17951714992523193 + ], + "credit_map": [ + 0.07525915652513504, + 0.02404777705669403, + 0.014212604612112045, + 0.01151918526738882, + 0.06579379737377167, + 0.16981437802314758, + 0.012015590444207191, + 0.02001187764108181, + 0.006808674894273281, + 0.005335013847798109, + 0.006802383344620466, + 0.004253692924976349, + 0.018112290650606155, + 0.007073833607137203, + 0.0073792641051113605, + 0.004733142908662558, + 0.00666497927159071, + 0.00902550108730793, + 0.009493060410022736, + 0.014058518223464489, + 0.024724947288632393, + 0.010899490676820278, + 0.009730895981192589, + 0.008736887946724892, + 0.0029855503235012293, + 0.010718729346990585, + 0.0031371964141726494, + 0.006402255035936832, + 0.006879641208797693, + 0.006266897078603506, + 0.011042522266507149, + 0.008117069490253925, + 0.0036051457282155752, + 0.006673919502645731, + 0.00394025631248951, + 0.0030232176650315523, + 0.004252850543707609, + 0.0034327704925090075, + 0.004504192154854536, + 0.0030381905380636454, + 0.006285218987613916, + 0.010192930698394775, + 0.012488511390984058, + 0.007870103232562542, + 0.012238336727023125, + 0.007319302298128605, + 0.010132319293916225, + 0.009362868033349514, + 0.004237307235598564, + 0.0034477233421057463, + 0.00548032158985734, + 0.003152546240016818, + 0.0040514604188501835, + 0.0036009575705975294, + 0.002899734303355217, + 0.0031040434259921312, + 0.00591794541105628, + 0.006200983189046383, + 0.0026850218418985605, + 0.004643552005290985, + 0.005635244306176901, + 0.005133595317602158, + 0.004757773596793413, + 0.002339098136872053, + 0.0049859099090099335, + 0.006945209112018347, + 0.005065108183771372, + 0.004372024908661842, + 0.005730651319026947, + 0.005692043341696262, + 0.00401611253619194, + 0.005971502047032118, + 0.017709219828248024, + 0.020316557958722115, + 0.01876484416425228, + 0.02058488130569458, + 0.030877767130732536, + 0.028869658708572388, + 0.014110567048192024, + 0.028283650055527687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.1270969808101654 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057763561606407 + ], + "opposite_logprob_tokens": [ + -3.8955776691436768 + ], + "logprob_reward": -0.02057763561606407, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057763561606407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1839512288570404, + 0.043021060526371, + 0.0661952942609787, + 0.051373910158872604, + 0.031146610155701637, + 0.05891036614775658, + 0.08306604623794556, + 0.11308139562606812, + 0.13603585958480835, + 0.233218252658844 + ], + "credit_map": [ + 0.03853962942957878, + 0.010646242648363113, + 0.006702370475977659, + 0.007130726240575314, + 0.02722317725419998, + 0.07642606645822525, + 0.006719940807670355, + 0.010563061572611332, + 0.005175896920263767, + 0.0037786350585520267, + 0.0033241710625588894, + 0.0038605928421020508, + 0.01468596886843443, + 0.006134038791060448, + 0.0033432934433221817, + 0.0027184635400772095, + 0.005251678638160229, + 0.007499029394239187, + 0.003164504887536168, + 0.007202500943094492, + 0.023961201310157776, + 0.008816124871373177, + 0.0055580222979187965, + 0.004742227960377932, + 0.004968162626028061, + 0.005358516238629818, + 0.004908173345029354, + 0.004247129429131746, + 0.007533247582614422, + 0.006475660018622875, + 0.008242718875408173, + 0.00964029785245657, + 0.003657546127215028, + 0.0031773631926625967, + 0.00450977124273777, + 0.003010733751580119, + 0.004574012476950884, + 0.005704570561647415, + 0.0030788318254053593, + 0.0034337788820266724, + 0.008703779429197311, + 0.010070077143609524, + 0.006121472921222448, + 0.0038889923598617315, + 0.010372571647167206, + 0.008104897104203701, + 0.006537516135722399, + 0.005111058708280325, + 0.01432354748249054, + 0.006233662832528353, + 0.01682877726852894, + 0.010301891714334488, + 0.007161705754697323, + 0.009277270175516605, + 0.009412636049091816, + 0.009526555426418781, + 0.00846022367477417, + 0.030014507472515106, + 0.00794016569852829, + 0.007862458936870098, + 0.006530277896672487, + 0.0324021577835083, + 0.010344334878027439, + 0.00952725950628519, + 0.01282753050327301, + 0.017693184316158295, + 0.017032284289598465, + 0.015640517696738243, + 0.020724769681692123, + 0.01671326719224453, + 0.015135704539716244, + 0.020268596708774567, + 0.025223925709724426, + 0.02532792277634144, + 0.026506783440709114, + 0.03207205981016159, + 0.03979426994919777, + 0.03632187098264694, + 0.0233212448656559, + 0.024650156497955322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.02057763561606407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the scene feature two distinct suns, one massive crimson giant and one smaller bright sapphire companion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the atmosphere thick, deep purple, and filled with slowly drifting iridescent clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ocean composed of liquid methane with a surface that moves in a slow, viscous, and oily manner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are enormous manta-ray-like creatures visible with smooth dark skin adorned with pulsating bioluminescent patterns in electric green, pink, and gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow tilt to the right followed by a steady roll?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31341424584388733 + ], + "opposite_logprob_tokens": [ + -1.313414216041565 + ], + "logprob_reward": -0.31341424584388733, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31341424584388733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2093300074338913, + 0.09793201833963394, + 0.07279116660356522, + 0.05309203267097473, + 0.05699794739484787, + 0.03825020790100098, + 0.046392470598220825, + 0.08699224889278412, + 0.10245122760534286, + 0.23577065765857697 + ], + "credit_map": [ + 0.04009286314249039, + 0.012970879673957825, + 0.010426176711916924, + 0.007144936826080084, + 0.0294022336602211, + 0.0855507105588913, + 0.008710963651537895, + 0.015031266026198864, + 0.010682918131351471, + 0.009513058699667454, + 0.009301266632974148, + 0.009461380541324615, + 0.025076422840356827, + 0.014981387183070183, + 0.008336160331964493, + 0.010579432360827923, + 0.004926924593746662, + 0.0064853341318666935, + 0.006048408802598715, + 0.009486316703259945, + 0.015475833788514137, + 0.010212142951786518, + 0.011374269612133503, + 0.008781944401562214, + 0.004895406775176525, + 0.007012458518147469, + 0.004571504890918732, + 0.005531109403818846, + 0.007778673432767391, + 0.00688158581033349, + 0.009632766246795654, + 0.00678853178396821, + 0.00351326959207654, + 0.0060928561724722385, + 0.007724483963102102, + 0.007573149167001247, + 0.007410272024571896, + 0.006989309098571539, + 0.010686258785426617, + 0.007008351385593414, + 0.004127583000808954, + 0.00547138461843133, + 0.005330632906407118, + 0.002999048912897706, + 0.005910904612392187, + 0.005127429962158203, + 0.003956148866564035, + 0.005327077582478523, + 0.011645893566310406, + 0.005461335647851229, + 0.006014510523527861, + 0.003482562256976962, + 0.006229544524103403, + 0.0043181744404137135, + 0.004330740310251713, + 0.0049097128212451935, + 0.007641816046088934, + 0.00904647633433342, + 0.007587041240185499, + 0.014000005088746548, + 0.01180760096758604, + 0.012830615974962711, + 0.01444733701646328, + 0.009631356224417686, + 0.009209529496729374, + 0.015125812031328678, + 0.00885670818388462, + 0.014918329194188118, + 0.014720550738275051, + 0.009958169423043728, + 0.01137640979140997, + 0.0182857196778059, + 0.023281505331397057, + 0.024700580164790154, + 0.021478477865457535, + 0.04233662039041519, + 0.03747989982366562, + 0.04393817484378815, + 0.016747888177633286, + 0.025807520374655724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.31341424584388733 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_11/index_tensor([4087, 4087])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_11/index_tensor([4087, 4087])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eed2245919777c77c9918197030b600ee33c72ed --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_11/index_tensor([4087, 4087])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a slowly advancing first-person perspective, we begin the film through the round headlights of a magical Volkswagen Type 2 De Luxe bus. Embracing the signature hand-painted quality of Hayao Miyazaki's animations, the bus boasts a soft creamy white upper body and a warm sky blue lower body, painted with a vintage matte sheen. It joyfully plows through the turquoise sea, its front gracefully parting the surface, stirring up circles of hand-painted white ripples and shimmering spray, while the bus gently rises and falls rhythmically with the waves. Ahead of the camera, a vast, dreamy ocean shimmers in the brilliant sunlight, while the sky is dotted with massive, puffy white cumulonimbus clouds, casting a warm, luminous hue over the world. As the bus steadily and deliberately navigates its path, the camera pans gently, capturing the magnificent and tranquil ocean panorama ahead, creating an immersive atmosphere of adventure filled with wonder and freedom.", + "seed": 10871, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_11/index_4087_seed_10871_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04686739584431052, + "vlm_logprob": -0.04686739584431052, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.04686739584431052, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08409782499074936 + ], + "opposite_logprob_tokens": [ + -2.7090978622436523 + ], + "logprob_reward": -0.08409782499074936, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08409782499074936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05736527219414711, + 0.07075804471969604, + 0.06261047720909119, + 0.05938907340168953, + 0.04609600454568863, + 0.06105908751487732, + 0.4542929232120514, + 0.13801921904087067, + 0.023281745612621307, + 0.027128102257847786 + ], + "credit_map": [ + 0.008194423280656338, + 0.005855461582541466, + 0.010144436731934547, + 0.011055302806198597, + 0.005192623473703861, + 0.005909262225031853, + 0.00501526053994894, + 0.005998503416776657, + 0.01045268401503563, + 0.009671619161963463, + 0.00914731714874506, + 0.0058868699707090855, + 0.007987819612026215, + 0.007812931202352047, + 0.010065537877380848, + 0.009733266197144985, + 0.018614349886775017, + 0.00625923927873373, + 0.020571019500494003, + 0.005023602861911058, + 0.002332803327590227, + 0.002193054649978876, + 0.0035289269872009754, + 0.004087480716407299, + 0.007581933867186308, + 0.00718298926949501, + 0.007786324713379145, + 0.005748386029154062, + 0.007187876850366592, + 0.00921480543911457, + 0.007115617860108614, + 0.007571139372885227, + 0.0067335618659853935, + 0.006753969006240368, + 0.0038876859471201897, + 0.003346950514242053, + 0.008462827652692795, + 0.008395008742809296, + 0.003736894577741623, + 0.004779108799993992, + 0.01612556166946888, + 0.005119381472468376, + 0.014101212844252586, + 0.005731748417019844, + 0.004746869206428528, + 0.005530413705855608, + 0.005773626267910004, + 0.003930270671844482, + 0.14839933812618256, + 0.0380467064678669, + 0.1758558601140976, + 0.02116589993238449, + 0.022844349965453148, + 0.010077978484332561, + 0.01784341223537922, + 0.02005937695503235, + 0.018515316769480705, + 0.019200563430786133, + 0.014597701840102673, + 0.015530329197645187, + 0.01274990662932396, + 0.027654578909277916, + 0.010625194758176804, + 0.01914563588798046, + 0.0038950000889599323, + 0.0032360260374844074, + 0.003015299327671528, + 0.0019077592296525836, + 0.00218651769682765, + 0.003865674138069153, + 0.001887234509922564, + 0.00328823272138834, + 0.005164287053048611, + 0.004966966807842255, + 0.0038661600556224585, + 0.002362704137340188, + 0.0038055137265473604, + 0.0033185919746756554, + 0.0014872730243951082, + 0.0021566064096987247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.08409782499074936 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589781925082207 + ], + "opposite_logprob_tokens": [ + -3.895589828491211 + ], + "logprob_reward": -0.020589781925082207, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589781925082207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1427793651819229, + 0.0533597469329834, + 0.1981630176305771, + 0.09009992331266403, + 0.025476770475506783, + 0.022467033937573433, + 0.31153565645217896, + 0.054401129484176636, + 0.01508602499961853, + 0.08663134276866913 + ], + "credit_map": [ + 0.0200858972966671, + 0.02233791910111904, + 0.019565220922231674, + 0.02049396187067032, + 0.018263384699821472, + 0.011590187437832355, + 0.018170906230807304, + 0.012271879240870476, + 0.006714735180139542, + 0.004625061992555857, + 0.005079679191112518, + 0.004200307186692953, + 0.010100358165800571, + 0.00944919977337122, + 0.006927992217242718, + 0.006262410432100296, + 0.04891057312488556, + 0.01782393828034401, + 0.0520271360874176, + 0.011544927023351192, + 0.011587820947170258, + 0.02726510539650917, + 0.011357785202562809, + 0.017645737156271935, + 0.010847201570868492, + 0.00997629202902317, + 0.009237712249159813, + 0.008446699008345604, + 0.009301080368459225, + 0.01969762332737446, + 0.011321371421217918, + 0.0112719452008605, + 0.003989498596638441, + 0.0031557066831737757, + 0.0022807198110967875, + 0.0020237162243574858, + 0.0039507742039859295, + 0.004522189497947693, + 0.00234902068041265, + 0.0032051459420472383, + 0.004986226093024015, + 0.0020131864584982395, + 0.005895128473639488, + 0.001997703220695257, + 0.002088912297040224, + 0.0026906970888376236, + 0.0013270709896460176, + 0.0014681098982691765, + 0.07519109547138214, + 0.020992537960410118, + 0.12778955698013306, + 0.01528223417699337, + 0.03822854906320572, + 0.008729403838515282, + 0.011958491057157516, + 0.013363786041736603, + 0.008892888203263283, + 0.00970500148832798, + 0.006830996833741665, + 0.0038800574839115143, + 0.006497087888419628, + 0.006717997137457132, + 0.0036386747378855944, + 0.008238430134952068, + 0.002327241934835911, + 0.0017126627499237657, + 0.0022864267230033875, + 0.0022378633730113506, + 0.001309093669988215, + 0.0013880956685170531, + 0.002225936623290181, + 0.0015987057704478502, + 0.012428992427885532, + 0.006369469221681356, + 0.020421432331204414, + 0.006495542358607054, + 0.01934005133807659, + 0.005672916304320097, + 0.006234249100089073, + 0.009668685495853424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.020589781925082207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897115498781204 + ], + "opposite_logprob_tokens": [ + -2.5789711475372314 + ], + "logprob_reward": -0.07897115498781204, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897115498781204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11719035357236862, + 0.1624092012643814, + 0.08058032393455505, + 0.04815863445401192, + 0.060924820601940155, + 0.10173185169696808, + 0.17246800661087036, + 0.11444015800952911, + 0.03981577977538109, + 0.10228091478347778 + ], + "credit_map": [ + 0.013793746009469032, + 0.019007757306098938, + 0.01553838700056076, + 0.014898736961185932, + 0.0073235188610851765, + 0.014306408353149891, + 0.014239304699003696, + 0.018082495778799057, + 0.021210817620158195, + 0.012141960673034191, + 0.023683298379182816, + 0.01675792969763279, + 0.01596127636730671, + 0.028462566435337067, + 0.02288544923067093, + 0.021305911242961884, + 0.01738112047314644, + 0.007377089932560921, + 0.018041422590613365, + 0.005630324129015207, + 0.00659605860710144, + 0.010319672524929047, + 0.006179690361022949, + 0.009054948575794697, + 0.008153872564435005, + 0.006269830744713545, + 0.005521878134459257, + 0.0045759305357933044, + 0.005406448617577553, + 0.005615847185254097, + 0.0061061750166118145, + 0.006508654914796352, + 0.008617951534688473, + 0.009490071795880795, + 0.006172446999698877, + 0.006194070912897587, + 0.006704873405396938, + 0.010861964896321297, + 0.006250828038901091, + 0.00663261441513896, + 0.025552954524755478, + 0.009446821175515652, + 0.025163833051919937, + 0.007171907927840948, + 0.007187283132225275, + 0.011443546041846275, + 0.010584580712020397, + 0.005180924665182829, + 0.06622937321662903, + 0.016303956508636475, + 0.030366335064172745, + 0.005044444929808378, + 0.03760743886232376, + 0.006672956049442291, + 0.00511733815073967, + 0.005126172211021185, + 0.019731778651475906, + 0.019077882170677185, + 0.013680322095751762, + 0.011745274998247623, + 0.013066221959888935, + 0.01505852397531271, + 0.008501032367348671, + 0.013579118065536022, + 0.006273149978369474, + 0.004301398992538452, + 0.004982443526387215, + 0.00456538051366806, + 0.0056276326067745686, + 0.006307173985987902, + 0.003879103111103177, + 0.00387949519790709, + 0.01725306361913681, + 0.009928951971232891, + 0.019252941012382507, + 0.006331702694296837, + 0.0231211856007576, + 0.009229297749698162, + 0.00797940045595169, + 0.009184366092085838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.07897115498781204 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762315234169364 + ], + "opposite_logprob_tokens": [ + -4.882623195648193 + ], + "logprob_reward": -0.00762315234169364, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762315234169364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1740928739309311, + 0.18155115842819214, + 0.052600130438804626, + 0.045989759266376495, + 0.042960505932569504, + 0.039084576070308685, + 0.1417614221572876, + 0.09324961155653, + 0.060860734432935715, + 0.16784921288490295 + ], + "credit_map": [ + 0.0317029245197773, + 0.026186194270849228, + 0.030110418796539307, + 0.02398361638188362, + 0.012364532798528671, + 0.020263126119971275, + 0.012905147857964039, + 0.016576917842030525, + 0.02886701561510563, + 0.022610314190387726, + 0.02567906118929386, + 0.015086480416357517, + 0.02094261907041073, + 0.020987581461668015, + 0.02620685286819935, + 0.021171241998672485, + 0.012495174072682858, + 0.0047888741828501225, + 0.013265026733279228, + 0.005054109264165163, + 0.0038816980086266994, + 0.004702569916844368, + 0.005094809457659721, + 0.0033178653102368116, + 0.0077225212007761, + 0.004243143368512392, + 0.0052941651083528996, + 0.005102983210235834, + 0.0037115125451236963, + 0.007048663683235645, + 0.005315009970217943, + 0.007551760412752628, + 0.007722544949501753, + 0.00585036538541317, + 0.0028217872604727745, + 0.0032077282667160034, + 0.010645695962011814, + 0.005456066224724054, + 0.0035186551976948977, + 0.003737666644155979, + 0.005859391298145056, + 0.004358187317848206, + 0.006955690681934357, + 0.0032233307138085365, + 0.00508025661110878, + 0.006940948311239481, + 0.0031629735603928566, + 0.0035037987399846315, + 0.030460894107818604, + 0.009367488324642181, + 0.05546913295984268, + 0.007178807631134987, + 0.015949638560414314, + 0.004333339631557465, + 0.007602485362440348, + 0.011399630457162857, + 0.011091085150837898, + 0.014339993707835674, + 0.009967035613954067, + 0.010275200009346008, + 0.010600071400403976, + 0.017408162355422974, + 0.007426853757351637, + 0.012141210027039051, + 0.007285611238330603, + 0.006674595642834902, + 0.007976011373102665, + 0.006997156888246536, + 0.008290102705359459, + 0.009484968148171902, + 0.005900438409298658, + 0.008251852355897427, + 0.024633273482322693, + 0.0194659773260355, + 0.031361572444438934, + 0.01890774816274643, + 0.0268251970410347, + 0.017705479636788368, + 0.010544233955442905, + 0.018405726179480553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.00762315234169364 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305506497621536 + ], + "opposite_logprob_tokens": [ + -3.1680550575256348 + ], + "logprob_reward": -0.04305506497621536, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305506497621536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030863188207149506, + 0.04311025142669678, + 0.11250564455986023, + 0.03990523889660835, + 0.047015320509672165, + 0.0499521903693676, + 0.4280230402946472, + 0.20063045620918274, + 0.025692744180560112, + 0.0223019290715456 + ], + "credit_map": [ + 0.004703355021774769, + 0.0037161665968596935, + 0.004705795086920261, + 0.004874801263213158, + 0.003282382385805249, + 0.0036197782028466463, + 0.002952304668724537, + 0.0030086063779890537, + 0.008161874487996101, + 0.007531775161623955, + 0.0070875259116292, + 0.004187410231679678, + 0.003296902868896723, + 0.004633730743080378, + 0.0032694432884454727, + 0.0049415891990065575, + 0.03711531683802605, + 0.011200212873518467, + 0.03941123187541962, + 0.008094828575849533, + 0.003614960238337517, + 0.0029316816944628954, + 0.004791854880750179, + 0.005345546640455723, + 0.007513943128287792, + 0.0037959986366331577, + 0.005571249406784773, + 0.0040879142470657825, + 0.00316453049890697, + 0.004909980576485395, + 0.00455824937671423, + 0.006303373258560896, + 0.008616989478468895, + 0.0075812614522874355, + 0.0034936026204377413, + 0.0030676256865262985, + 0.011618916876614094, + 0.00579296238720417, + 0.003136473475024104, + 0.0037074859719723463, + 0.012765537016093731, + 0.004951831419020891, + 0.010513857938349247, + 0.0040200212970376015, + 0.0052958386950194836, + 0.00474363099783659, + 0.004144652746617794, + 0.0035168188624083996, + 0.13569360971450806, + 0.034594591706991196, + 0.16764554381370544, + 0.020273525267839432, + 0.02176784723997116, + 0.008559425361454487, + 0.017424704506993294, + 0.02206377498805523, + 0.025248412042856216, + 0.02520970068871975, + 0.022805994376540184, + 0.02201230265200138, + 0.01863142102956772, + 0.04159848764538765, + 0.015054420568048954, + 0.03006972372531891, + 0.0036143939942121506, + 0.002750930143520236, + 0.003754084464162588, + 0.0023195045068860054, + 0.0029342814814299345, + 0.004005449824035168, + 0.001995750702917576, + 0.004318349529057741, + 0.0023688306100666523, + 0.0015132541302591562, + 0.005947408266365528, + 0.0021191828418523073, + 0.004369920119643211, + 0.0017017213394865394, + 0.0016490857815369964, + 0.002632526447996497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.04305506497621536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slowly advancing first-person perspective, we begin the film through the round headlights of a magical Volkswagen Type 2 De Luxe bus. Embracing the signature hand-painted quality of Hayao Miyazaki's animations, the bus boasts a soft creamy white upper body and a warm sky blue lower body, painted with a vintage matte sheen. It joyfully plows through the turquoise sea, its front gracefully parting the surface, stirring up circles of hand-painted white ripples and shimmering spray, while the bus gently rises and falls rhythmically with the waves. Ahead of the camera, a vast, dreamy ocean shimmers in the brilliant sunlight, while the sky is dotted with massive, puffy white cumulonimbus clouds, casting a warm, luminous hue over the world. As the bus steadily and deliberately navigates its path, the camera pans gently, capturing the magnificent and tranquil ocean panorama ahead, creating an immersive atmosphere of adventure filled with wonder and freedom.", + "seed": 10872, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_11/index_4087_seed_10872_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07369218207895756, + "vlm_logprob": -0.07369218207895756, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.07369218207895756, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12310929596424103 + ], + "opposite_logprob_tokens": [ + -2.3731093406677246 + ], + "logprob_reward": -0.12310929596424103, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12310929596424103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08700165152549744, + 0.07202128320932388, + 0.03687872737646103, + 0.10635008662939072, + 0.12357258796691895, + 0.01932631991803646, + 0.13714823126792908, + 0.07855424284934998, + 0.18484337627887726, + 0.15430352091789246 + ], + "credit_map": [ + 0.01209264900535345, + 0.007176980841904879, + 0.0107655581086874, + 0.010827071033418179, + 0.008238116279244423, + 0.008137434720993042, + 0.01941515877842903, + 0.010348674841225147, + 0.00800317246466875, + 0.005870176013559103, + 0.010359598323702812, + 0.010075763799250126, + 0.007649917621165514, + 0.006040864624083042, + 0.013907825574278831, + 0.010113957338035107, + 0.005702648777514696, + 0.0039707934483885765, + 0.0035745876375585794, + 0.003815106116235256, + 0.004546002484858036, + 0.003335140645503998, + 0.007788120303303003, + 0.004146327264606953, + 0.036385927349328995, + 0.011035677045583725, + 0.03483985736966133, + 0.00809736829251051, + 0.0040444424375891685, + 0.003966135438531637, + 0.004868448246270418, + 0.00311222649179399, + 0.04623454436659813, + 0.014198818244040012, + 0.0430879071354866, + 0.005621663760393858, + 0.003421560162678361, + 0.004312422126531601, + 0.003757193684577942, + 0.0029384794179350138, + 0.003276480594649911, + 0.002148703671991825, + 0.002465170808136463, + 0.002453613793477416, + 0.0021426433231681585, + 0.0020742188207805157, + 0.0032440712675452232, + 0.0015214158920571208, + 0.027144553139805794, + 0.014509372413158417, + 0.016890622675418854, + 0.010618436150252819, + 0.023621996864676476, + 0.011459523811936378, + 0.02134147845208645, + 0.01156223937869072, + 0.01079322025179863, + 0.010076059959828854, + 0.007505266927182674, + 0.008830699138343334, + 0.012801533564925194, + 0.007145441137254238, + 0.013796838000416756, + 0.007605172693729401, + 0.023651592433452606, + 0.021978572010993958, + 0.025996742770075798, + 0.020552603527903557, + 0.029328256845474243, + 0.016254616901278496, + 0.03181556612253189, + 0.01526542566716671, + 0.054222047328948975, + 0.012843243777751923, + 0.059601955115795135, + 0.0075853983871638775, + 0.004572375677525997, + 0.004069838672876358, + 0.00807972438633442, + 0.003328910330310464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.12310929596424103 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078457236289978 + ], + "opposite_logprob_tokens": [ + -4.141078472137451 + ], + "logprob_reward": -0.016078457236289978, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078457236289978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03710589557886124, + 0.04381706193089485, + 0.028055110946297646, + 0.533664345741272, + 0.07042504101991653, + 0.01365269348025322, + 0.05263229086995125, + 0.0940154492855072, + 0.05522450804710388, + 0.07140760123729706 + ], + "credit_map": [ + 0.006987058557569981, + 0.005203408654779196, + 0.005280951503664255, + 0.004949552472680807, + 0.00414015119895339, + 0.003236116375774145, + 0.00486755883321166, + 0.0024410956539213657, + 0.00585928512737155, + 0.004078540951013565, + 0.00921392347663641, + 0.004169001244008541, + 0.004034786485135555, + 0.0034027628134936094, + 0.00727665051817894, + 0.005782107822597027, + 0.004613690543919802, + 0.00334706692956388, + 0.003001015866175294, + 0.0031821492593735456, + 0.0030643115751445293, + 0.0022158604115247726, + 0.004909581039100885, + 0.0037214327603578568, + 0.20630668103694916, + 0.06703180074691772, + 0.18216776847839355, + 0.02662808448076248, + 0.015331325121223927, + 0.014520714059472084, + 0.008998716250061989, + 0.012679209001362324, + 0.024072736501693726, + 0.008134301751852036, + 0.023171380162239075, + 0.003580021671950817, + 0.0024255202151834965, + 0.0032773332204669714, + 0.0031574303284287453, + 0.0026063169352710247, + 0.0019374671392142773, + 0.002031070878729224, + 0.002027754206210375, + 0.0012333138147369027, + 0.0014701378531754017, + 0.0015068157808855176, + 0.002198664238676429, + 0.0012474688701331615, + 0.008792119100689888, + 0.0061468095518648624, + 0.008265186101198196, + 0.005166387185454369, + 0.008500033058226109, + 0.004450842272490263, + 0.007748457603156567, + 0.0035624506417661905, + 0.014081121422350407, + 0.010721788741648197, + 0.008732526563107967, + 0.014597326517105103, + 0.011642573401331902, + 0.012154889293015003, + 0.012688728980720043, + 0.009396483190357685, + 0.007143915630877018, + 0.007685719523578882, + 0.007125538773834705, + 0.005776196718215942, + 0.008084888570010662, + 0.007294353563338518, + 0.0077896821312606335, + 0.00432421313598752, + 0.021063195541501045, + 0.006257082801312208, + 0.022249452769756317, + 0.004839730449020863, + 0.0032133334316313267, + 0.0033286502584815025, + 0.007571645081043243, + 0.002884508343413472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 9 + ], + "selected_reward": -0.016078457236289978 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292041838169098 + ], + "opposite_logprob_tokens": [ + -2.2379205226898193 + ], + "logprob_reward": -0.11292041838169098, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292041838169098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05808960273861885, + 0.040466733276844025, + 0.01666916534304619, + 0.5249015092849731, + 0.044920045882463455, + 0.013702819123864174, + 0.05591784417629242, + 0.04020736366510391, + 0.11214012652635574, + 0.09298478811979294 + ], + "credit_map": [ + 0.012837338261306286, + 0.008199046365916729, + 0.007077772170305252, + 0.006562670227140188, + 0.006116922479122877, + 0.004344174172729254, + 0.008298642933368683, + 0.004653037991374731, + 0.00613428745418787, + 0.004171160981059074, + 0.00748335151001811, + 0.007636594120413065, + 0.002701124642044306, + 0.003654157742857933, + 0.005329188425093889, + 0.0033568714279681444, + 0.002498304471373558, + 0.001845294376835227, + 0.0022075281012803316, + 0.0016433052951470017, + 0.0016371910460293293, + 0.002156633185222745, + 0.0031142772641032934, + 0.0015666328836232424, + 0.2024732530117035, + 0.06473830342292786, + 0.18045344948768616, + 0.025463728234171867, + 0.01495532039552927, + 0.01392362080514431, + 0.0104402806609869, + 0.012453551404178143, + 0.015744920819997787, + 0.005229630041867495, + 0.014365392737090588, + 0.00239507551304996, + 0.001609285594895482, + 0.00170626409817487, + 0.0023810144048184156, + 0.0014884623233228922, + 0.001732444390654564, + 0.0012884309981018305, + 0.0017386676045134664, + 0.0018542228499427438, + 0.0013237141538411379, + 0.0010095592588186264, + 0.002916760044172406, + 0.001839019008912146, + 0.008793128654360771, + 0.007562792394310236, + 0.009468037635087967, + 0.008038480766117573, + 0.006725333631038666, + 0.004122087266296148, + 0.006774978246539831, + 0.004433006048202515, + 0.0046665552072227, + 0.004658367019146681, + 0.004137768410146236, + 0.00558500736951828, + 0.004366119857877493, + 0.004252246115356684, + 0.008872751146554947, + 0.0036685452796518803, + 0.01539980061352253, + 0.011354143731296062, + 0.0133089330047369, + 0.012632831931114197, + 0.018437301740050316, + 0.009962539188563824, + 0.022298116236925125, + 0.008746449835598469, + 0.03229173272848129, + 0.009368625469505787, + 0.03542318195104599, + 0.0046425387263298035, + 0.0028484167996793985, + 0.0022697055246680975, + 0.004260942805558443, + 0.0018796382937580347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.11292041838169098 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016059452667832375 + ], + "opposite_logprob_tokens": [ + -4.141059398651123 + ], + "logprob_reward": -0.016059452667832375, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016059452667832375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031668566167354584, + 0.019868163391947746, + 0.02047017402946949, + 0.6402896046638489, + 0.14224526286125183, + 0.04445609077811241, + 0.01058392133563757, + 0.01703808456659317, + 0.03856753557920456, + 0.0348125658929348 + ], + "credit_map": [ + 0.006387289613485336, + 0.005279491189867258, + 0.004846423864364624, + 0.005350240971893072, + 0.0025829521473497152, + 0.002180054085329175, + 0.0031792910303920507, + 0.001862822682596743, + 0.0025134726893156767, + 0.003474260214716196, + 0.0033283461816608906, + 0.003048558486625552, + 0.0012129668612033129, + 0.0023122513666749, + 0.0026179675478488207, + 0.0013603400439023972, + 0.003461651038378477, + 0.003508850699290633, + 0.003272377885878086, + 0.0024874883238226175, + 0.0020382062066346407, + 0.0014105262234807014, + 0.0026705937925726175, + 0.0016204796265810728, + 0.24461549520492554, + 0.07914537191390991, + 0.21899515390396118, + 0.032467883080244064, + 0.01917121559381485, + 0.019225062802433968, + 0.011066141538321972, + 0.015603296458721161, + 0.05387646704912186, + 0.017078734934329987, + 0.049087028950452805, + 0.005830670706927776, + 0.004302551969885826, + 0.004715512506663799, + 0.003986022435128689, + 0.003368270117789507, + 0.007913840934634209, + 0.009045077487826347, + 0.006559228524565697, + 0.006525845266878605, + 0.0037707050796598196, + 0.0038665200117975473, + 0.003394670784473419, + 0.0033802043180912733, + 0.0012396910460665822, + 0.000979328528046608, + 0.0011935561196878552, + 0.001651316531933844, + 0.0016618233639746904, + 0.0010306840995326638, + 0.0014824791578575969, + 0.0013450414407998323, + 0.002516771899536252, + 0.0019702506251633167, + 0.0016223132843151689, + 0.0027321316301822662, + 0.001740213599987328, + 0.0018532315734773874, + 0.0032698139548301697, + 0.0013333577662706375, + 0.004876449704170227, + 0.004700932651758194, + 0.004485598765313625, + 0.0042485729791224, + 0.006282708141952753, + 0.00334834773093462, + 0.007500631734728813, + 0.00312429154291749, + 0.008907252922654152, + 0.0030748157296329737, + 0.009626402519643307, + 0.0030768474098294973, + 0.002975311130285263, + 0.0018159786704927683, + 0.0037882865872234106, + 0.0015476696426048875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.016059452667832375 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029328614473343 + ], + "opposite_logprob_tokens": [ + -2.3502933979034424 + ], + "logprob_reward": -0.10029328614473343, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029328614473343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047722212970256805, + 0.03344682231545448, + 0.017294468358159065, + 0.5100607872009277, + 0.04535197839140892, + 0.027293497696518898, + 0.042518746107816696, + 0.10374773293733597, + 0.06675500422716141, + 0.10580883920192719 + ], + "credit_map": [ + 0.010006160475313663, + 0.007025313563644886, + 0.006112503819167614, + 0.005947669502347708, + 0.004353513475507498, + 0.003436916507780552, + 0.00697724474593997, + 0.0038628880865871906, + 0.0036412666086107492, + 0.003622486488893628, + 0.005072219762951136, + 0.005887521896511316, + 0.003226127242669463, + 0.0026857915800064802, + 0.005992287304252386, + 0.0033191151451319456, + 0.0026669763028621674, + 0.0019234227947890759, + 0.0027053349185734987, + 0.0016123492969200015, + 0.0019902160856872797, + 0.001387626864016056, + 0.003508805762976408, + 0.0014997333055362105, + 0.19638599455356598, + 0.0638827234506607, + 0.1740676760673523, + 0.02521352283656597, + 0.01417201291769743, + 0.014119167812168598, + 0.01003036368638277, + 0.012189257889986038, + 0.01286221668124199, + 0.004602006636559963, + 0.011453373357653618, + 0.002677986864000559, + 0.0031705861911177635, + 0.003005236154422164, + 0.00450548529624939, + 0.003075085114687681, + 0.003728085896000266, + 0.003985454794019461, + 0.004207051824778318, + 0.0031937125604599714, + 0.0030937271658331156, + 0.0018823316786438227, + 0.004930915776640177, + 0.0022722145076841116, + 0.007423232775181532, + 0.005626956932246685, + 0.005731642711907625, + 0.004692985676229, + 0.004370234441012144, + 0.0034769298508763313, + 0.007991470396518707, + 0.003205288201570511, + 0.01457236148416996, + 0.013744452968239784, + 0.008164754137396812, + 0.009912017732858658, + 0.018795309588313103, + 0.009769700467586517, + 0.01958235539495945, + 0.009206770919263363, + 0.00830767024308443, + 0.010713059455156326, + 0.005798446945846081, + 0.006875045131891966, + 0.011522697284817696, + 0.006951678078621626, + 0.011376485228538513, + 0.005209919530898333, + 0.033890243619680405, + 0.009983574040234089, + 0.03864592686295509, + 0.007782192900776863, + 0.003929799422621727, + 0.004824467469006777, + 0.00377679243683815, + 0.0029758282471448183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.10029328614473343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slowly advancing first-person perspective, we begin the film through the round headlights of a magical Volkswagen Type 2 De Luxe bus. Embracing the signature hand-painted quality of Hayao Miyazaki's animations, the bus boasts a soft creamy white upper body and a warm sky blue lower body, painted with a vintage matte sheen. It joyfully plows through the turquoise sea, its front gracefully parting the surface, stirring up circles of hand-painted white ripples and shimmering spray, while the bus gently rises and falls rhythmically with the waves. Ahead of the camera, a vast, dreamy ocean shimmers in the brilliant sunlight, while the sky is dotted with massive, puffy white cumulonimbus clouds, casting a warm, luminous hue over the world. As the bus steadily and deliberately navigates its path, the camera pans gently, capturing the magnificent and tranquil ocean panorama ahead, creating an immersive atmosphere of adventure filled with wonder and freedom.", + "seed": 10873, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_11/index_4087_seed_10873_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06050409264862537, + "vlm_logprob": -0.06050409264862537, + "vlm_yes_no_margin": 3.2749999523162843, + "vlm_token_credit": -0.06050409264862537, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1078362837433815 + ], + "opposite_logprob_tokens": [ + -2.7328362464904785 + ], + "logprob_reward": -0.1078362837433815, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.1078362837433815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0803154855966568, + 0.07828933000564575, + 0.060092754662036896, + 0.06944017112255096, + 0.0583471953868866, + 0.30415692925453186, + 0.08580601960420609, + 0.13237766921520233, + 0.10224470496177673, + 0.028929781168699265 + ], + "credit_map": [ + 0.011626293882727623, + 0.008449753746390343, + 0.007243725936859846, + 0.012041285634040833, + 0.009122155606746674, + 0.008091254159808159, + 0.01593194715678692, + 0.007809066213667393, + 0.012267799116671085, + 0.015123842284083366, + 0.013177846558392048, + 0.006978098303079605, + 0.005098754074424505, + 0.012137615121901035, + 0.006041930057108402, + 0.007463442627340555, + 0.008823767304420471, + 0.008794921450316906, + 0.006220279261469841, + 0.007720975670963526, + 0.00642978074029088, + 0.008569881319999695, + 0.006468466017395258, + 0.007064687553793192, + 0.010645232163369656, + 0.008320407941937447, + 0.006203331984579563, + 0.007979209534823895, + 0.005680035334080458, + 0.0123088788241148, + 0.0076816813088953495, + 0.010621392168104649, + 0.009448613040149212, + 0.008119061589241028, + 0.006325603928416967, + 0.00566305173560977, + 0.0056698257103562355, + 0.006430317182093859, + 0.010440179146826267, + 0.006250540725886822, + 0.10446937382221222, + 0.01964111439883709, + 0.11131232231855392, + 0.012814194895327091, + 0.012695475481450558, + 0.012395938858389854, + 0.009885787963867188, + 0.02094269171357155, + 0.013705547899007797, + 0.018764987587928772, + 0.011530292220413685, + 0.004729077219963074, + 0.008295300416648388, + 0.01411385741084814, + 0.00549195846542716, + 0.009175007231533527, + 0.01722797378897667, + 0.02123803272843361, + 0.012476927600800991, + 0.019881712272763252, + 0.010270733386278152, + 0.008668847382068634, + 0.015845809131860733, + 0.026767635717988014, + 0.00927053764462471, + 0.011938943527638912, + 0.009168241173028946, + 0.006996740587055683, + 0.010626263916492462, + 0.008098470978438854, + 0.025810606777668, + 0.020334895700216293, + 0.0037835470866411924, + 0.004294616170227528, + 0.002164575271308422, + 0.0015831264900043607, + 0.0030375891365110874, + 0.007578924763947725, + 0.0019270159536972642, + 0.00456038536503911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.1078362837433815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648676797747612 + ], + "opposite_logprob_tokens": [ + -4.758648872375488 + ], + "logprob_reward": -0.008648676797747612, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648676797747612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07013778388500214, + 0.1121540516614914, + 0.03573654592037201, + 0.06775777041912079, + 0.04945209249854088, + 0.1895776391029358, + 0.23465797305107117, + 0.14227181673049927, + 0.07261538505554199, + 0.025638984516263008 + ], + "credit_map": [ + 0.009846244938671589, + 0.010163501836359501, + 0.010037851519882679, + 0.011546529829502106, + 0.006518350914120674, + 0.007786482572555542, + 0.00855259783565998, + 0.0056862253695726395, + 0.01900121383368969, + 0.0095593873411417, + 0.014937715604901314, + 0.016834894195199013, + 0.009494934231042862, + 0.010679996572434902, + 0.017665734514594078, + 0.013980165123939514, + 0.004871871788054705, + 0.00595384556800127, + 0.002727421233430505, + 0.00556632224470377, + 0.003974908031523228, + 0.0033667422831058502, + 0.005690016783773899, + 0.0035854170564562082, + 0.01219247654080391, + 0.011896541342139244, + 0.004709393717348576, + 0.005516875069588423, + 0.005825644824653864, + 0.015084159560501575, + 0.005500155966728926, + 0.0070325229316949844, + 0.007350811269134283, + 0.006988159380853176, + 0.005579711403697729, + 0.006068344693630934, + 0.004481661599129438, + 0.006341413129121065, + 0.005439718719571829, + 0.007202269043773413, + 0.05085282400250435, + 0.01577966846525669, + 0.051301613450050354, + 0.012695306912064552, + 0.015072443522512913, + 0.012827193364501, + 0.012770026922225952, + 0.01827855408191681, + 0.0730852410197258, + 0.026328612118959427, + 0.06801509112119675, + 0.011609390377998352, + 0.011136777698993683, + 0.019861964508891106, + 0.009724017232656479, + 0.014896883629262447, + 0.017769107595086098, + 0.020195934921503067, + 0.012779613956809044, + 0.02813110686838627, + 0.010061829350888729, + 0.007692595943808556, + 0.013984386809170246, + 0.03165723755955696, + 0.009165461175143719, + 0.008432718925178051, + 0.006485410034656525, + 0.006444255355745554, + 0.00826178677380085, + 0.0036124917678534985, + 0.022506769746541977, + 0.007706489879637957, + 0.005487562157213688, + 0.0031022026669234037, + 0.0019428174709901214, + 0.0014967895112931728, + 0.0032638453412801027, + 0.005185084417462349, + 0.0020025584381073713, + 0.0031581243965774775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.008648676797747612 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291328817605972 + ], + "opposite_logprob_tokens": [ + -2.2379133701324463 + ], + "logprob_reward": -0.11291328817605972, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291328817605972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10081300139427185, + 0.15406787395477295, + 0.04145900160074234, + 0.11540935933589935, + 0.06088807061314583, + 0.19808699190616608, + 0.1537470817565918, + 0.0829215794801712, + 0.0723758116364479, + 0.020231187343597412 + ], + "credit_map": [ + 0.01616823486983776, + 0.01232854649424553, + 0.01872318983078003, + 0.011029984802007675, + 0.006211047992110252, + 0.015312176197767258, + 0.008620841428637505, + 0.012418991886079311, + 0.019826659932732582, + 0.020611019805073738, + 0.018596632406115532, + 0.01903451979160309, + 0.01297442615032196, + 0.024299519136548042, + 0.018619174137711525, + 0.02010595239698887, + 0.006874138489365578, + 0.006244212854653597, + 0.00438213674351573, + 0.004648867063224316, + 0.0028344946913421154, + 0.007531663402915001, + 0.004812934435904026, + 0.0041305553168058395, + 0.01985855959355831, + 0.019128646701574326, + 0.010235111229121685, + 0.009509636089205742, + 0.007655404508113861, + 0.0224679633975029, + 0.009459083899855614, + 0.017094964161515236, + 0.009988117031753063, + 0.010401635430753231, + 0.008792116306722164, + 0.005166532006114721, + 0.0050555686466395855, + 0.006068292073905468, + 0.006581934168934822, + 0.00883388239890337, + 0.06857598572969437, + 0.013514433987438679, + 0.07060667127370834, + 0.008396351709961891, + 0.007638274691998959, + 0.00817799661308527, + 0.006262368522584438, + 0.014914928935468197, + 0.04128742218017578, + 0.011259093880653381, + 0.05565643310546875, + 0.007001006975769997, + 0.007361568044871092, + 0.01428500097244978, + 0.006421939004212618, + 0.01047462411224842, + 0.01068708673119545, + 0.016606120392680168, + 0.00774290319532156, + 0.013357589952647686, + 0.005292631220072508, + 0.004133263602852821, + 0.007674913387745619, + 0.017427077516913414, + 0.009191378951072693, + 0.009546258486807346, + 0.0052108922973275185, + 0.0049388594925403595, + 0.006294718477874994, + 0.0060479845851659775, + 0.016873693093657494, + 0.014272035099565983, + 0.002413212787359953, + 0.004635939374566078, + 0.0017127996543422341, + 0.0011096058879047632, + 0.001847207429818809, + 0.004728587809950113, + 0.0015317591605708003, + 0.002252077916637063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.11291328817605972 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018175363540649414 + ], + "opposite_logprob_tokens": [ + -4.01817512512207 + ], + "logprob_reward": -0.018175363540649414, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018175363540649414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16935046017169952, + 0.044462330639362335, + 0.2389271855354309, + 0.06766267120838165, + 0.04320019856095314, + 0.16699464619159698, + 0.04392975568771362, + 0.05521108955144882, + 0.07907615602016449, + 0.09118551015853882 + ], + "credit_map": [ + 0.031083211302757263, + 0.022652780637145042, + 0.03068551793694496, + 0.01622815988957882, + 0.008107445202767849, + 0.02821335382759571, + 0.011482683941721916, + 0.020897308364510536, + 0.006836913526058197, + 0.005039679352194071, + 0.0065138377249240875, + 0.005706430412828922, + 0.005671214312314987, + 0.004146815277636051, + 0.005476818419992924, + 0.005070623476058245, + 0.0357125885784626, + 0.028261132538318634, + 0.025103146210312843, + 0.02631603367626667, + 0.015771841630339622, + 0.04327099025249481, + 0.029577551409602165, + 0.034913916140794754, + 0.010474812239408493, + 0.006042874418199062, + 0.0065125953406095505, + 0.007856923155486584, + 0.004467407241463661, + 0.008539590053260326, + 0.00798545591533184, + 0.01578301005065441, + 0.007562547456473112, + 0.010197102092206478, + 0.006103168707340956, + 0.004070445895195007, + 0.003285911399871111, + 0.003589909989386797, + 0.0051838550716638565, + 0.00320726097561419, + 0.04916964843869209, + 0.023968789726495743, + 0.04634682089090347, + 0.00916888564825058, + 0.008173305541276932, + 0.007514557335525751, + 0.010031688958406448, + 0.01262094359844923, + 0.013367903418838978, + 0.007853640243411064, + 0.009487707167863846, + 0.003015258815139532, + 0.0021556743886321783, + 0.002218629000708461, + 0.0023853732272982597, + 0.0034455708228051662, + 0.007244564127177, + 0.007497153244912624, + 0.0054514650255441666, + 0.008162721060216427, + 0.0071411654353141785, + 0.004542839713394642, + 0.009756162762641907, + 0.005415019113570452, + 0.012080962769687176, + 0.009114816784858704, + 0.007740973494946957, + 0.007590745110064745, + 0.01051862072199583, + 0.004750835243612528, + 0.019993245601654053, + 0.0072859544306993484, + 0.018006175756454468, + 0.011530325748026371, + 0.008615199476480484, + 0.005334561690688133, + 0.013002253137528896, + 0.01632572151720524, + 0.008379669860005379, + 0.009991604834794998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.018175363540649414 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494685098528862 + ], + "opposite_logprob_tokens": [ + -2.9299468994140625 + ], + "logprob_reward": -0.05494685098528862, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494685098528862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0752580389380455, + 0.06397035717964172, + 0.02579805999994278, + 0.036962106823921204, + 0.040381792932748795, + 0.13292591273784637, + 0.3745386600494385, + 0.11340437084436417, + 0.05765913799405098, + 0.0791015550494194 + ], + "credit_map": [ + 0.011434419080615044, + 0.009985269047319889, + 0.007319452241063118, + 0.010789016261696815, + 0.008899984881281853, + 0.008467118255794048, + 0.010258462280035019, + 0.008104330860078335, + 0.009559576399624348, + 0.009844586253166199, + 0.009995228610932827, + 0.005901200231164694, + 0.006124704144895077, + 0.01114482618868351, + 0.006333609111607075, + 0.005066624376922846, + 0.0041428483091294765, + 0.005811805371195078, + 0.002172974869608879, + 0.0029624300077557564, + 0.0025317573454231024, + 0.0035065605770796537, + 0.0022012076806277037, + 0.0024684779345989227, + 0.0057548522017896175, + 0.005557825788855553, + 0.002503401366993785, + 0.005011650733649731, + 0.003991700243204832, + 0.005542050581425428, + 0.004609412048012018, + 0.003991215489804745, + 0.0059873731806874275, + 0.006152378860861063, + 0.005495843011885881, + 0.003926025237888098, + 0.004382299724966288, + 0.0033211277332156897, + 0.00615547364577651, + 0.004961276892572641, + 0.044962506741285324, + 0.010897847823798656, + 0.045045286417007446, + 0.006188065279275179, + 0.0050916532054543495, + 0.0054430970922112465, + 0.005445168819278479, + 0.009852292016148567, + 0.12829844653606415, + 0.029790043830871582, + 0.14327512681484222, + 0.016316192224621773, + 0.011128909885883331, + 0.012677245773375034, + 0.015871314331889153, + 0.017181411385536194, + 0.014106365852057934, + 0.02249726466834545, + 0.008843044750392437, + 0.010405899956822395, + 0.012273920699954033, + 0.008312515914440155, + 0.013823212124407291, + 0.02314215525984764, + 0.006945640780031681, + 0.007938406430184841, + 0.0060921888798475266, + 0.006866867188364267, + 0.007171680685132742, + 0.00571097619831562, + 0.011594749055802822, + 0.005338631570339203, + 0.011622235178947449, + 0.019060663878917694, + 0.007856165058910847, + 0.006286278832703829, + 0.008550111204385757, + 0.012309776619076729, + 0.005176777020096779, + 0.00823955237865448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.05494685098528862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slowly advancing first-person perspective, we begin the film through the round headlights of a magical Volkswagen Type 2 De Luxe bus. Embracing the signature hand-painted quality of Hayao Miyazaki's animations, the bus boasts a soft creamy white upper body and a warm sky blue lower body, painted with a vintage matte sheen. It joyfully plows through the turquoise sea, its front gracefully parting the surface, stirring up circles of hand-painted white ripples and shimmering spray, while the bus gently rises and falls rhythmically with the waves. Ahead of the camera, a vast, dreamy ocean shimmers in the brilliant sunlight, while the sky is dotted with massive, puffy white cumulonimbus clouds, casting a warm, luminous hue over the world. As the bus steadily and deliberately navigates its path, the camera pans gently, capturing the magnificent and tranquil ocean panorama ahead, creating an immersive atmosphere of adventure filled with wonder and freedom.", + "seed": 10874, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_11/index_4087_seed_10874_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06415150463581085, + "vlm_logprob": -0.06415150463581085, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.06415150463581085, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11434315890073776 + ], + "opposite_logprob_tokens": [ + -2.3643431663513184 + ], + "logprob_reward": -0.11434315890073776, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11434315890073776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1777891218662262, + 0.07437795400619507, + 0.0527222640812397, + 0.05317065119743347, + 0.08505736291408539, + 0.0768449455499649, + 0.20899879932403564, + 0.08659953624010086, + 0.09239863604307175, + 0.09204071015119553 + ], + "credit_map": [ + 0.022599345073103905, + 0.01824183575809002, + 0.019403871148824692, + 0.025253983214497566, + 0.027199232950806618, + 0.022772016003727913, + 0.020701149478554726, + 0.02161768637597561, + 0.010036478750407696, + 0.006555076222866774, + 0.0092476112768054, + 0.01390750426799059, + 0.007731572724878788, + 0.008118902333080769, + 0.010766275227069855, + 0.00801453460007906, + 0.006938815116882324, + 0.00845949724316597, + 0.00612533139064908, + 0.005494197364896536, + 0.006317420396953821, + 0.005304743070155382, + 0.009639403782784939, + 0.004442852456122637, + 0.007903818972408772, + 0.005411464720964432, + 0.005391043610870838, + 0.0067714848555624485, + 0.004737559240311384, + 0.0055597927421331406, + 0.010627584531903267, + 0.006767904851585627, + 0.017889421433210373, + 0.0086830398067832, + 0.01404629647731781, + 0.008156755939126015, + 0.010500297881662846, + 0.008447223342955112, + 0.011079623363912106, + 0.00625470420345664, + 0.014577994123101234, + 0.009582027792930603, + 0.011980880051851273, + 0.006299412809312344, + 0.011527140624821186, + 0.00981252919882536, + 0.00788162462413311, + 0.005183338187634945, + 0.051873285323381424, + 0.02568577043712139, + 0.03281271457672119, + 0.02458369918167591, + 0.019572975113987923, + 0.02810710482299328, + 0.016169963404536247, + 0.010193292051553726, + 0.012244590558111668, + 0.007995971478521824, + 0.011575790122151375, + 0.015851765871047974, + 0.007455579470843077, + 0.00939809437841177, + 0.013079791329801083, + 0.008997958153486252, + 0.011818943545222282, + 0.009830474853515625, + 0.012200983241200447, + 0.014303877018392086, + 0.008889290504157543, + 0.011487610638141632, + 0.011242241598665714, + 0.012625215575098991, + 0.01701599545776844, + 0.014579908922314644, + 0.014380154199898243, + 0.0073302933014929295, + 0.006088730879127979, + 0.01131333876401186, + 0.008830045349895954, + 0.012502247467637062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.11434315890073776 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079159453511238 + ], + "opposite_logprob_tokens": [ + -4.141078948974609 + ], + "logprob_reward": -0.016079159453511238, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079159453511238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13181336224079132, + 0.04110903665423393, + 0.06910163164138794, + 0.05491482466459274, + 0.055254098027944565, + 0.09462012350559235, + 0.05040227249264717, + 0.09943288564682007, + 0.08668603748083115, + 0.3166656792163849 + ], + "credit_map": [ + 0.019495587795972824, + 0.013745314441621304, + 0.023418698459863663, + 0.014367274008691311, + 0.012265673838555813, + 0.012702701613307, + 0.02172606810927391, + 0.014092041179537773, + 0.005093222483992577, + 0.0060049123130738735, + 0.004931761417537928, + 0.004726849962025881, + 0.004369401838630438, + 0.004183861892670393, + 0.008738583885133266, + 0.0030604470521211624, + 0.0077402931638062, + 0.005944584961980581, + 0.007835889235138893, + 0.011818514205515385, + 0.007898933254182339, + 0.005815829150378704, + 0.015398938208818436, + 0.006648648530244827, + 0.007933116517961025, + 0.006877315230667591, + 0.006294366903603077, + 0.005656863562762737, + 0.004859030246734619, + 0.0045727952383458614, + 0.009082670323550701, + 0.00963867362588644, + 0.010131839662790298, + 0.004626438952982426, + 0.007131213787943125, + 0.0045306496322155, + 0.00777554651722312, + 0.005158071406185627, + 0.01004830189049244, + 0.0058520399034023285, + 0.014613142237067223, + 0.011106514371931553, + 0.01032213494181633, + 0.006702997721731663, + 0.012037988752126694, + 0.009293729439377785, + 0.01718866638839245, + 0.013354958966374397, + 0.006035510916262865, + 0.004700018092989922, + 0.005432148929685354, + 0.004055991303175688, + 0.00989829283207655, + 0.005851139780133963, + 0.010519047267735004, + 0.003910126630216837, + 0.011656053364276886, + 0.010775353759527206, + 0.011557145975530148, + 0.009758629836142063, + 0.0140156215056777, + 0.012296373024582863, + 0.018695130944252014, + 0.010678578168153763, + 0.016757871955633163, + 0.010494998656213284, + 0.00866986345499754, + 0.011357349343597889, + 0.00826912559568882, + 0.013628990389406681, + 0.011798733845353127, + 0.00570910656824708, + 0.04576246067881584, + 0.034254029393196106, + 0.03935869038105011, + 0.030163554474711418, + 0.03725123777985573, + 0.045992884784936905, + 0.0544164814054966, + 0.029466353356838226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.016079159453511238 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030212998390198 + ], + "opposite_logprob_tokens": [ + -2.350302219390869 + ], + "logprob_reward": -0.10030212998390198, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030212998390198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22322167456150055, + 0.08283860236406326, + 0.07410593330860138, + 0.09747244417667389, + 0.04052869975566864, + 0.07042475044727325, + 0.10819797217845917, + 0.095781609416008, + 0.07584761083126068, + 0.1315806806087494 + ], + "credit_map": [ + 0.025433070957660675, + 0.024307530373334885, + 0.029748842120170593, + 0.0393357016146183, + 0.022936465218663216, + 0.018899744376540184, + 0.03340218588709831, + 0.029158126562833786, + 0.011933722533285618, + 0.007391270250082016, + 0.013675190508365631, + 0.01190523523837328, + 0.007736175321042538, + 0.011236035265028477, + 0.012051103636622429, + 0.006909865420311689, + 0.010888456366956234, + 0.006845243740826845, + 0.009578739292919636, + 0.015155206434428692, + 0.008713452145457268, + 0.007238268852233887, + 0.009251441806554794, + 0.0064351242035627365, + 0.017402857542037964, + 0.011660520918667316, + 0.013307251036167145, + 0.01456616073846817, + 0.010612069629132748, + 0.007660360541194677, + 0.012392695993185043, + 0.009870529174804688, + 0.005413453094661236, + 0.004746418446302414, + 0.005343306809663773, + 0.003675263375043869, + 0.006224273703992367, + 0.0051440224051475525, + 0.005378150846809149, + 0.004603812005370855, + 0.015512061305344105, + 0.009390129707753658, + 0.010301093570888042, + 0.005616694688796997, + 0.007938455790281296, + 0.008182240650057793, + 0.008428133092820644, + 0.0050559453666210175, + 0.0235628429800272, + 0.013169038109481335, + 0.014856110326945782, + 0.011884107254445553, + 0.01504241768270731, + 0.011862131766974926, + 0.011662716045975685, + 0.00615861639380455, + 0.01137307845056057, + 0.011595932766795158, + 0.009358089417219162, + 0.010342921130359173, + 0.01439629215747118, + 0.011308004148304462, + 0.0187250804156065, + 0.008682211861014366, + 0.009794714860618114, + 0.010285857133567333, + 0.007654842920601368, + 0.00937002431601286, + 0.008201143704354763, + 0.008984598331153393, + 0.011824766173958778, + 0.009731660597026348, + 0.019822673872113228, + 0.01612279750406742, + 0.019685206934809685, + 0.010846893303096294, + 0.014103112742304802, + 0.013897793367505074, + 0.020826896652579308, + 0.016275299713015556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.10030212998390198 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067314073443413 + ], + "opposite_logprob_tokens": [ + -4.5110673904418945 + ], + "logprob_reward": -0.011067314073443413, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067314073443413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1296989917755127, + 0.07071865350008011, + 0.05755350738763809, + 0.11038868874311447, + 0.1432052105665207, + 0.11607567965984344, + 0.02901844121515751, + 0.19337712228298187, + 0.06172192096710205, + 0.08824177086353302 + ], + "credit_map": [ + 0.015020953491330147, + 0.014745562337338924, + 0.01356141734868288, + 0.020577631890773773, + 0.012911448255181313, + 0.01792879030108452, + 0.019159536808729172, + 0.01579364761710167, + 0.011664672754704952, + 0.008346357382833958, + 0.00811963900923729, + 0.009594433940947056, + 0.008835348300635815, + 0.009571122005581856, + 0.009290453046560287, + 0.00529663497582078, + 0.007095017936080694, + 0.0045385039411485195, + 0.007257756311446428, + 0.011993428692221642, + 0.006479436531662941, + 0.006748414132744074, + 0.00797248724848032, + 0.005468458868563175, + 0.02404862269759178, + 0.015062708407640457, + 0.013755167834460735, + 0.007824220694601536, + 0.011372268199920654, + 0.013111776672303677, + 0.014527453109622002, + 0.010686475783586502, + 0.025890782475471497, + 0.014787483029067516, + 0.024584539234638214, + 0.015123688615858555, + 0.02081519179046154, + 0.017234664410352707, + 0.015810582786798477, + 0.008958266116678715, + 0.028529314324259758, + 0.013169479556381702, + 0.02007501758635044, + 0.011393316090106964, + 0.014058709144592285, + 0.013655425980687141, + 0.009111101739108562, + 0.006083320360630751, + 0.0036592097021639347, + 0.0032293705735355616, + 0.0036779148504137993, + 0.003953463863581419, + 0.0035921649541705847, + 0.004329914692789316, + 0.004275194834917784, + 0.0023012112360447645, + 0.030671507120132446, + 0.029261911287903786, + 0.02543765679001808, + 0.025990035384893417, + 0.015190145932137966, + 0.031041191890835762, + 0.02074883133172989, + 0.015035826712846756, + 0.011759373359382153, + 0.010417923331260681, + 0.00726940855383873, + 0.006320016458630562, + 0.006710828747600317, + 0.006966705434024334, + 0.006029651500284672, + 0.006248010788112879, + 0.01315287221223116, + 0.009636672213673592, + 0.010304325260221958, + 0.010775339789688587, + 0.008470766246318817, + 0.014958503656089306, + 0.011603502556681633, + 0.009339784272015095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.011067314073443413 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a slowly advancing first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a Volkswagen Type 2 De Luxe bus with a creamy white upper body and sky blue lower body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the bus have a hand-painted matte sheen and is it moving through a turquoise sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there massive puffy white cumulonimbus clouds in the sky casting a warm luminous hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan gently to capture the ocean panorama ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896576076745987 + ], + "opposite_logprob_tokens": [ + -2.578965663909912 + ], + "logprob_reward": -0.07896576076745987, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896576076745987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15588338673114777, + 0.10495097935199738, + 0.06722895801067352, + 0.06319534778594971, + 0.07985099405050278, + 0.03462718054652214, + 0.20759493112564087, + 0.07154479622840881, + 0.12292249500751495, + 0.09220090508460999 + ], + "credit_map": [ + 0.02094370312988758, + 0.019272224977612495, + 0.017299389466643333, + 0.015817955136299133, + 0.01985211856663227, + 0.017926063388586044, + 0.028573887422680855, + 0.016198059543967247, + 0.015732653439044952, + 0.01019218098372221, + 0.015762435272336006, + 0.01724955439567566, + 0.010819299146533012, + 0.00984362605959177, + 0.016971057280898094, + 0.008380187675356865, + 0.011429695412516594, + 0.005456333048641682, + 0.011195546947419643, + 0.010673834010958672, + 0.007198443170636892, + 0.007024014368653297, + 0.00978044979274273, + 0.0044706519693136215, + 0.011829579249024391, + 0.005963248200714588, + 0.009825930930674076, + 0.0061292569153010845, + 0.006247698795050383, + 0.004590792581439018, + 0.010736324824392796, + 0.007872522808611393, + 0.01834530010819435, + 0.008224666118621826, + 0.012203283607959747, + 0.006832528859376907, + 0.00932677648961544, + 0.009318799711763859, + 0.009499134495854378, + 0.006100508850067854, + 0.004963033366948366, + 0.005260903388261795, + 0.004018877167254686, + 0.003477469552308321, + 0.0053936634212732315, + 0.004736873786896467, + 0.003971258644014597, + 0.002805102616548538, + 0.05016376078128815, + 0.024171113967895508, + 0.031714748591184616, + 0.02558056078851223, + 0.02163371816277504, + 0.025367088615894318, + 0.018062913790345192, + 0.010901032015681267, + 0.0086065623909235, + 0.0073470319621264935, + 0.007474834565073252, + 0.011108011938631535, + 0.008415748365223408, + 0.008161424659192562, + 0.014368830248713493, + 0.006062357686460018, + 0.021237187087535858, + 0.015809454023838043, + 0.014499682933092117, + 0.013985938392579556, + 0.013895061798393726, + 0.01591724343597889, + 0.014907246455550194, + 0.01267069112509489, + 0.01559509988874197, + 0.011330443434417248, + 0.011957013048231602, + 0.00898345559835434, + 0.012083887122571468, + 0.012306276708841324, + 0.01190277375280857, + 0.008041963912546635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.07896576076745987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_12/index_tensor([1910, 1910])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_12/index_tensor([1910, 1910])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ee7dc8e2ca172a97b66ca29ada0d6650307203aa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_12/index_tensor([1910, 1910])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide shot begins from an underwater upward perspective, the camera slowly tilting downward to follow a massive humpback whale as it leisurely descends into the depths. Beams of golden jewels (the Tyndall effect) penetrate the shimmering azure water, casting a shifting, shimmering, and shimmering dappled light and shadow across its vast, smooth, blue-black skin. These dapples of light dance gracefully with its slow movements. The beast's hide is covered in white scars from age and sparsely covered with barnacles. Its signature long pectoral fins stretch gracefully like wings through the clear water, barely disturbing the water. As it steadily and leisurely descends, the illuminated world above gradually diminishes, and the surroundings transition from brilliant light blue to a deep indigo. The entire image is imbued with a tranquil, majestic, and awe-inspiring sense of life, creating an atmosphere of grandeur and tranquility.", + "seed": 11095, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_12/index_1910_seed_11095_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.036248343717306855, + "vlm_logprob": -0.036248343717306855, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.036248343717306855, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00561387138441205 + ], + "opposite_logprob_tokens": [ + -6.505613803863525 + ], + "logprob_reward": -0.00561387138441205, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.00561387138441205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14047633111476898, + 0.09392084926366806, + 0.11359488219022751, + 0.052354853600263596, + 0.032520800828933716, + 0.06082858517765999, + 0.06447219848632812, + 0.045140281319618225, + 0.13474352657794952, + 0.2619476914405823 + ], + "credit_map": [ + 0.009452578611671925, + 0.01881580799818039, + 0.012832378968596458, + 0.017134008929133415, + 0.02278056927025318, + 0.023818958550691605, + 0.018037129193544388, + 0.017604900524020195, + 0.007823862135410309, + 0.006874464452266693, + 0.028881456702947617, + 0.023952966555953026, + 0.0071115195751190186, + 0.008376648649573326, + 0.006558053661137819, + 0.00434187101200223, + 0.004066041205078363, + 0.005119615234434605, + 0.04501696676015854, + 0.04292386770248413, + 0.003780781989917159, + 0.0037437835708260536, + 0.005640868563205004, + 0.0033029532060027122, + 0.00506714778020978, + 0.006498245522379875, + 0.01143554411828518, + 0.009105907753109932, + 0.005423859693109989, + 0.004083885811269283, + 0.006847561802715063, + 0.003892698558047414, + 0.002868975279852748, + 0.002976535353809595, + 0.004083576146513224, + 0.004483474418520927, + 0.004211781546473503, + 0.00680151954293251, + 0.004025720525532961, + 0.003069215454161167, + 0.004415770992636681, + 0.00850751530379057, + 0.006318903528153896, + 0.011801589280366898, + 0.009633369743824005, + 0.006558584980666637, + 0.007556305732578039, + 0.006036552134901285, + 0.006417365744709969, + 0.006916300859302282, + 0.009227107279002666, + 0.010371169075369835, + 0.009610752575099468, + 0.006710519082844257, + 0.006392533425241709, + 0.008826450444757938, + 0.004594008903950453, + 0.008008571341633797, + 0.005140862427651882, + 0.00474191177636385, + 0.006310637574642897, + 0.005321586970239878, + 0.003868639934808016, + 0.00715406471863389, + 0.013829219155013561, + 0.016949333250522614, + 0.015479443594813347, + 0.011194449849426746, + 0.015853172168135643, + 0.031037352979183197, + 0.016314614564180374, + 0.014085937291383743, + 0.026154106482863426, + 0.06571486592292786, + 0.022442186251282692, + 0.03355693444609642, + 0.04304685443639755, + 0.027624091133475304, + 0.02096470445394516, + 0.02244396135210991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.00561387138441205 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023330582305788994 + ], + "opposite_logprob_tokens": [ + -3.7733306884765625 + ], + "logprob_reward": -0.023330582305788994, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023330582305788994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09374850243330002, + 0.10237327963113785, + 0.22391153872013092, + 0.06501172482967377, + 0.05785500258207321, + 0.09184033423662186, + 0.09605687856674194, + 0.07117556780576706, + 0.09232401847839355, + 0.10570314526557922 + ], + "credit_map": [ + 0.00659025227651, + 0.012934764847159386, + 0.008197429589927197, + 0.0146315423771739, + 0.013280128128826618, + 0.013511080294847488, + 0.013290423899888992, + 0.011312887072563171, + 0.0071525597013533115, + 0.0077323815785348415, + 0.028406422585248947, + 0.022770347073674202, + 0.009783083572983742, + 0.009276819415390491, + 0.010428844951093197, + 0.006822820287197828, + 0.010622418485581875, + 0.012731869705021381, + 0.07949019223451614, + 0.07157749682664871, + 0.01065913401544094, + 0.01470203697681427, + 0.014847074635326862, + 0.009281305596232414, + 0.004487089812755585, + 0.006085309665650129, + 0.012579898349940777, + 0.011422999203205109, + 0.009435821324586868, + 0.006839579902589321, + 0.008817974478006363, + 0.005343051161617041, + 0.00433421740308404, + 0.004639344289898872, + 0.006978640332818031, + 0.00860447809100151, + 0.008164896629750729, + 0.011206076480448246, + 0.007994876243174076, + 0.0059324693866074085, + 0.007885635830461979, + 0.009650023654103279, + 0.010660379193723202, + 0.015142909251153469, + 0.012503332458436489, + 0.011360335163772106, + 0.01234331727027893, + 0.012294397689402103, + 0.009080419316887856, + 0.00977237056940794, + 0.01672421209514141, + 0.010309615172445774, + 0.01573074609041214, + 0.012005235999822617, + 0.012064228765666485, + 0.010370055213570595, + 0.006779014132916927, + 0.012655301950871944, + 0.00711709912866354, + 0.0062627592124044895, + 0.010718252509832382, + 0.009370948188006878, + 0.006386078894138336, + 0.01188611425459385, + 0.00762183777987957, + 0.019937846809625626, + 0.007091314997524023, + 0.009763865731656551, + 0.013103608042001724, + 0.01127971988171339, + 0.01571238972246647, + 0.007813427597284317, + 0.011904486455023289, + 0.019510431215167046, + 0.011384699493646622, + 0.014792388305068016, + 0.015751348808407784, + 0.011074071750044823, + 0.008787490427494049, + 0.012498227879405022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.023330582305788994 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198708347976208 + ], + "opposite_logprob_tokens": [ + -4.2641987800598145 + ], + "logprob_reward": -0.014198708347976208, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198708347976208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1323874294757843, + 0.17135433852672577, + 0.10953158140182495, + 0.11106055974960327, + 0.06094836816191673, + 0.10814588516950607, + 0.06907328218221664, + 0.08245517313480377, + 0.0691424161195755, + 0.08590096235275269 + ], + "credit_map": [ + 0.014476978220045567, + 0.013882562518119812, + 0.03369705379009247, + 0.02023012563586235, + 0.01335019338876009, + 0.01209228951483965, + 0.01219526119530201, + 0.012462972663342953, + 0.013995486311614513, + 0.023005401715636253, + 0.045533839613199234, + 0.035395145416259766, + 0.012759287841618061, + 0.014256096445024014, + 0.016209354624152184, + 0.010199722833931446, + 0.005541790742427111, + 0.009002933278679848, + 0.030597420409321785, + 0.03133492171764374, + 0.006776990834623575, + 0.012245065532624722, + 0.008256400004029274, + 0.005776054225862026, + 0.011251814663410187, + 0.017864862456917763, + 0.02914765477180481, + 0.023700805380940437, + 0.008616846054792404, + 0.006717409007251263, + 0.007956886664032936, + 0.00580427423119545, + 0.005969409830868244, + 0.007517688442021608, + 0.008181178942322731, + 0.0091920867562294, + 0.006380838807672262, + 0.008643273264169693, + 0.00833016075193882, + 0.006733729504048824, + 0.012177450582385063, + 0.00840829499065876, + 0.01745501719415188, + 0.010032208636403084, + 0.010061335749924183, + 0.02608529105782509, + 0.012955321930348873, + 0.010970965027809143, + 0.005925686564296484, + 0.007568074855953455, + 0.0076315696351230145, + 0.012591221369802952, + 0.009324660524725914, + 0.006587138399481773, + 0.006946021690964699, + 0.012498903088271618, + 0.008221104741096497, + 0.009033508598804474, + 0.015380006283521652, + 0.012539011426270008, + 0.008213702589273453, + 0.010960513725876808, + 0.00910240039229393, + 0.009004922583699226, + 0.004818324465304613, + 0.00976119190454483, + 0.00834276806563139, + 0.009459936991333961, + 0.008228354156017303, + 0.010136296972632408, + 0.010167146101593971, + 0.008228400722146034, + 0.006804619915783405, + 0.01903628744184971, + 0.007388696074485779, + 0.012565310113132, + 0.016556693241000175, + 0.00805673561990261, + 0.008942625485360622, + 0.006549992598593235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.014198708347976208 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074859648942947 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011074859648942947, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074859648942947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08294122666120529, + 0.07016317546367645, + 0.207697331905365, + 0.03461173549294472, + 0.032020725309848785, + 0.04487055912613869, + 0.07844622433185577, + 0.047509294003248215, + 0.3250114321708679, + 0.07672827690839767 + ], + "credit_map": [ + 0.005378510802984238, + 0.013456104323267937, + 0.007319748401641846, + 0.010825930163264275, + 0.01454243902117014, + 0.011235784739255905, + 0.009584746323525906, + 0.010597970336675644, + 0.005113033577799797, + 0.004836692474782467, + 0.02149626985192299, + 0.016769712790846825, + 0.004711315035820007, + 0.006790496874600649, + 0.006522078067064285, + 0.003923576790839434, + 0.008447601459920406, + 0.007843324914574623, + 0.08356226235628128, + 0.07756885141134262, + 0.00617429381236434, + 0.008531611412763596, + 0.009827358648180962, + 0.005742015317082405, + 0.0027695270255208015, + 0.0035959728993475437, + 0.009296255186200142, + 0.007055768743157387, + 0.0034072548151016235, + 0.0023121796548366547, + 0.0031841606833040714, + 0.0029906155541539192, + 0.0031577094923704863, + 0.004110150039196014, + 0.0028763937298208475, + 0.004566564690321684, + 0.005327226594090462, + 0.004129697568714619, + 0.004480169620364904, + 0.003372811945155263, + 0.003949714824557304, + 0.004932789597660303, + 0.007028231862932444, + 0.0055332002229988575, + 0.006033970508724451, + 0.008744337595999241, + 0.005054752808064222, + 0.00359356333501637, + 0.00580316549167037, + 0.008442049846053123, + 0.011882968246936798, + 0.011600185185670853, + 0.013713798485696316, + 0.008468076586723328, + 0.00853282306343317, + 0.010003159753978252, + 0.00568475853651762, + 0.008166472427546978, + 0.005211859475821257, + 0.005680615082383156, + 0.003863064805045724, + 0.0054276143200695515, + 0.004514746833592653, + 0.008960161358118057, + 0.03157364949584007, + 0.03801210597157478, + 0.04242388531565666, + 0.022841352969408035, + 0.030391432344913483, + 0.08603660017251968, + 0.036195769906044006, + 0.03753663972020149, + 0.005719712935388088, + 0.018560878932476044, + 0.005330245941877365, + 0.013150352984666824, + 0.012649109587073326, + 0.006584200542420149, + 0.008363792672753334, + 0.006369982846081257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.011074859648942947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702369689941406 + ], + "opposite_logprob_tokens": [ + -2.127023696899414 + ], + "logprob_reward": -0.12702369689941406, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702369689941406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07490778714418411, + 0.07415501028299332, + 0.46919310092926025, + 0.038518134504556656, + 0.022136539220809937, + 0.025958912447094917, + 0.033827073872089386, + 0.07811261713504791, + 0.14225995540618896, + 0.040930889546871185 + ], + "credit_map": [ + 0.006805763579905033, + 0.013412226922810078, + 0.009686736389994621, + 0.007498683407902718, + 0.011938115581870079, + 0.008450482040643692, + 0.009153279475867748, + 0.00796249695122242, + 0.006213940214365721, + 0.006133401300758123, + 0.028385469689965248, + 0.020753290504217148, + 0.002726050326600671, + 0.0028174258768558502, + 0.0046490696258842945, + 0.002476358087733388, + 0.020352639257907867, + 0.01655314490199089, + 0.20435790717601776, + 0.1880464255809784, + 0.008216500282287598, + 0.00845220685005188, + 0.01512539479881525, + 0.00808886531740427, + 0.0026592237409204245, + 0.0037186965346336365, + 0.009731552563607693, + 0.007864742539823055, + 0.004457387141883373, + 0.0029510436579585075, + 0.003913423046469688, + 0.0032220659777522087, + 0.0019817331340163946, + 0.0025781793519854546, + 0.0027158199809491634, + 0.003212342271581292, + 0.003227333538234234, + 0.002973882481455803, + 0.0030019129626452923, + 0.0024453343357890844, + 0.0018302056705579162, + 0.0028529963456094265, + 0.0031244533602148294, + 0.004136686213314533, + 0.00447266548871994, + 0.0038346387445926666, + 0.002961598103865981, + 0.0027456681709736586, + 0.0037536846939474344, + 0.003974388353526592, + 0.004369252827018499, + 0.004627674352377653, + 0.004982760176062584, + 0.003626360557973385, + 0.003658896079286933, + 0.004834055434912443, + 0.005940644070506096, + 0.012862743809819221, + 0.009270226582884789, + 0.006756626535207033, + 0.01044356357306242, + 0.008874856866896152, + 0.008127166889607906, + 0.015836790204048157, + 0.010290264151990414, + 0.021052759140729904, + 0.01563052088022232, + 0.012482348829507828, + 0.016994323581457138, + 0.03039497323334217, + 0.019314205273985863, + 0.016100559383630753, + 0.0037224693223834038, + 0.007818684913218021, + 0.004536445252597332, + 0.005736179184168577, + 0.004724315367639065, + 0.0046369340270757675, + 0.0035427920520305634, + 0.0062130712904036045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.12702369689941406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide shot begins from an underwater upward perspective, the camera slowly tilting downward to follow a massive humpback whale as it leisurely descends into the depths. Beams of golden jewels (the Tyndall effect) penetrate the shimmering azure water, casting a shifting, shimmering, and shimmering dappled light and shadow across its vast, smooth, blue-black skin. These dapples of light dance gracefully with its slow movements. The beast's hide is covered in white scars from age and sparsely covered with barnacles. Its signature long pectoral fins stretch gracefully like wings through the clear water, barely disturbing the water. As it steadily and leisurely descends, the illuminated world above gradually diminishes, and the surroundings transition from brilliant light blue to a deep indigo. The entire image is imbued with a tranquil, majestic, and awe-inspiring sense of life, creating an atmosphere of grandeur and tranquility.", + "seed": 11096, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_12/index_1910_seed_11096_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025506847631186248, + "vlm_logprob": -0.025506847631186248, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.025506847631186248, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0033143609762191772 + ], + "opposite_logprob_tokens": [ + -7.62831449508667 + ], + "logprob_reward": -0.0033143609762191772, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0033143609762191772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17766231298446655, + 0.12347230315208435, + 0.06144873797893524, + 0.04459812119603157, + 0.05036506801843643, + 0.07567177712917328, + 0.06950496882200241, + 0.11608197540044785, + 0.08728396892547607, + 0.19391079246997833 + ], + "credit_map": [ + 0.020912954583764076, + 0.021161360666155815, + 0.014653787948191166, + 0.04080893099308014, + 0.016002269461750984, + 0.017820464447140694, + 0.02073243260383606, + 0.02557011879980564, + 0.013687942177057266, + 0.011876578442752361, + 0.013594093732535839, + 0.023008350282907486, + 0.013460411690175533, + 0.01557754073292017, + 0.010667132213711739, + 0.021600252017378807, + 0.004300420172512531, + 0.006337176542729139, + 0.013486577197909355, + 0.009919092059135437, + 0.0067026689648628235, + 0.005950626451522112, + 0.007754921447485685, + 0.006997253745794296, + 0.00323952897451818, + 0.006280002649873495, + 0.007432496175169945, + 0.007732086814939976, + 0.003964647185057402, + 0.006408342160284519, + 0.003946186043322086, + 0.005594830960035324, + 0.00442493986338377, + 0.008235232904553413, + 0.005706657189875841, + 0.004711712710559368, + 0.00978512316942215, + 0.006345535162836313, + 0.006024595815688372, + 0.005131271202117205, + 0.0049569979310035706, + 0.01306814793497324, + 0.007680284790694714, + 0.008957665413618088, + 0.012699681334197521, + 0.01034140307456255, + 0.01075485348701477, + 0.0072127440944314, + 0.0076164850033819675, + 0.011021027341485023, + 0.0077115558087825775, + 0.012110564857721329, + 0.007171077653765678, + 0.008500155992805958, + 0.006478326395153999, + 0.008895778097212315, + 0.013709226623177528, + 0.03443082422018051, + 0.014219826087355614, + 0.011601927690207958, + 0.008177080191671848, + 0.012852896004915237, + 0.008302251808345318, + 0.012787949293851852, + 0.007137082517147064, + 0.005720422137528658, + 0.01899990625679493, + 0.007949626073241234, + 0.005595455411821604, + 0.01979273371398449, + 0.014756908640265465, + 0.007331835571676493, + 0.01702306605875492, + 0.01584085449576378, + 0.03236161544919014, + 0.023291485384106636, + 0.018968159332871437, + 0.02464897930622101, + 0.03620890900492668, + 0.02556771971285343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.0033143609762191772 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020625397562980652 + ], + "opposite_logprob_tokens": [ + -3.895625352859497 + ], + "logprob_reward": -0.020625397562980652, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020625397562980652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25948426127433777, + 0.12007360905408859, + 0.09703046083450317, + 0.1648283302783966, + 0.07303071767091751, + 0.06871231645345688, + 0.029330817982554436, + 0.050787460058927536, + 0.03151373565196991, + 0.10520832240581512 + ], + "credit_map": [ + 0.02656150981783867, + 0.022410811856389046, + 0.020940711721777916, + 0.053022030740976334, + 0.03351258486509323, + 0.04553750157356262, + 0.025636157020926476, + 0.031862951815128326, + 0.009137538261711597, + 0.011508874595165253, + 0.01756218448281288, + 0.01889408566057682, + 0.018467379733920097, + 0.01688048057258129, + 0.014502959325909615, + 0.01312009897083044, + 0.0069905491545796394, + 0.004385245963931084, + 0.0296862181276083, + 0.00877920351922512, + 0.005771405063569546, + 0.00922587513923645, + 0.005525590851902962, + 0.026666371151804924, + 0.013684708625078201, + 0.023272451013326645, + 0.03476710990071297, + 0.016650399193167686, + 0.016398712992668152, + 0.022798404097557068, + 0.016505245119333267, + 0.020751308649778366, + 0.005650361068546772, + 0.006631260737776756, + 0.011372745037078857, + 0.006500035524368286, + 0.016226816922426224, + 0.013475845567882061, + 0.006225832272320986, + 0.006947821471840143, + 0.004612347576767206, + 0.006318770349025726, + 0.006708170752972364, + 0.0075562987476587296, + 0.010308484546840191, + 0.01506104227155447, + 0.010034282691776752, + 0.008112918585538864, + 0.002342933788895607, + 0.0026506264694035053, + 0.004692738875746727, + 0.003991149365901947, + 0.003192777978256345, + 0.004789931699633598, + 0.002016992773860693, + 0.005653667263686657, + 0.003003884805366397, + 0.00576234282925725, + 0.005699352826923132, + 0.006969223730266094, + 0.005112830083817244, + 0.006985194981098175, + 0.006995401345193386, + 0.010259230621159077, + 0.0018409003969281912, + 0.0029406119138002396, + 0.00594861526042223, + 0.003509353846311569, + 0.002669860143214464, + 0.005940019153058529, + 0.004919033031910658, + 0.003745344001799822, + 0.010960105806589127, + 0.009786995127797127, + 0.017231453210115433, + 0.013006237335503101, + 0.009945547208189964, + 0.012372820638120174, + 0.018776539713144302, + 0.013128623366355896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.020625397562980652 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949999671429396 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.005949999671429396, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949999671429396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45999875664711, + 0.05983075127005577, + 0.202565997838974, + 0.06976914405822754, + 0.04013723134994507, + 0.0362335667014122, + 0.025957027450203896, + 0.05005897209048271, + 0.03235262632369995, + 0.023095926269888878 + ], + "credit_map": [ + 0.056072451174259186, + 0.057437025010585785, + 0.04333820939064026, + 0.09641415625810623, + 0.03324482962489128, + 0.035209085792303085, + 0.05410677194595337, + 0.0841762125492096, + 0.005413896404206753, + 0.00396130234003067, + 0.005007447674870491, + 0.007307338993996382, + 0.008374620229005814, + 0.010233310982584953, + 0.009453889913856983, + 0.010078942403197289, + 0.017222512513399124, + 0.004691842012107372, + 0.07932701706886292, + 0.014413735829293728, + 0.010927406139671803, + 0.03384530544281006, + 0.006696383003145456, + 0.03544178977608681, + 0.006097862496972084, + 0.01031570602208376, + 0.010640380904078484, + 0.006129641551524401, + 0.007104534190148115, + 0.008517889305949211, + 0.010141052305698395, + 0.01082207728177309, + 0.005001728422939777, + 0.004108126275241375, + 0.00538237951695919, + 0.003980325534939766, + 0.00755460886284709, + 0.005752176977694035, + 0.004101094789803028, + 0.004256790969520807, + 0.0028739587869495153, + 0.005487686954438686, + 0.0036641659680753946, + 0.0027571721002459526, + 0.004374362528324127, + 0.0072496673092246056, + 0.006018155254423618, + 0.0038084001280367374, + 0.0029846071265637875, + 0.0036452538333833218, + 0.0037544379010796547, + 0.003382713533937931, + 0.003520508296787739, + 0.0027759866788983345, + 0.0023306277580559254, + 0.0035628918558359146, + 0.0037541899364441633, + 0.009652463719248772, + 0.005157796666026115, + 0.0050348457880318165, + 0.005712190642952919, + 0.005411328282207251, + 0.006510809995234013, + 0.008825348690152168, + 0.0026302863843739033, + 0.003003388876095414, + 0.004575861617922783, + 0.003630381077528, + 0.003715612692758441, + 0.004544182680547237, + 0.004911509342491627, + 0.005341402720659971, + 0.0023939546663314104, + 0.0023179883137345314, + 0.003475374076515436, + 0.002257060259580612, + 0.0025277582462877035, + 0.0030026687309145927, + 0.00420802179723978, + 0.002913100179284811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.005949999671429396 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635440841317177 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635440841317177, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635440841317177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1532503217458725, + 0.10435983538627625, + 0.15579962730407715, + 0.03210524097084999, + 0.04763079062104225, + 0.08054055273532867, + 0.07597154378890991, + 0.225900337100029, + 0.09327910840511322, + 0.031162653118371964 + ], + "credit_map": [ + 0.024754445999860764, + 0.016830813139677048, + 0.012615279294550419, + 0.021227136254310608, + 0.01657705195248127, + 0.02400767244398594, + 0.010991443879902363, + 0.026246480643749237, + 0.011697757989168167, + 0.008477995172142982, + 0.010298908688127995, + 0.01504715345799923, + 0.009052468463778496, + 0.014677822589874268, + 0.009659312665462494, + 0.025448424741625786, + 0.01327801588922739, + 0.008108112029731274, + 0.054739706218242645, + 0.01245855912566185, + 0.01057322695851326, + 0.028589587658643723, + 0.007631372660398483, + 0.02042103186249733, + 0.0024111217353492975, + 0.003593981731683016, + 0.007041278760880232, + 0.0034277578815817833, + 0.003962481394410133, + 0.0048998682759702206, + 0.0023779766634106636, + 0.004390774294734001, + 0.004442387726157904, + 0.0046125492081046104, + 0.005106809549033642, + 0.0046142516657710075, + 0.010171541012823582, + 0.00801151618361473, + 0.006224285811185837, + 0.004447454120963812, + 0.005013944115489721, + 0.009662432596087456, + 0.007389795500785112, + 0.008426266722381115, + 0.011026343330740929, + 0.014150536619126797, + 0.016219455748796463, + 0.008651779033243656, + 0.008086266927421093, + 0.010337243787944317, + 0.011509772390127182, + 0.013470087200403214, + 0.00839577242732048, + 0.009805507957935333, + 0.007187406066805124, + 0.007179491687566042, + 0.016353217884898186, + 0.0235599372535944, + 0.02742275968194008, + 0.026047255843877792, + 0.02317897230386734, + 0.030986925587058067, + 0.03231362625956535, + 0.04603764787316322, + 0.009912537410855293, + 0.012015413492918015, + 0.010422274470329285, + 0.008760998956859112, + 0.010909141972661018, + 0.010101576335728168, + 0.012803718447685242, + 0.018353445455431938, + 0.0030706538818776608, + 0.0034297662787139416, + 0.005807808600366116, + 0.0030100068543106318, + 0.003447695169597864, + 0.00436003552749753, + 0.004029503557831049, + 0.0040071820840239525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.008635440841317177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900903910398483 + ], + "opposite_logprob_tokens": [ + -2.4640090465545654 + ], + "logprob_reward": -0.08900903910398483, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900903910398483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18449494242668152, + 0.08273541182279587, + 0.29121872782707214, + 0.05677316337823868, + 0.050849299877882004, + 0.0913032591342926, + 0.13543841242790222, + 0.03204336389899254, + 0.034864991903305054, + 0.040278367698192596 + ], + "credit_map": [ + 0.023865843191742897, + 0.023029964417219162, + 0.02395942434668541, + 0.026578037068247795, + 0.014187618158757687, + 0.02049957774579525, + 0.01636572554707527, + 0.03600876033306122, + 0.006523867603391409, + 0.01190948486328125, + 0.008942220360040665, + 0.01845250278711319, + 0.009051836095750332, + 0.010194630362093449, + 0.009483102709054947, + 0.008177769370377064, + 0.025931332260370255, + 0.013728257268667221, + 0.10523926466703415, + 0.022282881662249565, + 0.01953834854066372, + 0.053716592490673065, + 0.012033503502607346, + 0.03874856233596802, + 0.004296062048524618, + 0.007733368780463934, + 0.011555769480764866, + 0.0054382141679525375, + 0.005806801840662956, + 0.007775343954563141, + 0.006906981114298105, + 0.007260623387992382, + 0.003535842290148139, + 0.004888907540589571, + 0.008192959241569042, + 0.004525232128798962, + 0.009091846644878387, + 0.006997579243034124, + 0.007702595088630915, + 0.00591433746740222, + 0.006770974025130272, + 0.010660561732947826, + 0.00913547445088625, + 0.008694183081388474, + 0.014283443801105022, + 0.014797340147197247, + 0.017097890377044678, + 0.009863387793302536, + 0.015071661211550236, + 0.022189563140273094, + 0.015619208104908466, + 0.015098833478987217, + 0.02128598652780056, + 0.015446560457348824, + 0.01572849415242672, + 0.014998107217252254, + 0.0026887254789471626, + 0.002840290544554591, + 0.004027926828712225, + 0.005044793244451284, + 0.0042263842187821865, + 0.003199540311470628, + 0.004451530519872904, + 0.005564174614846706, + 0.00342397578060627, + 0.0069132172502577305, + 0.0035070362500846386, + 0.0034557925537228584, + 0.002888578688725829, + 0.0037914100103080273, + 0.004368182271718979, + 0.006516803056001663, + 0.004812608007341623, + 0.0037901538889855146, + 0.010084751062095165, + 0.0032893652096390724, + 0.003366404678672552, + 0.005557458382099867, + 0.0046083396300673485, + 0.0047692847438156605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.08900903910398483 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide shot begins from an underwater upward perspective, the camera slowly tilting downward to follow a massive humpback whale as it leisurely descends into the depths. Beams of golden jewels (the Tyndall effect) penetrate the shimmering azure water, casting a shifting, shimmering, and shimmering dappled light and shadow across its vast, smooth, blue-black skin. These dapples of light dance gracefully with its slow movements. The beast's hide is covered in white scars from age and sparsely covered with barnacles. Its signature long pectoral fins stretch gracefully like wings through the clear water, barely disturbing the water. As it steadily and leisurely descends, the illuminated world above gradually diminishes, and the surroundings transition from brilliant light blue to a deep indigo. The entire image is imbued with a tranquil, majestic, and awe-inspiring sense of life, creating an atmosphere of grandeur and tranquility.", + "seed": 11097, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_12/index_1910_seed_11097_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04865239448845386, + "vlm_logprob": -0.04865239448845386, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.04865239448845386, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0027581527829170227 + ], + "opposite_logprob_tokens": [ + -7.127758026123047 + ], + "logprob_reward": -0.0027581527829170227, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0027581527829170227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07054675370454788, + 0.07858578860759735, + 0.11219889670610428, + 0.07062402367591858, + 0.03857089951634407, + 0.09602237492799759, + 0.0980457216501236, + 0.04124634340405464, + 0.34044718742370605, + 0.053712017834186554 + ], + "credit_map": [ + 0.006576839368790388, + 0.010033623315393925, + 0.008111084811389446, + 0.009802318178117275, + 0.013434589840471745, + 0.00938472244888544, + 0.007668917067348957, + 0.005534662399441004, + 0.007536984048783779, + 0.009128391742706299, + 0.007890814915299416, + 0.009617957286536694, + 0.017270857468247414, + 0.010696676559746265, + 0.0093193044885993, + 0.0071247960440814495, + 0.012526457197964191, + 0.015695936977863312, + 0.010377888567745686, + 0.014054392464458942, + 0.01841275952756405, + 0.01415549498051405, + 0.011260945349931717, + 0.01571502536535263, + 0.005323965102434158, + 0.008611960336565971, + 0.007331589702516794, + 0.011000427417457104, + 0.011535202153027058, + 0.011343160644173622, + 0.008046403527259827, + 0.007431304547935724, + 0.003929855301976204, + 0.005328577943146229, + 0.004651384428143501, + 0.006343638524413109, + 0.003159276442602277, + 0.004444831516593695, + 0.004736191127449274, + 0.005977143999189138, + 0.011015155352652073, + 0.011310641653835773, + 0.013263129629194736, + 0.016748059540987015, + 0.007588027976453304, + 0.011362170800566673, + 0.01214034017175436, + 0.012594847939908504, + 0.010254885070025921, + 0.014358480460941792, + 0.005851774010807276, + 0.011026302352547646, + 0.019390610978007317, + 0.014641647227108479, + 0.012594920583069324, + 0.009927100501954556, + 0.0039319563657045364, + 0.005177807062864304, + 0.005391502287238836, + 0.004739031661301851, + 0.00425884360447526, + 0.007286099251359701, + 0.004542235285043716, + 0.005918866489082575, + 0.013492959551513195, + 0.014260995201766491, + 0.018012413755059242, + 0.014568854123353958, + 0.1372147798538208, + 0.06240050122141838, + 0.016210325062274933, + 0.06428636610507965, + 0.004330878145992756, + 0.003601177129894495, + 0.004719614516943693, + 0.00497537013143301, + 0.012695473618805408, + 0.013877799734473228, + 0.005454278085380793, + 0.0040574269369244576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.0027581527829170227 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638811059296131 + ], + "opposite_logprob_tokens": [ + -3.651388168334961 + ], + "logprob_reward": -0.02638811059296131, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638811059296131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10957504063844681, + 0.052649445831775665, + 0.20943062007427216, + 0.09115435183048248, + 0.0534263551235199, + 0.1981872320175171, + 0.04187720641493797, + 0.06278969347476959, + 0.04892430827021599, + 0.13198572397232056 + ], + "credit_map": [ + 0.0079274857416749, + 0.015918325632810593, + 0.012485145591199398, + 0.015202166512608528, + 0.025837106630206108, + 0.015854788944125175, + 0.009741628542542458, + 0.0066083925776183605, + 0.004387245047837496, + 0.004762270022183657, + 0.008706382475793362, + 0.005362542811781168, + 0.008218254894018173, + 0.0069201928563416, + 0.008024854585528374, + 0.006267701741307974, + 0.011126965284347534, + 0.014049599878489971, + 0.021341269835829735, + 0.040857281535863876, + 0.025695012882351875, + 0.01868566684424877, + 0.038665320724248886, + 0.03900950402021408, + 0.007566845044493675, + 0.009019879624247551, + 0.012824748642742634, + 0.019267981871962547, + 0.011172754690051079, + 0.010295411571860313, + 0.010208622552454472, + 0.010798106901347637, + 0.0065978397615253925, + 0.005139231216162443, + 0.00688294367864728, + 0.00884956493973732, + 0.004872748162597418, + 0.005985961761325598, + 0.006943260319530964, + 0.00815480574965477, + 0.024088041856884956, + 0.018384834751486778, + 0.02710813097655773, + 0.029765233397483826, + 0.017169596627354622, + 0.027666795998811722, + 0.030356252565979958, + 0.023648347705602646, + 0.003580569289624691, + 0.004869360011070967, + 0.004326961934566498, + 0.005435916129499674, + 0.0056612431071698666, + 0.005187241360545158, + 0.007414027117192745, + 0.005401886999607086, + 0.004827581811696291, + 0.005726825445890427, + 0.009725109674036503, + 0.00944710336625576, + 0.006976200267672539, + 0.010772029869258404, + 0.008966967463493347, + 0.0063478779047727585, + 0.003197413869202137, + 0.00521782785654068, + 0.0037545873783528805, + 0.004995080176740885, + 0.01053050346672535, + 0.005810215603560209, + 0.003700275905430317, + 0.011718404479324818, + 0.008365059271454811, + 0.01140429824590683, + 0.009890764951705933, + 0.016752824187278748, + 0.028751682490110397, + 0.0333985909819603, + 0.011196578852832317, + 0.012225928716361523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 9 + ], + "selected_reward": -0.02638811059296131 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073528677225113 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073528677225113, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073528677225113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19704881310462952, + 0.09609786421060562, + 0.06884414702653885, + 0.08413437753915787, + 0.1573512852191925, + 0.08588128536939621, + 0.05599358677864075, + 0.025612521916627884, + 0.13062934577465057, + 0.09840674698352814 + ], + "credit_map": [ + 0.017432404682040215, + 0.0351828970015049, + 0.01873568259179592, + 0.02974734827876091, + 0.03933532536029816, + 0.026406304910779, + 0.017840318381786346, + 0.012368534691631794, + 0.00845731794834137, + 0.011984167620539665, + 0.01082151010632515, + 0.010857785120606422, + 0.01742546074092388, + 0.01087493821978569, + 0.015028025954961777, + 0.010648656636476517, + 0.004995383787900209, + 0.006587110459804535, + 0.005337883252650499, + 0.01040589902549982, + 0.009790082462131977, + 0.006649016868323088, + 0.011076523922383785, + 0.014002248644828796, + 0.007194195874035358, + 0.013957126997411251, + 0.008363260887563229, + 0.010615095496177673, + 0.016130292788147926, + 0.010471241548657417, + 0.009636636823415756, + 0.007766527123749256, + 0.016280220821499825, + 0.0243967417627573, + 0.014397602528333664, + 0.020722955465316772, + 0.014819033443927765, + 0.022835221141576767, + 0.019596152007579803, + 0.024303369224071503, + 0.008985737338662148, + 0.009395497851073742, + 0.011272115632891655, + 0.01363284606486559, + 0.009269661270081997, + 0.009789553470909595, + 0.011393489316105843, + 0.012142387218773365, + 0.00449055852368474, + 0.0067530786618590355, + 0.004921294283121824, + 0.007225526962429285, + 0.008729222230613232, + 0.007891488261520863, + 0.007868052460253239, + 0.00811437051743269, + 0.0024383491836488247, + 0.0031562813092023134, + 0.0021190757397562265, + 0.004320982843637466, + 0.0031560747884213924, + 0.002491256920620799, + 0.004485492128878832, + 0.003445009933784604, + 0.0070297596976161, + 0.008201725780963898, + 0.010667414404451847, + 0.009099187329411507, + 0.03842439129948616, + 0.0225093774497509, + 0.013937056064605713, + 0.020760424435138702, + 0.006561995483934879, + 0.012125824578106403, + 0.007969542406499386, + 0.009358801878988743, + 0.02172355353832245, + 0.021373821422457695, + 0.008733558468520641, + 0.010559648275375366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.016073528677225113 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184494227170944 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.018184494227170944, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184494227170944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09950199723243713, + 0.043957605957984924, + 0.1750512719154358, + 0.07990020513534546, + 0.05659688636660576, + 0.10073725134134293, + 0.11400867253541946, + 0.08938055485486984, + 0.16128511726856232, + 0.07958047091960907 + ], + "credit_map": [ + 0.007497468497604132, + 0.01277854572981596, + 0.011908046901226044, + 0.012292031198740005, + 0.027176272124052048, + 0.01394535880535841, + 0.007245410233736038, + 0.006658864673227072, + 0.004252688493579626, + 0.003975686617195606, + 0.006124530453234911, + 0.004883716348558664, + 0.009040992707014084, + 0.006628885865211487, + 0.004555575083941221, + 0.004495527129620314, + 0.012933229096233845, + 0.011956067755818367, + 0.017919596284627914, + 0.029358208179473877, + 0.019397148862481117, + 0.01670725643634796, + 0.03050212189555168, + 0.036277640610933304, + 0.0065912241116166115, + 0.007799540646374226, + 0.011192944832146168, + 0.015785913914442062, + 0.00967459287494421, + 0.008380375802516937, + 0.010606827214360237, + 0.009868786670267582, + 0.00741686811670661, + 0.005432491190731525, + 0.006186234764754772, + 0.00955877173691988, + 0.004949552472680807, + 0.007112965919077396, + 0.008150606416165829, + 0.007789392489939928, + 0.01026115007698536, + 0.010022017173469067, + 0.014607555232942104, + 0.015466700308024883, + 0.009489179588854313, + 0.013560635969042778, + 0.013105117715895176, + 0.014224893413484097, + 0.01095192413777113, + 0.017869479954242706, + 0.007246606517583132, + 0.012540939263999462, + 0.01814061403274536, + 0.016606176272034645, + 0.01579301245510578, + 0.014859912917017937, + 0.007131201680749655, + 0.006709028501063585, + 0.011896442621946335, + 0.009389660321176052, + 0.01089226733893156, + 0.02541545033454895, + 0.008921370841562748, + 0.009025128558278084, + 0.009299332275986671, + 0.015326610766351223, + 0.021458154544234276, + 0.013425981625914574, + 0.03608096390962601, + 0.028532518073916435, + 0.021832160651683807, + 0.015329403802752495, + 0.007251653354614973, + 0.007494890131056309, + 0.0058503239415585995, + 0.007770571857690811, + 0.01870875246822834, + 0.019118620082736015, + 0.005984821356832981, + 0.0074008372612297535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.018184494227170944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985768616199493 + ], + "opposite_logprob_tokens": [ + -1.8048577308654785 + ], + "logprob_reward": -0.17985768616199493, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985768616199493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1389058232307434, + 0.08113370835781097, + 0.09076522290706635, + 0.028594262897968292, + 0.10421493649482727, + 0.15705129504203796, + 0.061558615416288376, + 0.1469714641571045, + 0.1612362116575241, + 0.029568424448370934 + ], + "credit_map": [ + 0.011217496357858181, + 0.01998394913971424, + 0.016263149678707123, + 0.013241984881460667, + 0.03992878273129463, + 0.021015258505940437, + 0.009820660576224327, + 0.007434528321027756, + 0.006869632285088301, + 0.006649329327046871, + 0.01016394142061472, + 0.008799823001027107, + 0.015334502793848515, + 0.009125693701207638, + 0.01456573698669672, + 0.009625055827200413, + 0.004878371488302946, + 0.007039480842649937, + 0.008990596979856491, + 0.015973567962646484, + 0.014534888789057732, + 0.008955386467278004, + 0.014773966744542122, + 0.015618961304426193, + 0.002662578597664833, + 0.0029982721898704767, + 0.003617506707087159, + 0.00683547044172883, + 0.002677198266610503, + 0.0028404586482793093, + 0.003472429234534502, + 0.003490351140499115, + 0.012768935412168503, + 0.01093997061252594, + 0.015042286366224289, + 0.020659133791923523, + 0.006801433861255646, + 0.009944457560777664, + 0.014919289387762547, + 0.013139431364834309, + 0.018002158030867577, + 0.014374385587871075, + 0.023955659940838814, + 0.024202270433306694, + 0.014662738889455795, + 0.021169103682041168, + 0.01939617656171322, + 0.021288804709911346, + 0.0063278693705797195, + 0.006927727721631527, + 0.007199740502983332, + 0.006625915411859751, + 0.009730386547744274, + 0.008180264383554459, + 0.0071204062551259995, + 0.009446308948099613, + 0.012207357212901115, + 0.011976971291005611, + 0.020778562873601913, + 0.01628805696964264, + 0.01927538774907589, + 0.03224403038620949, + 0.01810428686439991, + 0.016096824780106544, + 0.007007466163486242, + 0.01006940845400095, + 0.01687074825167656, + 0.009183699265122414, + 0.0471239909529686, + 0.024353960528969765, + 0.016286486759781837, + 0.030340466648340225, + 0.0021356220822781324, + 0.0026202742010354996, + 0.0023878479842096567, + 0.0020627821795642376, + 0.007440425921231508, + 0.007533090654760599, + 0.0023434953764081, + 0.0030448848847299814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.17985768616199493 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide shot begins from an underwater upward perspective, the camera slowly tilting downward to follow a massive humpback whale as it leisurely descends into the depths. Beams of golden jewels (the Tyndall effect) penetrate the shimmering azure water, casting a shifting, shimmering, and shimmering dappled light and shadow across its vast, smooth, blue-black skin. These dapples of light dance gracefully with its slow movements. The beast's hide is covered in white scars from age and sparsely covered with barnacles. Its signature long pectoral fins stretch gracefully like wings through the clear water, barely disturbing the water. As it steadily and leisurely descends, the illuminated world above gradually diminishes, and the surroundings transition from brilliant light blue to a deep indigo. The entire image is imbued with a tranquil, majestic, and awe-inspiring sense of life, creating an atmosphere of grandeur and tranquility.", + "seed": 11098, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_12/index_1910_seed_11098_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020478158490732312, + "vlm_logprob": -0.020478158490732312, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.020478158490732312, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0038636098615825176 + ], + "opposite_logprob_tokens": [ + -6.87886381149292 + ], + "logprob_reward": -0.0038636098615825176, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0038636098615825176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11318199336528778, + 0.3069536089897156, + 0.04828955605626106, + 0.02932627499103546, + 0.1852540820837021, + 0.05122901499271393, + 0.07378613948822021, + 0.040621303021907806, + 0.08223293721675873, + 0.06912517547607422 + ], + "credit_map": [ + 0.006341311614960432, + 0.007011775858700275, + 0.011705066077411175, + 0.019576268270611763, + 0.006464151665568352, + 0.012076640501618385, + 0.03699835389852524, + 0.013008431531488895, + 0.05917768552899361, + 0.09697578102350235, + 0.02839766815304756, + 0.0246418546885252, + 0.0352589376270771, + 0.013413545675575733, + 0.017662178725004196, + 0.03142595663666725, + 0.004838724620640278, + 0.003758420469239354, + 0.00417165644466877, + 0.005749586038291454, + 0.005196281708776951, + 0.007383076008409262, + 0.008665991015732288, + 0.008525820448994637, + 0.002874576486647129, + 0.0024384569842368364, + 0.002877353224903345, + 0.004182401578873396, + 0.0038791580591350794, + 0.00464597437530756, + 0.00499998964369297, + 0.0034283646382391453, + 0.025974949821829796, + 0.017306679859757423, + 0.011575810611248016, + 0.016203919425606728, + 0.013273945078253746, + 0.018976666033267975, + 0.03828740119934082, + 0.043654706329107285, + 0.003637016983702779, + 0.0033236502204090357, + 0.002821184927597642, + 0.005780541338026524, + 0.0025089902337640524, + 0.0059653292410075665, + 0.013845459558069706, + 0.01334683783352375, + 0.004074529279023409, + 0.004674283787608147, + 0.003924532327800989, + 0.008726445026695728, + 0.011923812329769135, + 0.022806618362665176, + 0.010561532340943813, + 0.007094386965036392, + 0.002926873741671443, + 0.00489299651235342, + 0.0028909260872751474, + 0.005055370274931192, + 0.009910878725349903, + 0.004726465791463852, + 0.005652164574712515, + 0.004565624985843897, + 0.007131192833185196, + 0.010689864866435528, + 0.00534459762275219, + 0.009197431616485119, + 0.010558894835412502, + 0.017888491973280907, + 0.0085920924320817, + 0.012830367311835289, + 0.004832330159842968, + 0.004559606779366732, + 0.0032167036551982164, + 0.005988997872918844, + 0.01424403116106987, + 0.018120450899004936, + 0.009973469190299511, + 0.008189581334590912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0038636098615825176 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822569966316223 + ], + "opposite_logprob_tokens": [ + -4.01822566986084 + ], + "logprob_reward": -0.01822569966316223, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822569966316223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07072148472070694, + 0.09348803013563156, + 0.029924621805548668, + 0.08519691973924637, + 0.29609036445617676, + 0.03847257420420647, + 0.08594431728124619, + 0.20930296182632446, + 0.03642716258764267, + 0.05443158000707626 + ], + "credit_map": [ + 0.004315425641834736, + 0.006176244001835585, + 0.008117490448057652, + 0.018407559022307396, + 0.005389951169490814, + 0.0060366131365299225, + 0.014397102408111095, + 0.007881094701588154, + 0.008168082684278488, + 0.009319756180047989, + 0.013716530986130238, + 0.00820332020521164, + 0.0064714215695858, + 0.007273889146745205, + 0.01276516355574131, + 0.027569865807890892, + 0.003642383264377713, + 0.0047557540237903595, + 0.002859819447621703, + 0.006085236091166735, + 0.002888859249651432, + 0.0027989184018224478, + 0.0037426643539220095, + 0.003150986973196268, + 0.011328508146107197, + 0.028420208021998405, + 0.004801945295184851, + 0.0067248838022351265, + 0.01525599230080843, + 0.004715844057500362, + 0.005564497783780098, + 0.008385036140680313, + 0.04133767634630203, + 0.02411186695098877, + 0.017566733062267303, + 0.027256987988948822, + 0.01791575737297535, + 0.028196072205901146, + 0.0646195113658905, + 0.07508572936058044, + 0.00278635136783123, + 0.0036036204546689987, + 0.002563133370131254, + 0.005240547005087137, + 0.0031704744324088097, + 0.005225064232945442, + 0.007909653708338737, + 0.007973725907504559, + 0.004522533155977726, + 0.005129816941916943, + 0.00544313620775938, + 0.013954048044979572, + 0.016815269365906715, + 0.020712049677968025, + 0.011258814483880997, + 0.008108649402856827, + 0.015441522933542728, + 0.022940943017601967, + 0.015330567955970764, + 0.0228690505027771, + 0.04879432171583176, + 0.026820072904229164, + 0.030766349285840988, + 0.02634013257920742, + 0.0027460427954792976, + 0.002959707984700799, + 0.003068731166422367, + 0.004083109553903341, + 0.007554356940090656, + 0.006381521001458168, + 0.004395951051265001, + 0.005237739998847246, + 0.0027909635100513697, + 0.0033557938877493143, + 0.002386778127402067, + 0.005710534751415253, + 0.014254460111260414, + 0.009958470240235329, + 0.00775162223726511, + 0.00822295993566513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.01822569966316223 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002202943665906787 + ], + "opposite_logprob_tokens": [ + -6.127202987670898 + ], + "logprob_reward": -0.002202943665906787, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002202943665906787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09985329210758209, + 0.32108765840530396, + 0.05698775872588158, + 0.05295611172914505, + 0.1685187816619873, + 0.06266942620277405, + 0.055761437863111496, + 0.09488366544246674, + 0.03182622045278549, + 0.05545569956302643 + ], + "credit_map": [ + 0.00798575859516859, + 0.00842349324375391, + 0.01267915964126587, + 0.016216911375522614, + 0.0067838700488209724, + 0.009376304224133492, + 0.02642221562564373, + 0.011965583078563213, + 0.04790259152650833, + 0.09130353480577469, + 0.020675068721175194, + 0.023563504219055176, + 0.036878060549497604, + 0.015500706620514393, + 0.03144614398479462, + 0.05381804332137108, + 0.006168704945594072, + 0.006670869886875153, + 0.00633356673642993, + 0.00975870992988348, + 0.008054787293076515, + 0.009737289510667324, + 0.005280786659568548, + 0.004983046092092991, + 0.004523364827036858, + 0.008164027705788612, + 0.0038240794092416763, + 0.0053782216273248196, + 0.007137951906770468, + 0.007126584183424711, + 0.00757810240611434, + 0.009223781526088715, + 0.02148878388106823, + 0.012750877998769283, + 0.008430243469774723, + 0.01125386729836464, + 0.009173668920993805, + 0.018683258444070816, + 0.039675239473581314, + 0.047062840312719345, + 0.004243419971317053, + 0.0025157686322927475, + 0.0030995546840131283, + 0.005476077552884817, + 0.0031604995019733906, + 0.007431678008288145, + 0.01932927779853344, + 0.017413143068552017, + 0.0028324280865490437, + 0.003723304020240903, + 0.002707325154915452, + 0.004667544271796942, + 0.011838025413453579, + 0.014377845451235771, + 0.008882297202944756, + 0.006732664071023464, + 0.006138631608337164, + 0.011067375540733337, + 0.007609593216329813, + 0.012867647223174572, + 0.017926698550581932, + 0.012448970228433609, + 0.015005762688815594, + 0.011818981729447842, + 0.0037750909104943275, + 0.002668367698788643, + 0.00267052068375051, + 0.00523022236302495, + 0.00511753698810935, + 0.004576939158141613, + 0.003575709881260991, + 0.004211834631860256, + 0.0034232090692967176, + 0.004623359069228172, + 0.003094644518569112, + 0.005621880292892456, + 0.009896504692733288, + 0.008922378532588482, + 0.013463088311254978, + 0.006410636473447084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.002202943665906787 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069188714027405 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.016069188714027405, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069188714027405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11451460421085358, + 0.05288587510585785, + 0.03400668129324913, + 0.08676361292600632, + 0.28981146216392517, + 0.13203132152557373, + 0.09488581866025925, + 0.09777200222015381, + 0.03331947699189186, + 0.0640091598033905 + ], + "credit_map": [ + 0.008274313062429428, + 0.011866724118590355, + 0.012412653304636478, + 0.018708305433392525, + 0.01066625863313675, + 0.014561331830918789, + 0.025514259934425354, + 0.012510746717453003, + 0.0054605863988399506, + 0.009835179895162582, + 0.004676360171288252, + 0.010494638234376907, + 0.005656964611262083, + 0.005111485254019499, + 0.006448002532124519, + 0.005202657077461481, + 0.0030066058970987797, + 0.003883921541273594, + 0.0025081641506403685, + 0.006154732778668404, + 0.005553076509386301, + 0.004084593616425991, + 0.004811603110283613, + 0.004003981593996286, + 0.008117292076349258, + 0.02202964946627617, + 0.003856073599308729, + 0.007384201046079397, + 0.012350373901426792, + 0.005458641331642866, + 0.010001186281442642, + 0.017566192895174026, + 0.036269739270210266, + 0.0257205069065094, + 0.013850937597453594, + 0.02275989204645157, + 0.01725177653133869, + 0.027825504541397095, + 0.068210169672966, + 0.07792291790246964, + 0.008153911679983139, + 0.006756909657269716, + 0.007023122627288103, + 0.011798227205872536, + 0.007821346633136272, + 0.01871470920741558, + 0.038471851497888565, + 0.03329123556613922, + 0.005352284759283066, + 0.007542937528342009, + 0.005213680677115917, + 0.015481112524867058, + 0.011233994737267494, + 0.024003347381949425, + 0.013570123352110386, + 0.012488327920436859, + 0.0047984374687075615, + 0.00969529990106821, + 0.006872628349810839, + 0.011367501690983772, + 0.01975848898291588, + 0.01645348221063614, + 0.014724135398864746, + 0.014102021232247353, + 0.0030092240776866674, + 0.0032585570588707924, + 0.002652084222063422, + 0.004181272350251675, + 0.005039006005972624, + 0.00736323744058609, + 0.003387311240658164, + 0.0044287811033427715, + 0.0037100957706570625, + 0.005925224162638187, + 0.003239657962694764, + 0.009511541575193405, + 0.011441989801824093, + 0.009664179757237434, + 0.012281255796551704, + 0.008235214278101921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.016069188714027405 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a massive humpback whale as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the whale have blue-black skin covered in white scars and sparsely with barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are beams of golden light (Tyndall effect) visible penetrating the azure water and casting dappled light on the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale descend slowly into the depths while its long pectoral fins stretch like wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with an upward perspective and slowly tilt downward to follow the whale?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202935054898262 + ], + "opposite_logprob_tokens": [ + -2.8120293617248535 + ], + "logprob_reward": -0.06202935054898262, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202935054898262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13910888135433197, + 0.11708557605743408, + 0.03074050508439541, + 0.13337856531143188, + 0.13693249225616455, + 0.13258865475654602, + 0.08721011877059937, + 0.07185869663953781, + 0.054688211530447006, + 0.09640834480524063 + ], + "credit_map": [ + 0.009387223981320858, + 0.009654299356043339, + 0.018236080184578896, + 0.0343003049492836, + 0.009395423345267773, + 0.015452850610017776, + 0.027550382539629936, + 0.015132302418351173, + 0.01263140793889761, + 0.028623994439840317, + 0.014077203348279, + 0.01416308619081974, + 0.015033980831503868, + 0.012147627770900726, + 0.010451793670654297, + 0.009956478141248226, + 0.0028985324315726757, + 0.0028777732513844967, + 0.0031425131019204855, + 0.006070232484489679, + 0.0031315116211771965, + 0.0022575806360691786, + 0.006353466771543026, + 0.0040088919922709465, + 0.022462157532572746, + 0.048591289669275284, + 0.00837047677487135, + 0.009099951945245266, + 0.023809954524040222, + 0.006941819563508034, + 0.006413402035832405, + 0.007689522113651037, + 0.013415122404694557, + 0.014353563077747822, + 0.005587723571807146, + 0.027746621519327164, + 0.010608467273414135, + 0.008913207799196243, + 0.028462767601013184, + 0.027845017611980438, + 0.009815131314098835, + 0.007825919426977634, + 0.006751634646207094, + 0.011950270272791386, + 0.011677855625748634, + 0.023047635331749916, + 0.03388992324471474, + 0.027630286291241646, + 0.004927688743919134, + 0.007697199005633593, + 0.005896452814340591, + 0.009087644517421722, + 0.01658138819038868, + 0.019732380285859108, + 0.014326623640954494, + 0.008960745297372341, + 0.0037788720801472664, + 0.008541463874280453, + 0.005293332971632481, + 0.010110286995768547, + 0.014109629206359386, + 0.00987080205231905, + 0.010977174155414104, + 0.009177139960229397, + 0.0038046601694077253, + 0.004610100761055946, + 0.0031013682018965483, + 0.007911984808743, + 0.008687403053045273, + 0.011525915004312992, + 0.007443931419402361, + 0.00760284997522831, + 0.005970760248601437, + 0.006373319774866104, + 0.004839187487959862, + 0.01059167180210352, + 0.018644629046320915, + 0.016363848000764847, + 0.021686794236302376, + 0.011938133276998997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.06202935054898262 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_13/index_tensor([1415, 1415])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_13/index_tensor([1415, 1415])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d082cf76b085a190376c8e1cfbdf70e0a983937a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_13/index_tensor([1415, 1415])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This medium, eye-level shot, captured with a handheld camera, subtly panning to heighten the chaos, shows a man in his 30s at the wheel of a gleaming, aluminum, adult-sized tricycle. He stands on a wide suburban street, slick and dark beneath a gray, cloudy sky from a recent rain. The man, his brown hair plastered to his forehead, wears a vibrant yellow rain hoodie over jeans. His face perfectly conceals a comical panic, eyes wide and mouth open. He is in the midst of a frantic fishtail; both rear wheels lose traction on the wet asphalt and slide sharply to one side, setting off a dramatic arc of water. He desperately twists the handlebars in opposite directions, his body leaning hard to balance the slide, his legs flailing as he struggles to keep his feet on the pedals. The tricycle's polished aluminum frame catches the diffused light, shimmering even as it careens uncontrollably. In the background, neat suburban houses and manicured lawns, slightly out of focus, provide a calming backdrop to the unfolding chaos. The entire shot is filled with subtle film grain, adding a cinematic and nostalgic quality to a chaotic and fun moment of pure, effortless panic.", + "seed": 23246, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_13/index_1415_seed_23246_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1917499590665102, + "vlm_logprob": -0.1917499590665102, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.1917499590665102, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10231516510248184 + ], + "opposite_logprob_tokens": [ + -3.1023151874542236 + ], + "logprob_reward": -0.10231516510248184, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.10231516510248184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17760492861270905, + 0.04635622724890709, + 0.06402313709259033, + 0.045791830867528915, + 0.10917415469884872, + 0.08570677787065506, + 0.12824025750160217, + 0.2348494678735733, + 0.05679094418883324, + 0.051462288945913315 + ], + "credit_map": [ + 0.01370837353169918, + 0.020508544519543648, + 0.026719417423009872, + 0.04082776606082916, + 0.014448912814259529, + 0.02109864540398121, + 0.020981566980481148, + 0.019311703741550446, + 0.005886406637728214, + 0.0042238314636051655, + 0.006680834107100964, + 0.008642743341624737, + 0.004182426258921623, + 0.003988868091255426, + 0.006904282607138157, + 0.005846827756613493, + 0.00712669687345624, + 0.009457984939217567, + 0.011766756884753704, + 0.00644107349216938, + 0.004118191543966532, + 0.013684784062206745, + 0.004553044680505991, + 0.006874597165733576, + 0.006223964970558882, + 0.003924656193703413, + 0.005917178466916084, + 0.009912229143083096, + 0.005081055220216513, + 0.004158253315836191, + 0.005312254652380943, + 0.005262237042188644, + 0.008838562294840813, + 0.01207713782787323, + 0.01423919852823019, + 0.01522385235875845, + 0.019704163074493408, + 0.011829172261059284, + 0.01509233471006155, + 0.012169718742370605, + 0.008517790585756302, + 0.006732328329235315, + 0.013636029325425625, + 0.008363054133951664, + 0.011575407348573208, + 0.01109083741903305, + 0.013356820680201054, + 0.012434503063559532, + 0.035719357430934906, + 0.011210514232516289, + 0.016407044604420662, + 0.007901572622358799, + 0.00673525920137763, + 0.010839413851499557, + 0.007294563110917807, + 0.03213253244757652, + 0.02884230948984623, + 0.02394414320588112, + 0.04208458214998245, + 0.03029000014066696, + 0.017891258001327515, + 0.01715969666838646, + 0.030012473464012146, + 0.044624995440244675, + 0.008492456749081612, + 0.007609168067574501, + 0.008220282383263111, + 0.006315364968031645, + 0.0064524151384830475, + 0.006012324709445238, + 0.007453496102243662, + 0.006235431879758835, + 0.006403106264770031, + 0.005541842896491289, + 0.005320179276168346, + 0.0067588575184345245, + 0.008924026973545551, + 0.0038250244688242674, + 0.009835351258516312, + 0.004853895399719477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.10231516510248184 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328511238098145 + ], + "opposite_logprob_tokens": [ + -0.7578511238098145 + ], + "logprob_reward": -0.6328511238098145, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328511238098145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1797219216823578, + 0.05643453076481819, + 0.24850468337535858, + 0.03458773344755173, + 0.040636636316776276, + 0.03186103329062462, + 0.05922764167189598, + 0.266893595457077, + 0.05741020664572716, + 0.024722013622522354 + ], + "credit_map": [ + 0.014935103245079517, + 0.020184246823191643, + 0.02898910455405712, + 0.045943159610033035, + 0.012474039569497108, + 0.01409658882766962, + 0.019232017919421196, + 0.0238676555454731, + 0.005934986751526594, + 0.006138191092759371, + 0.008714184165000916, + 0.010025358758866787, + 0.004653316922485828, + 0.005619968753308058, + 0.007244469132274389, + 0.008104055188596249, + 0.02862565591931343, + 0.023007569834589958, + 0.032236311584711075, + 0.016646213829517365, + 0.021130327135324478, + 0.07353727519512177, + 0.023440487682819366, + 0.02988084964454174, + 0.003038524417206645, + 0.002629053546115756, + 0.004804633092135191, + 0.007794677745550871, + 0.003916385117918253, + 0.002535926178097725, + 0.005061522591859102, + 0.004807011224329472, + 0.005147880408912897, + 0.006387967616319656, + 0.006129178684204817, + 0.006820698734372854, + 0.0026988391764461994, + 0.0032533262856304646, + 0.004837299231439829, + 0.005361449904739857, + 0.0034726434387266636, + 0.002898975508287549, + 0.003696603002026677, + 0.004334565717726946, + 0.003145344089716673, + 0.004357299767434597, + 0.004937190096825361, + 0.00501841027289629, + 0.005842593498528004, + 0.005146614741533995, + 0.012986037880182266, + 0.006397231947630644, + 0.006315616425126791, + 0.0073259309865534306, + 0.0064780754037201405, + 0.008735541254281998, + 0.0322461873292923, + 0.02413434162735939, + 0.043773457407951355, + 0.02677190862596035, + 0.025995926931500435, + 0.033449724316596985, + 0.02887605130672455, + 0.051645997911691666, + 0.008569471538066864, + 0.007724333088845015, + 0.008819027803838253, + 0.007900971919298172, + 0.005004125647246838, + 0.0061740451492369175, + 0.007324233651161194, + 0.005893997848033905, + 0.003353965701535344, + 0.0027946566697210073, + 0.0030519042629748583, + 0.0035987920127809048, + 0.0032893044408410788, + 0.0020224922336637974, + 0.003548590699210763, + 0.003062306670472026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.6328511238098145 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059571098536252975 + ], + "opposite_logprob_tokens": [ + -5.130957126617432 + ], + "logprob_reward": -0.0059571098536252975, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059571098536252975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19871190190315247, + 0.056138668209314346, + 0.1630055457353592, + 0.0650036633014679, + 0.060712818056344986, + 0.03245161846280098, + 0.06443074345588684, + 0.19080321490764618, + 0.1029062494635582, + 0.0658356249332428 + ], + "credit_map": [ + 0.013088097795844078, + 0.017071109265089035, + 0.030464820563793182, + 0.05326255410909653, + 0.019805217161774635, + 0.015791425481438637, + 0.028445027768611908, + 0.020783647894859314, + 0.005306628067046404, + 0.00612413976341486, + 0.008089232258498669, + 0.010431598871946335, + 0.005594499874860048, + 0.005079719237983227, + 0.0077756126411259174, + 0.007737235631793737, + 0.011324710212647915, + 0.013962650671601295, + 0.035994723439216614, + 0.015244156122207642, + 0.014236724935472012, + 0.039568688720464706, + 0.014503608457744122, + 0.018170278519392014, + 0.005924161057919264, + 0.005583782214671373, + 0.00880932342261076, + 0.01113478746265173, + 0.007249901536852121, + 0.005093654617667198, + 0.010328637436032295, + 0.010879413224756718, + 0.004518029745668173, + 0.005454150028526783, + 0.0075541152618825436, + 0.011590513400733471, + 0.008366873487830162, + 0.006111052352935076, + 0.0073528531938791275, + 0.009765228256583214, + 0.004179102834314108, + 0.0031371237710118294, + 0.0033294453751295805, + 0.0042103673331439495, + 0.004807346034795046, + 0.004918144084513187, + 0.004608409479260445, + 0.003261677920818329, + 0.007503495551645756, + 0.0067183575592935085, + 0.0094398632645607, + 0.006544400472193956, + 0.005458883009850979, + 0.006641921121627092, + 0.006871586199849844, + 0.015252232551574707, + 0.020620133727788925, + 0.023065682500600815, + 0.02988472953438759, + 0.022010430693626404, + 0.018208038061857224, + 0.026679590344429016, + 0.019943641498684883, + 0.030390966683626175, + 0.026219751685857773, + 0.005831234157085419, + 0.035566240549087524, + 0.005848115775734186, + 0.006255874410271645, + 0.004939731676131487, + 0.008428678847849369, + 0.009816624224185944, + 0.00519572664052248, + 0.006194877438247204, + 0.005850994493812323, + 0.009671377018094063, + 0.013207100331783295, + 0.006012201774865389, + 0.013863775879144669, + 0.005839567631483078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0059571098536252975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156565308570862 + ], + "opposite_logprob_tokens": [ + -1.7015656232833862 + ], + "logprob_reward": -0.20156565308570862, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156565308570862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19050288200378418, + 0.04510575160384178, + 0.1081874892115593, + 0.017652902752161026, + 0.09525778144598007, + 0.06561513990163803, + 0.07527238130569458, + 0.23621954023838043, + 0.10650202631950378, + 0.05968410521745682 + ], + "credit_map": [ + 0.014931024052202702, + 0.018681008368730545, + 0.022151164710521698, + 0.032937634736299515, + 0.031043395400047302, + 0.022833766415715218, + 0.025519173592329025, + 0.022405698895454407, + 0.005376341752707958, + 0.005165713373571634, + 0.006003492511808872, + 0.007345567923039198, + 0.005396260414272547, + 0.004529563244432211, + 0.006015591323375702, + 0.005273221060633659, + 0.01086676586419344, + 0.01869260147213936, + 0.02078159712255001, + 0.011632995679974556, + 0.008860628120601177, + 0.016376124694943428, + 0.00799933448433876, + 0.01297744270414114, + 0.002368729095906019, + 0.001703559304587543, + 0.0025051175616681576, + 0.0034479021560400724, + 0.0011714092688634992, + 0.0012510360684245825, + 0.0028364018071442842, + 0.0023687465582042933, + 0.009436509571969509, + 0.009432604536414146, + 0.01354192290455103, + 0.014056195504963398, + 0.015295261517167091, + 0.010141613893210888, + 0.014287994243204594, + 0.009065675549209118, + 0.00534078711643815, + 0.005478045903146267, + 0.007898849435150623, + 0.0096385283395648, + 0.012252682819962502, + 0.00576442526653409, + 0.01252077892422676, + 0.006721045821905136, + 0.007642870303243399, + 0.007254389580339193, + 0.010957087390124798, + 0.005694929976016283, + 0.005132820922881365, + 0.011708000674843788, + 0.006289263255894184, + 0.02059302106499672, + 0.028438635170459747, + 0.020861929282546043, + 0.042462997138500214, + 0.03260631486773491, + 0.015934156253933907, + 0.02276676520705223, + 0.03315793722867966, + 0.03999081254005432, + 0.020139925181865692, + 0.00983361341059208, + 0.030625442042946815, + 0.010303884744644165, + 0.00907011330127716, + 0.0053102923557162285, + 0.011386189609766006, + 0.009832564741373062, + 0.009411688894033432, + 0.005973913706839085, + 0.01084029208868742, + 0.006749166175723076, + 0.007122171577066183, + 0.003372766776010394, + 0.011014382354915142, + 0.0051997266709804535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.20156565308570862 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606074348092079 + ], + "opposite_logprob_tokens": [ + -4.141060829162598 + ], + "logprob_reward": -0.01606074348092079, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606074348092079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.215078666806221, + 0.10636742413043976, + 0.05750584974884987, + 0.056529540568590164, + 0.14218291640281677, + 0.04014948010444641, + 0.07259300351142883, + 0.15512743592262268, + 0.09218952059745789, + 0.06227623671293259 + ], + "credit_map": [ + 0.015546065755188465, + 0.02635747194290161, + 0.02551853656768799, + 0.04774167016148567, + 0.02939918264746666, + 0.020853009074926376, + 0.029133832082152367, + 0.020528895780444145, + 0.009656252339482307, + 0.015942256897687912, + 0.0140409916639328, + 0.0152610894292593, + 0.008925061672925949, + 0.011117804795503616, + 0.014267237856984138, + 0.017156727612018585, + 0.005208533722907305, + 0.008863775059580803, + 0.013142063282430172, + 0.006281730718910694, + 0.00569566385820508, + 0.005652583669871092, + 0.005315394140779972, + 0.0073461071588099, + 0.0072904787957668304, + 0.00529329851269722, + 0.007899099960923195, + 0.010989208705723286, + 0.004940926097333431, + 0.0036059031262993813, + 0.006248939782381058, + 0.010261684656143188, + 0.015325083397328854, + 0.029316503554582596, + 0.022281039506196976, + 0.0227645356208086, + 0.01143160741776228, + 0.01240981463342905, + 0.01554566714912653, + 0.01310865767300129, + 0.003059610491618514, + 0.0035325370263308287, + 0.004205981735140085, + 0.005255746189504862, + 0.005828749854117632, + 0.00513949291780591, + 0.009354074485599995, + 0.003773288568481803, + 0.006603460293263197, + 0.007697361055761576, + 0.01315345149487257, + 0.010819939896464348, + 0.0068105715326964855, + 0.007879882119596004, + 0.009266948327422142, + 0.010361388325691223, + 0.02221447415649891, + 0.018393712118268013, + 0.028951292857527733, + 0.01649000495672226, + 0.011740599758923054, + 0.014863604679703712, + 0.019260141998529434, + 0.02321360632777214, + 0.021519536152482033, + 0.00828954204916954, + 0.029715487733483315, + 0.006536270026117563, + 0.0055766343139112, + 0.005045586731284857, + 0.00642680237069726, + 0.009079654701054096, + 0.012897792272269726, + 0.009090381674468517, + 0.009504704736173153, + 0.007757248822599649, + 0.004766765050590038, + 0.004996183328330517, + 0.007837331853806973, + 0.00542583130300045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.01606074348092079 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This medium, eye-level shot, captured with a handheld camera, subtly panning to heighten the chaos, shows a man in his 30s at the wheel of a gleaming, aluminum, adult-sized tricycle. He stands on a wide suburban street, slick and dark beneath a gray, cloudy sky from a recent rain. The man, his brown hair plastered to his forehead, wears a vibrant yellow rain hoodie over jeans. His face perfectly conceals a comical panic, eyes wide and mouth open. He is in the midst of a frantic fishtail; both rear wheels lose traction on the wet asphalt and slide sharply to one side, setting off a dramatic arc of water. He desperately twists the handlebars in opposite directions, his body leaning hard to balance the slide, his legs flailing as he struggles to keep his feet on the pedals. The tricycle's polished aluminum frame catches the diffused light, shimmering even as it careens uncontrollably. In the background, neat suburban houses and manicured lawns, slightly out of focus, provide a calming backdrop to the unfolding chaos. The entire shot is filled with subtle film grain, adding a cinematic and nostalgic quality to a chaotic and fun moment of pure, effortless panic.", + "seed": 23247, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_13/index_1415_seed_23247_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2608204678632319, + "vlm_logprob": -0.2608204678632319, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.2608204678632319, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15608058869838715 + ], + "opposite_logprob_tokens": [ + -2.531080484390259 + ], + "logprob_reward": -0.15608058869838715, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.15608058869838715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18325909972190857, + 0.07953015714883804, + 0.04769514128565788, + 0.03549041971564293, + 0.04885964095592499, + 0.052129920572042465, + 0.07816962152719498, + 0.1403641402721405, + 0.20142295956611633, + 0.13307888805866241 + ], + "credit_map": [ + 0.030426258221268654, + 0.019191846251487732, + 0.019117863848805428, + 0.025281405076384544, + 0.017185533419251442, + 0.023450274020433426, + 0.01930626854300499, + 0.029299665242433548, + 0.010509083047509193, + 0.0079282121732831, + 0.007063631433993578, + 0.010456482879817486, + 0.01796838641166687, + 0.0063943443819880486, + 0.012560405768454075, + 0.006649618502706289, + 0.006130178924649954, + 0.003627209924161434, + 0.0063799722120165825, + 0.008626227267086506, + 0.004385882057249546, + 0.005055442918092012, + 0.006266388576477766, + 0.007223839405924082, + 0.00439605163410306, + 0.004234963562339544, + 0.004337925463914871, + 0.004049153998494148, + 0.004528249613940716, + 0.00412660650908947, + 0.0049470956437289715, + 0.004870375618338585, + 0.004081795457750559, + 0.004776421934366226, + 0.005569243337959051, + 0.008876488544046879, + 0.006833298597484827, + 0.004734587389975786, + 0.006466784980148077, + 0.007521024439483881, + 0.00722888670861721, + 0.0045752800069749355, + 0.007223045919090509, + 0.01073271594941616, + 0.00464321905747056, + 0.007545934524387121, + 0.004407359287142754, + 0.005773482378572226, + 0.005869573447853327, + 0.006344721186906099, + 0.009978176094591618, + 0.008322491310536861, + 0.013244585134088993, + 0.010866713710129261, + 0.015485058538615704, + 0.008058306761085987, + 0.009935286827385426, + 0.013104152865707874, + 0.020562520250678062, + 0.05730001628398895, + 0.006597842089831829, + 0.008231556043028831, + 0.013432304374873638, + 0.01120045967400074, + 0.014446014538407326, + 0.01575075089931488, + 0.02773568220436573, + 0.05423920601606369, + 0.03229592368006706, + 0.015350211411714554, + 0.020047802478075027, + 0.02155737392604351, + 0.015550571493804455, + 0.010146625339984894, + 0.025927938520908356, + 0.019559558480978012, + 0.014195464551448822, + 0.013215389102697372, + 0.01700020767748356, + 0.017483137547969818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.15608058869838715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8262374401092529 + ], + "opposite_logprob_tokens": [ + -0.5762374401092529 + ], + "logprob_reward": -0.8262374401092529, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8262374401092529, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22118982672691345, + 0.07717008888721466, + 0.07745467871427536, + 0.10675507038831711, + 0.08818421512842178, + 0.0669238492846489, + 0.04147181287407875, + 0.09370293468236923, + 0.13542325794696808, + 0.09172429144382477 + ], + "credit_map": [ + 0.021093958988785744, + 0.018123183399438858, + 0.029176056385040283, + 0.05462735891342163, + 0.013618933036923409, + 0.023844344541430473, + 0.018571391701698303, + 0.0421346053481102, + 0.009320931509137154, + 0.007792071904987097, + 0.008497213013470173, + 0.01021196972578764, + 0.010758600197732449, + 0.010727067478001118, + 0.009434047155082226, + 0.010428191162645817, + 0.01285778172314167, + 0.006844493560492992, + 0.007982492446899414, + 0.014055941253900528, + 0.005727686453610659, + 0.013603177852928638, + 0.006630481220781803, + 0.009752624668180943, + 0.013452950865030289, + 0.010015789419412613, + 0.01266021654009819, + 0.02180648036301136, + 0.009086184203624725, + 0.011679261922836304, + 0.012385331094264984, + 0.015668855980038643, + 0.009992783889174461, + 0.011000572703778744, + 0.009194144047796726, + 0.012928958050906658, + 0.008085938170552254, + 0.01116736140102148, + 0.01049221120774746, + 0.015322245657444, + 0.006515918765217066, + 0.006798364222049713, + 0.008799946866929531, + 0.010526075959205627, + 0.006383934058248997, + 0.006021813955157995, + 0.012198606505990028, + 0.009679196402430534, + 0.004150521941483021, + 0.0032149003818631172, + 0.005761012434959412, + 0.006034838035702705, + 0.006200889591127634, + 0.005109876394271851, + 0.006248325575143099, + 0.0047514489851891994, + 0.009136120788753033, + 0.007982082664966583, + 0.012951175682246685, + 0.030339377000927925, + 0.005538873374462128, + 0.007065418176352978, + 0.011145848780870438, + 0.009544036351144314, + 0.01168437022715807, + 0.011582153849303722, + 0.019058801233768463, + 0.04219404235482216, + 0.015697916969656944, + 0.010582287795841694, + 0.0112374909222126, + 0.013386180624365807, + 0.01147494651377201, + 0.007405638229101896, + 0.015337650664150715, + 0.016277611255645752, + 0.007798109669238329, + 0.010319543071091175, + 0.012727445922791958, + 0.010383345186710358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.8262374401092529 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067553711123764515 + ], + "opposite_logprob_tokens": [ + -5.006755352020264 + ], + "logprob_reward": -0.0067553711123764515, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067553711123764515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2081698626279831, + 0.04619532823562622, + 0.08095673471689224, + 0.07560163736343384, + 0.0635978952050209, + 0.24290840327739716, + 0.08179192245006561, + 0.038034435361623764, + 0.06288107484579086, + 0.09986268728971481 + ], + "credit_map": [ + 0.02521674893796444, + 0.024079540744423866, + 0.023366142064332962, + 0.03440127149224281, + 0.019132796674966812, + 0.02828667312860489, + 0.020142314955592155, + 0.0335443876683712, + 0.0068884254433214664, + 0.004315667785704136, + 0.00525976438075304, + 0.005722937639802694, + 0.007792593911290169, + 0.004410068970173597, + 0.0046599810011684895, + 0.007145888637751341, + 0.013882044702768326, + 0.007240463979542255, + 0.00820883922278881, + 0.010953903198242188, + 0.006619814783334732, + 0.015304474160075188, + 0.007356023881584406, + 0.011391176842153072, + 0.011052502319216728, + 0.006939788348972797, + 0.009576614946126938, + 0.01510666310787201, + 0.0056852130219340324, + 0.008163864724338055, + 0.00758946780115366, + 0.011487527750432491, + 0.007669239770621061, + 0.005912392400205135, + 0.007791205774992704, + 0.009497115388512611, + 0.005872995126992464, + 0.007120834197849035, + 0.008355184458196163, + 0.011378930881619453, + 0.026926858350634575, + 0.018958397209644318, + 0.040569450706243515, + 0.06768772751092911, + 0.012963490560650826, + 0.01914079487323761, + 0.030490051954984665, + 0.026171663776040077, + 0.008399740792810917, + 0.007841328158974648, + 0.009158967062830925, + 0.012212269008159637, + 0.011351531371474266, + 0.010621875524520874, + 0.01270949374884367, + 0.009496727958321571, + 0.00681006396189332, + 0.0028935286682099104, + 0.005313887260854244, + 0.004779960494488478, + 0.004164513200521469, + 0.0032692099921405315, + 0.006463934201747179, + 0.0043393406085669994, + 0.005590467248111963, + 0.005194804631173611, + 0.010587469674646854, + 0.01820112019777298, + 0.006713423877954483, + 0.00559420557692647, + 0.005323989316821098, + 0.005675598047673702, + 0.010303925722837448, + 0.008919360116124153, + 0.01673438772559166, + 0.013754311017692089, + 0.012142753228545189, + 0.013460307382047176, + 0.013083208352327347, + 0.011464443989098072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.0067553711123764515 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813347578048706 + ], + "opposite_logprob_tokens": [ + -1.4063347578048706 + ], + "logprob_reward": -0.2813347578048706, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813347578048706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14387322962284088, + 0.10025891661643982, + 0.05195683613419533, + 0.1406990885734558, + 0.09070337563753128, + 0.1503172069787979, + 0.060265250504016876, + 0.12087580561637878, + 0.06465887278318405, + 0.07639139890670776 + ], + "credit_map": [ + 0.017305321991443634, + 0.017077777534723282, + 0.0141889164224267, + 0.025229360908269882, + 0.009934727102518082, + 0.0253602322191, + 0.012923972681164742, + 0.021852921694517136, + 0.007749251555651426, + 0.011500569991767406, + 0.0112760616466403, + 0.01980089209973812, + 0.01822228729724884, + 0.008179780095815659, + 0.013007935136556625, + 0.010522136464715004, + 0.005757312756031752, + 0.004216263070702553, + 0.00521487183868885, + 0.009360491298139095, + 0.005359479691833258, + 0.0059066107496619225, + 0.008483152836561203, + 0.007658654358237982, + 0.011989369988441467, + 0.014649026095867157, + 0.015739096328616142, + 0.017168430611491203, + 0.026753395795822144, + 0.017419064417481422, + 0.02198784612119198, + 0.014992864802479744, + 0.008555922657251358, + 0.010027625598013401, + 0.01014024019241333, + 0.016495447605848312, + 0.013011232018470764, + 0.008843254297971725, + 0.01275869831442833, + 0.010870955884456635, + 0.017099179327487946, + 0.011991529725492, + 0.025403503328561783, + 0.03605138137936592, + 0.010187861509621143, + 0.01600499451160431, + 0.015678556635975838, + 0.017900193110108376, + 0.006766406819224358, + 0.005145212635397911, + 0.007576374802738428, + 0.010772809386253357, + 0.004537793807685375, + 0.00643174396827817, + 0.012352949939668179, + 0.0066819582134485245, + 0.01066557690501213, + 0.011104373261332512, + 0.017079059034585953, + 0.03715552017092705, + 0.008423768915235996, + 0.013381565921008587, + 0.011241447180509567, + 0.011824499815702438, + 0.005216248799115419, + 0.0052030328661203384, + 0.007869686931371689, + 0.020461555570364, + 0.005900613963603973, + 0.00516669824719429, + 0.007910299114882946, + 0.00693074194714427, + 0.00954696349799633, + 0.005867488216608763, + 0.014713341370224953, + 0.013245471753180027, + 0.007952866144478321, + 0.005253805313259363, + 0.011006093584001064, + 0.008805370889604092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.2813347578048706 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033694181591272354 + ], + "opposite_logprob_tokens": [ + -3.408694267272949 + ], + "logprob_reward": -0.033694181591272354, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033694181591272354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15597651898860931, + 0.09317497909069061, + 0.04148717224597931, + 0.05889631435275078, + 0.07492858171463013, + 0.12985284626483917, + 0.11226695775985718, + 0.1981123983860016, + 0.07545933127403259, + 0.05984489619731903 + ], + "credit_map": [ + 0.022982215508818626, + 0.013796384446322918, + 0.01803368702530861, + 0.03352786600589752, + 0.010011387057602406, + 0.016639141365885735, + 0.0132378488779068, + 0.027747996151447296, + 0.009837810881435871, + 0.012088589370250702, + 0.010039771907031536, + 0.012599313631653786, + 0.014238635078072548, + 0.010483987629413605, + 0.011528857052326202, + 0.012358007952570915, + 0.0070798760280013084, + 0.004952796269208193, + 0.004840151872485876, + 0.004895023535937071, + 0.002999774180352688, + 0.0074484082870185375, + 0.0034288770984858274, + 0.005842264275997877, + 0.006572004407644272, + 0.004901543725281954, + 0.006684121675789356, + 0.009100599214434624, + 0.008439579047262669, + 0.00721529358997941, + 0.008461281657218933, + 0.007521888241171837, + 0.009065438061952591, + 0.0074671972543001175, + 0.008586037904024124, + 0.01139109581708908, + 0.0078055113554000854, + 0.009490437805652618, + 0.0067627206444740295, + 0.014360143803060055, + 0.017236780375242233, + 0.011738683097064495, + 0.021389935165643692, + 0.028033960610628128, + 0.010190154425799847, + 0.009420015849173069, + 0.014370742253959179, + 0.01747256889939308, + 0.019147595390677452, + 0.011528993025422096, + 0.014979665167629719, + 0.021740293130278587, + 0.006434734910726547, + 0.01308705285191536, + 0.012150718830525875, + 0.013197901658713818, + 0.011505390517413616, + 0.017403677105903625, + 0.031209640204906464, + 0.08596032112836838, + 0.008284127339720726, + 0.009859664365649223, + 0.017539260908961296, + 0.01635032519698143, + 0.00637034373357892, + 0.006541373208165169, + 0.008269832469522953, + 0.02053465135395527, + 0.009615177288651466, + 0.005383041687309742, + 0.008233031257987022, + 0.010511879809200764, + 0.0077957939356565475, + 0.005693706218153238, + 0.01187547855079174, + 0.01063117291778326, + 0.005900981370359659, + 0.004608387593179941, + 0.007052137050777674, + 0.00628723856061697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.033694181591272354 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This medium, eye-level shot, captured with a handheld camera, subtly panning to heighten the chaos, shows a man in his 30s at the wheel of a gleaming, aluminum, adult-sized tricycle. He stands on a wide suburban street, slick and dark beneath a gray, cloudy sky from a recent rain. The man, his brown hair plastered to his forehead, wears a vibrant yellow rain hoodie over jeans. His face perfectly conceals a comical panic, eyes wide and mouth open. He is in the midst of a frantic fishtail; both rear wheels lose traction on the wet asphalt and slide sharply to one side, setting off a dramatic arc of water. He desperately twists the handlebars in opposite directions, his body leaning hard to balance the slide, his legs flailing as he struggles to keep his feet on the pedals. The tricycle's polished aluminum frame catches the diffused light, shimmering even as it careens uncontrollably. In the background, neat suburban houses and manicured lawns, slightly out of focus, provide a calming backdrop to the unfolding chaos. The entire shot is filled with subtle film grain, adding a cinematic and nostalgic quality to a chaotic and fun moment of pure, effortless panic.", + "seed": 23248, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_13/index_1415_seed_23248_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2607596785761416, + "vlm_logprob": -0.2607596785761416, + "vlm_yes_no_margin": 2.4250000059604644, + "vlm_token_credit": -0.2607596785761416, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17515774071216583 + ], + "opposite_logprob_tokens": [ + -2.4251577854156494 + ], + "logprob_reward": -0.17515774071216583, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.17515774071216583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5769367814064026, + 0.05594230815768242, + 0.05624662712216377, + 0.022915638983249664, + 0.0301970262080431, + 0.03538346663117409, + 0.04362509027123451, + 0.04050564393401146, + 0.0833924263715744, + 0.05485493317246437 + ], + "credit_map": [ + 0.020863430574536324, + 0.021749576553702354, + 0.22782927751541138, + 0.23938684165477753, + 0.012971807271242142, + 0.010807705111801624, + 0.030750703066587448, + 0.012577516958117485, + 0.005200203508138657, + 0.004469277802854776, + 0.012079456821084023, + 0.007500160485506058, + 0.007302928250283003, + 0.0043464587070047855, + 0.008177673444151878, + 0.006866150069981813, + 0.006004642229527235, + 0.004365677945315838, + 0.011593425646424294, + 0.006874975748360157, + 0.005459246691316366, + 0.005687141325324774, + 0.009438903070986271, + 0.006822613533586264, + 0.0019242363050580025, + 0.001613563857972622, + 0.005656849127262831, + 0.003350241342559457, + 0.0027880521956831217, + 0.0017011136515066028, + 0.002754650078713894, + 0.0031269309110939503, + 0.0024564475752413273, + 0.0013847916852682829, + 0.009324404411017895, + 0.007203965913504362, + 0.0021812545601278543, + 0.002319562016054988, + 0.002718223026022315, + 0.002608375158160925, + 0.004539944231510162, + 0.0038365996442735195, + 0.007267609238624573, + 0.0077352458611130714, + 0.0022583722602576017, + 0.002661657053977251, + 0.003738590283319354, + 0.003345449222251773, + 0.004416387528181076, + 0.00440747058019042, + 0.0069414786994457245, + 0.005787438713014126, + 0.0038685393519699574, + 0.004666258115321398, + 0.007474760990589857, + 0.006062755361199379, + 0.004784141201525927, + 0.0034413218963891268, + 0.0053877015598118305, + 0.00529257208108902, + 0.005910936277359724, + 0.0050652772188186646, + 0.005834376905113459, + 0.004789312370121479, + 0.007921874523162842, + 0.008199148811399937, + 0.011143566109240055, + 0.008780639618635178, + 0.010783941484987736, + 0.009899749420583248, + 0.010991684161126614, + 0.01567182131111622, + 0.0037148436531424522, + 0.005259068217128515, + 0.0129878930747509, + 0.008539516478776932, + 0.004031622316688299, + 0.005992484744638205, + 0.006115110591053963, + 0.008214390836656094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.17515774071216583 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743417501449585 + ], + "opposite_logprob_tokens": [ + -0.4743417799472809 + ], + "logprob_reward": -0.9743417501449585, + "yes_no_margin_reward": -0.4999999701976776, + "token_credit_reward": -0.9743417501449585, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.8637462258338928, + 0.02358458936214447, + 0.03232067823410034, + 0.003322150092571974, + 0.009197264909744263, + 0.016894927248358727, + 0.009070973843336105, + 0.009106128476560116, + 0.009725922718644142, + 0.023031078279018402 + ], + "credit_map": [ + 0.026992468163371086, + 0.026003187522292137, + 0.35382989048957825, + 0.37213510274887085, + 0.0094422847032547, + 0.015334352850914001, + 0.043460529297590256, + 0.016548505052924156, + 0.0043121203780174255, + 0.0032878555357456207, + 0.006053767632693052, + 0.0034598989877849817, + 0.00158947822637856, + 0.0012454156531021, + 0.0014540591510012746, + 0.0021819963585585356, + 0.003235782962292433, + 0.003350822487846017, + 0.007253739517182112, + 0.003156458493322134, + 0.0016682660207152367, + 0.004504214506596327, + 0.0029807358514517546, + 0.006170661188662052, + 0.00029056950006633997, + 0.0003043841861654073, + 0.0007803967455402017, + 0.0005154546233825386, + 0.0003159971965942532, + 0.0002273295249324292, + 0.0003552238631527871, + 0.0005327949183993042, + 0.0005647602956742048, + 0.0005798156489618123, + 0.0026087502483278513, + 0.0021834177896380424, + 0.0006861452129669487, + 0.0007490181014873087, + 0.0009706027340143919, + 0.0008547552861273289, + 0.0017613840755075216, + 0.001379206427372992, + 0.005769893527030945, + 0.005417433567345142, + 0.00044369601528160274, + 0.000563935493119061, + 0.0008470752509310842, + 0.0007123042596504092, + 0.0009084321209229529, + 0.0005608880892395973, + 0.0010754953837022185, + 0.0013893133727833629, + 0.0007231931085698307, + 0.001340953167527914, + 0.001442350447177887, + 0.0016303493175655603, + 0.0012014381354674697, + 0.0008361382060684264, + 0.0013851263793185353, + 0.001479651196859777, + 0.0010500609641894698, + 0.0008502643904648721, + 0.0013217153027653694, + 0.0009817344835028052, + 0.0011602900922298431, + 0.001191360061056912, + 0.001589908730238676, + 0.0011106115998700261, + 0.0007332649547606707, + 0.0009264075197279453, + 0.0007739400025457144, + 0.0022401411551982164, + 0.0014547163154929876, + 0.0016472873976454139, + 0.0043909079395234585, + 0.005092089995741844, + 0.0017402446828782558, + 0.002007348695769906, + 0.0031613828614354134, + 0.0035371033009141684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.9743417501449585 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007650007028132677 + ], + "opposite_logprob_tokens": [ + -4.882649898529053 + ], + "logprob_reward": -0.007650007028132677, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007650007028132677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7599621415138245, + 0.02862902171909809, + 0.050531912595033646, + 0.01158838625997305, + 0.01679186336696148, + 0.031128691509366035, + 0.017876671627163887, + 0.013750297017395496, + 0.032706305384635925, + 0.03703469783067703 + ], + "credit_map": [ + 0.023905616253614426, + 0.0231031347066164, + 0.3082953095436096, + 0.3256284296512604, + 0.010892676189541817, + 0.014687059447169304, + 0.03871900960803032, + 0.01473089400678873, + 0.003708010073751211, + 0.004145101644098759, + 0.006178395356982946, + 0.003937731962651014, + 0.003154351143166423, + 0.002301553264260292, + 0.0028186943382024765, + 0.00238518207333982, + 0.006995056755840778, + 0.0053990683518350124, + 0.009635831229388714, + 0.005554137285798788, + 0.004393893759697676, + 0.004461648873984814, + 0.006073838099837303, + 0.008018437772989273, + 0.0007992987520992756, + 0.0008738832548260689, + 0.002740351017564535, + 0.00201556202955544, + 0.0011670415988191962, + 0.0008059314568527043, + 0.001394675811752677, + 0.0017916428623721004, + 0.0019839522428810596, + 0.001240817946381867, + 0.003509795293211937, + 0.002709613647311926, + 0.0013984892284497619, + 0.0016462475759908557, + 0.0023163349833339453, + 0.0019866114016622305, + 0.0041519394144415855, + 0.002947536762803793, + 0.009517526254057884, + 0.008956780657172203, + 0.0010257851099595428, + 0.0010512396693229675, + 0.0018123971531167626, + 0.0016654874198138714, + 0.001828921725973487, + 0.0022944731172174215, + 0.002367720240727067, + 0.0025523838121443987, + 0.0016153446631506085, + 0.00181129842530936, + 0.0026785230729728937, + 0.002728006336838007, + 0.002032836200669408, + 0.0011830395087599754, + 0.0025208715815097094, + 0.0022737092804163694, + 0.0012086918577551842, + 0.0015042367158457637, + 0.001720960601232946, + 0.0013059518532827497, + 0.0028970553539693356, + 0.003981285262852907, + 0.0038896403275430202, + 0.00433686189353466, + 0.003396068001165986, + 0.0039048881735652685, + 0.0048082685098052025, + 0.005492236465215683, + 0.0026534644421190023, + 0.00280504347756505, + 0.006656032986938953, + 0.006180461496114731, + 0.002583286724984646, + 0.004579546395689249, + 0.006065115798264742, + 0.00551174720749259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.007650007028132677 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295321583747864 + ], + "opposite_logprob_tokens": [ + -2.2379531860351562 + ], + "logprob_reward": -0.11295321583747864, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295321583747864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1798817664384842, + 0.08062943071126938, + 0.06222565099596977, + 0.02454461343586445, + 0.05369091033935547, + 0.16005107760429382, + 0.058877069503068924, + 0.05118153244256973, + 0.047765009105205536, + 0.28115296363830566 + ], + "credit_map": [ + 0.010916251689195633, + 0.014413139782845974, + 0.047916971147060394, + 0.0496375672519207, + 0.01702413335442543, + 0.017333023250102997, + 0.011581477709114552, + 0.011059219017624855, + 0.010206105187535286, + 0.010173043236136436, + 0.021142782643437386, + 0.012363001704216003, + 0.0052316682413220406, + 0.004970787093043327, + 0.008995948359370232, + 0.007546095177531242, + 0.005839433521032333, + 0.006388347130268812, + 0.016674015671014786, + 0.005707480013370514, + 0.004457976669073105, + 0.006018462125211954, + 0.008424594067037106, + 0.008715348318219185, + 0.0031946063973009586, + 0.002327344147488475, + 0.006147975102066994, + 0.004391903057694435, + 0.0013621298130601645, + 0.001912205945700407, + 0.001815106370486319, + 0.003393342485651374, + 0.0029039992950856686, + 0.0027333328034728765, + 0.017346372827887535, + 0.016791388392448425, + 0.002781445160508156, + 0.003509608330205083, + 0.0034722310956567526, + 0.004152530338615179, + 0.019564032554626465, + 0.014704308472573757, + 0.053829289972782135, + 0.04484470933675766, + 0.004848250187933445, + 0.006549932528287172, + 0.00853808131068945, + 0.007172472774982452, + 0.005754546262323856, + 0.004199638031423092, + 0.008240635506808758, + 0.010134231299161911, + 0.00729962345212698, + 0.011169946752488613, + 0.007001252379268408, + 0.005077195819467306, + 0.007009752094745636, + 0.005333186127245426, + 0.007786617614328861, + 0.009339850395917892, + 0.004406378138810396, + 0.0062788077630102634, + 0.006486314348876476, + 0.004540624562650919, + 0.005405513569712639, + 0.005133498460054398, + 0.007144330535084009, + 0.007256196811795235, + 0.004450464155524969, + 0.004921363666653633, + 0.004619211424142122, + 0.008834430947899818, + 0.015751678496599197, + 0.02105562575161457, + 0.054675064980983734, + 0.051463473588228226, + 0.02200000174343586, + 0.02559318207204342, + 0.049383990466594696, + 0.04122992977499962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.11295321583747864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695679157972336 + ], + "opposite_logprob_tokens": [ + -3.408695697784424 + ], + "logprob_reward": -0.033695679157972336, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033695679157972336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7091934084892273, + 0.07128854095935822, + 0.019773269072175026, + 0.01367229875177145, + 0.03794005513191223, + 0.02701726369559765, + 0.00972575694322586, + 0.008544627577066422, + 0.007905934937298298, + 0.09493884444236755 + ], + "credit_map": [ + 0.022624049335718155, + 0.021567190065979958, + 0.2893582284450531, + 0.3060755431652069, + 0.008246525190770626, + 0.012052512727677822, + 0.035535067319869995, + 0.013734295964241028, + 0.009769568219780922, + 0.007234104443341494, + 0.025486556813120842, + 0.012294997461140156, + 0.002577819162979722, + 0.0036195656284689903, + 0.004788070917129517, + 0.005517861340194941, + 0.0016797296702861786, + 0.0016909496625885367, + 0.004688989836722612, + 0.0017318063182756305, + 0.0010343281319364905, + 0.002573638455942273, + 0.002031768672168255, + 0.004342059604823589, + 0.00138986820820719, + 0.0013365373015403748, + 0.002401486039161682, + 0.00309751252643764, + 0.0013018033932894468, + 0.0010294949170202017, + 0.0012043400201946497, + 0.001911256113089621, + 0.0042487033642828465, + 0.0024477620609104633, + 0.012617867439985275, + 0.009734150022268295, + 0.0013774875551462173, + 0.001755708595737815, + 0.00272752670571208, + 0.0030308528803288937, + 0.004124470055103302, + 0.0032523025292903185, + 0.008718322031199932, + 0.004257816821336746, + 0.0013542972737923265, + 0.001438300358131528, + 0.0017491374164819717, + 0.0021226173266768456, + 0.0010576806962490082, + 0.0011790093267336488, + 0.002048822585493326, + 0.0013179138768464327, + 0.0006920016603544354, + 0.000967234605923295, + 0.0012751063331961632, + 0.0011879887897521257, + 0.0016490321140736341, + 0.0007376120775006711, + 0.0015084667829796672, + 0.0011217998107895255, + 0.0006543748313561082, + 0.0006904495530761778, + 0.000673303846269846, + 0.0015095884446054697, + 0.0009066596976481378, + 0.0010810700478032231, + 0.0010202425764873624, + 0.0010350419906899333, + 0.0007079873466864228, + 0.001001998083665967, + 0.0009223169763572514, + 0.0012306181015446782, + 0.007298303302377462, + 0.005270989146083593, + 0.025664275512099266, + 0.008552095852792263, + 0.007334794849157333, + 0.009676270186901093, + 0.016150308772921562, + 0.014991809614002705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.033695679157972336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This medium, eye-level shot, captured with a handheld camera, subtly panning to heighten the chaos, shows a man in his 30s at the wheel of a gleaming, aluminum, adult-sized tricycle. He stands on a wide suburban street, slick and dark beneath a gray, cloudy sky from a recent rain. The man, his brown hair plastered to his forehead, wears a vibrant yellow rain hoodie over jeans. His face perfectly conceals a comical panic, eyes wide and mouth open. He is in the midst of a frantic fishtail; both rear wheels lose traction on the wet asphalt and slide sharply to one side, setting off a dramatic arc of water. He desperately twists the handlebars in opposite directions, his body leaning hard to balance the slide, his legs flailing as he struggles to keep his feet on the pedals. The tricycle's polished aluminum frame catches the diffused light, shimmering even as it careens uncontrollably. In the background, neat suburban houses and manicured lawns, slightly out of focus, provide a calming backdrop to the unfolding chaos. The entire shot is filled with subtle film grain, adding a cinematic and nostalgic quality to a chaotic and fun moment of pure, effortless panic.", + "seed": 23249, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_13/index_1415_seed_23249_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.28688414925709366, + "vlm_logprob": -0.28688414925709366, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.28688414925709366, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23681971430778503 + ], + "opposite_logprob_tokens": [ + -1.8618197441101074 + ], + "logprob_reward": -0.23681971430778503, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.23681971430778503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045170027762651443, + 0.07206330448389053, + 0.026989439502358437, + 0.03352418541908264, + 0.03017960675060749, + 0.23905061185359955, + 0.08041571825742722, + 0.025899123400449753, + 0.07131168991327286, + 0.3753962218761444 + ], + "credit_map": [ + 0.006358286831527948, + 0.008279921486973763, + 0.004424555227160454, + 0.005874725058674812, + 0.005471248645335436, + 0.005736263003200293, + 0.004694953095167875, + 0.004330074414610863, + 0.008732556365430355, + 0.015318887308239937, + 0.019314778968691826, + 0.006935496814548969, + 0.007563746068626642, + 0.005294115748256445, + 0.003930054605007172, + 0.004973666276782751, + 0.004762961994856596, + 0.004118110053241253, + 0.0039975945837795734, + 0.003994959872215986, + 0.0026341176126152277, + 0.001935295877046883, + 0.0035195581149309874, + 0.0020268410444259644, + 0.003534311195835471, + 0.004450624343007803, + 0.006025428883731365, + 0.009304345585405827, + 0.0027079719584435225, + 0.0017274790443480015, + 0.0035563854034990072, + 0.002217639470472932, + 0.0027162411715835333, + 0.002676574047654867, + 0.0037890037056058645, + 0.0054005528800189495, + 0.006322200875729322, + 0.003339580027386546, + 0.003816369455307722, + 0.002119082026183605, + 0.020713621750473976, + 0.009000925347208977, + 0.08745720982551575, + 0.010596930049359798, + 0.01201347541064024, + 0.07749859243631363, + 0.011478321626782417, + 0.010291538201272488, + 0.011564208194613457, + 0.010007428005337715, + 0.013876684941351414, + 0.010750754736363888, + 0.01636619307100773, + 0.005862675607204437, + 0.006634374614804983, + 0.005353404209017754, + 0.0029692156240344048, + 0.0018934594700112939, + 0.0036364723928272724, + 0.0033268153201788664, + 0.004609499592334032, + 0.0036928041372448206, + 0.0032769660465419292, + 0.0024938900023698807, + 0.011251111514866352, + 0.006412946619093418, + 0.010266613215208054, + 0.007525699678808451, + 0.009233137592673302, + 0.007372172549366951, + 0.008318169973790646, + 0.010931845754384995, + 0.04129651561379433, + 0.02187865599989891, + 0.1327364295721054, + 0.09096623957157135, + 0.016719073057174683, + 0.02850642241537571, + 0.016299434006214142, + 0.026993460953235626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.23681971430778503 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8984261751174927 + ], + "opposite_logprob_tokens": [ + -0.5234261751174927 + ], + "logprob_reward": -0.8984261751174927, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8984261751174927, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18451116979122162, + 0.15044434368610382, + 0.02770492248237133, + 0.151204913854599, + 0.03711717948317528, + 0.19227010011672974, + 0.03241341561079025, + 0.026796050369739532, + 0.03810432553291321, + 0.15943363308906555 + ], + "credit_map": [ + 0.015818633139133453, + 0.02175547555088997, + 0.019138874486088753, + 0.022897332906723022, + 0.0166496392339468, + 0.02523612231016159, + 0.03984251990914345, + 0.023172559216618538, + 0.021290063858032227, + 0.02918790839612484, + 0.037609439343214035, + 0.015152736566960812, + 0.015506158582866192, + 0.010137729346752167, + 0.009078332222998142, + 0.012481965124607086, + 0.0038629905320703983, + 0.0027269558049738407, + 0.005153700243681669, + 0.0032766384538263083, + 0.003984441049396992, + 0.004118574783205986, + 0.0027355945203453302, + 0.0018460251158103347, + 0.013526915572583675, + 0.011368381790816784, + 0.028372151777148247, + 0.04848380759358406, + 0.016238320618867874, + 0.009791363961994648, + 0.01325684692710638, + 0.010167118161916733, + 0.0036808280274271965, + 0.004404653329402208, + 0.004691446200013161, + 0.004914080258458853, + 0.007557987235486507, + 0.0049062310717999935, + 0.004021840635687113, + 0.002940107136964798, + 0.014492672868072987, + 0.011061078868806362, + 0.06708265095949173, + 0.010414200834929943, + 0.010316832922399044, + 0.05554365739226341, + 0.00967433676123619, + 0.013684649951756, + 0.0038653130177408457, + 0.005014950875192881, + 0.004352835472673178, + 0.0036941636353731155, + 0.0042970613576471806, + 0.004946921952068806, + 0.003972067963331938, + 0.002270095981657505, + 0.004504386335611343, + 0.0025527707766741514, + 0.0021623370703309774, + 0.004238794092088938, + 0.0026682300958782434, + 0.004050714895129204, + 0.003963639494031668, + 0.0026551734190434217, + 0.0067879315465688705, + 0.004183330573141575, + 0.005563295911997557, + 0.0034114900045096874, + 0.004403760191053152, + 0.00408112071454525, + 0.003305245190858841, + 0.0063681467436254025, + 0.020959073677659035, + 0.010468064807355404, + 0.049325086176395416, + 0.03512578457593918, + 0.007826770655810833, + 0.01389570813626051, + 0.007889493368566036, + 0.013943638652563095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.8984261751174927 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005272056441754103 + ], + "opposite_logprob_tokens": [ + -5.255271911621094 + ], + "logprob_reward": -0.005272056441754103, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005272056441754103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06715355068445206, + 0.14101244509220123, + 0.04385126009583473, + 0.09833329916000366, + 0.02835025265812874, + 0.2828262448310852, + 0.05042232573032379, + 0.026790069416165352, + 0.02847897633910179, + 0.23278158903121948 + ], + "credit_map": [ + 0.010138537734746933, + 0.009218296967446804, + 0.006608055904507637, + 0.007900068536400795, + 0.007873665541410446, + 0.01150671299546957, + 0.007041707634925842, + 0.00686650350689888, + 0.02004345692694187, + 0.027583183720707893, + 0.03562922403216362, + 0.014522629790008068, + 0.014536725357174873, + 0.010426424443721771, + 0.007723706308752298, + 0.010547090321779251, + 0.008021335117518902, + 0.007700351532548666, + 0.006714901886880398, + 0.005530830938369036, + 0.005654546897858381, + 0.00442888168618083, + 0.003229070222005248, + 0.0025713376235216856, + 0.008397333323955536, + 0.0073364838026463985, + 0.017843419685959816, + 0.028633451089262962, + 0.011742901988327503, + 0.007114370819181204, + 0.00942373275756836, + 0.007841608487069607, + 0.0031554503366351128, + 0.0034162483643740416, + 0.003239097073674202, + 0.003503458108752966, + 0.005097340792417526, + 0.0037258195225149393, + 0.0039559234865009785, + 0.0022569133434444666, + 0.021995050832629204, + 0.01067589782178402, + 0.10390929877758026, + 0.014283871278166771, + 0.014987380243837833, + 0.0887564867734909, + 0.015814533457159996, + 0.012403694912791252, + 0.006851887796074152, + 0.005203237757086754, + 0.011357128620147705, + 0.007847782224416733, + 0.00557798333466053, + 0.005444120615720749, + 0.00469438498839736, + 0.003445797134190798, + 0.004551566205918789, + 0.002990695647895336, + 0.0036879831459373236, + 0.003731711534783244, + 0.003632695646956563, + 0.002414489397779107, + 0.003282488090917468, + 0.002498437650501728, + 0.004385376814752817, + 0.0037526225205510855, + 0.003702928079292178, + 0.0031746590975672007, + 0.00400881003588438, + 0.0025357776321470737, + 0.0024985962081700563, + 0.004420202691107988, + 0.02593487873673439, + 0.014144131913781166, + 0.07863716781139374, + 0.058385420590639114, + 0.010296682827174664, + 0.018759485334157944, + 0.009390250779688358, + 0.017233557999134064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.005272056441754103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.281362384557724 + ], + "opposite_logprob_tokens": [ + -1.4063624143600464 + ], + "logprob_reward": -0.281362384557724, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.281362384557724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0655980110168457, + 0.20778310298919678, + 0.03007102571427822, + 0.06731723994016647, + 0.0787603110074997, + 0.2647693157196045, + 0.03755693882703781, + 0.0369555838406086, + 0.023953571915626526, + 0.18723489344120026 + ], + "credit_map": [ + 0.005134419538080692, + 0.008070818148553371, + 0.0076100886799395084, + 0.006055061239749193, + 0.013935462571680546, + 0.01193192508071661, + 0.0078116473741829395, + 0.005048584192991257, + 0.025281602516770363, + 0.049290046095848083, + 0.058416251093149185, + 0.022743916139006615, + 0.02541857771575451, + 0.008810382336378098, + 0.008107442408800125, + 0.009714904241263866, + 0.006498982198536396, + 0.005935228895395994, + 0.003893383778631687, + 0.003395064501091838, + 0.003480999730527401, + 0.0026370787527412176, + 0.0021161180920898914, + 0.002114170230925083, + 0.004950887057930231, + 0.006228501908481121, + 0.012644619680941105, + 0.01802426017820835, + 0.008486596867442131, + 0.005159386899322271, + 0.006229329388588667, + 0.005593650508671999, + 0.006193832494318485, + 0.00928475521504879, + 0.01570633053779602, + 0.025526288896799088, + 0.005420851521193981, + 0.005928483325988054, + 0.005166579503566027, + 0.005533184390515089, + 0.019993307068943977, + 0.007700094021856785, + 0.10294059664011002, + 0.012739904224872589, + 0.01247728057205677, + 0.07999133318662643, + 0.012222529388964176, + 0.016704289242625237, + 0.005170480813831091, + 0.005208133719861507, + 0.006316946353763342, + 0.004189124330878258, + 0.0040449476800858974, + 0.0052132089622318745, + 0.004675989970564842, + 0.002738107228651643, + 0.004908944945782423, + 0.0050775399431586266, + 0.005202985368669033, + 0.005918973125517368, + 0.006074960343539715, + 0.002847437048330903, + 0.003755266312509775, + 0.0031694751232862473, + 0.004355587065219879, + 0.0020845234394073486, + 0.0029591929633170366, + 0.0029406012035906315, + 0.0032489297445863485, + 0.002527794800698757, + 0.0023382636718451977, + 0.003498678794130683, + 0.013668129220604897, + 0.008342282846570015, + 0.07191113382577896, + 0.04144555330276489, + 0.006141632329672575, + 0.010021383874118328, + 0.006449101492762566, + 0.02925567515194416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.281362384557724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a man in his 30s riding an adult-sized tricycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tricycle made of gleaming aluminum with a polished frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a vibrant yellow rain hoodie over jeans with brown hair plastered to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man perform a frantic fishtail where the rear wheels lose traction and slide on wet asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured with a handheld camera that subtly pans while the background shows out-of-focus suburban houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540415860712528 + ], + "opposite_logprob_tokens": [ + -4.387540340423584 + ], + "logprob_reward": -0.012540415860712528, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540415860712528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18337532877922058, + 0.15442445874214172, + 0.08385792374610901, + 0.05762190744280815, + 0.038386888802051544, + 0.14111214876174927, + 0.0349382646381855, + 0.04987003281712532, + 0.038892779499292374, + 0.21752029657363892 + ], + "credit_map": [ + 0.021859217435121536, + 0.0444767028093338, + 0.016354868188500404, + 0.021451294422149658, + 0.026893412694334984, + 0.017989899963140488, + 0.01902877166867256, + 0.015321159735321999, + 0.021050510928034782, + 0.03041354939341545, + 0.036233507096767426, + 0.01534239947795868, + 0.015759745612740517, + 0.010683564469218254, + 0.009418969973921776, + 0.015522206202149391, + 0.013033028692007065, + 0.014202306047081947, + 0.018670890480279922, + 0.011848298832774162, + 0.006571119651198387, + 0.006391564849764109, + 0.007435664534568787, + 0.005705051589757204, + 0.005207946058362722, + 0.004454796202480793, + 0.00985981710255146, + 0.0159860048443079, + 0.007220734376460314, + 0.005708349868655205, + 0.005095607601106167, + 0.004088651854544878, + 0.003132073674350977, + 0.0038316166028380394, + 0.005436946637928486, + 0.00799546018242836, + 0.006314101163297892, + 0.0035626927856355906, + 0.004686333704739809, + 0.0034276647493243217, + 0.012044072151184082, + 0.010422231629490852, + 0.04049983620643616, + 0.011166354641318321, + 0.011460023000836372, + 0.038961149752140045, + 0.009867100045084953, + 0.006691376678645611, + 0.005362166557461023, + 0.0037471093237400055, + 0.00648007495328784, + 0.0035028753336519003, + 0.0048698424361646175, + 0.00530640035867691, + 0.0032177651301026344, + 0.002452034503221512, + 0.006834793835878372, + 0.004373559728264809, + 0.005343572236597538, + 0.006677547469735146, + 0.005803333595395088, + 0.008041001856327057, + 0.007395211607217789, + 0.005401008762419224, + 0.006065797060728073, + 0.00422856118530035, + 0.004080924205482006, + 0.0048719788901507854, + 0.005020679906010628, + 0.004069856833666563, + 0.004400561563670635, + 0.0061544207856059074, + 0.021440213546156883, + 0.012798984535038471, + 0.06671037524938583, + 0.06179432198405266, + 0.009342525154352188, + 0.016533229500055313, + 0.009320409037172794, + 0.019580241292715073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.012540415860712528 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_14/index_tensor([4987, 4987])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_14/index_tensor([4987, 4987])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5289b94743b36735607f4e55bfac62d5f84ce7fd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_14/index_tensor([4987, 4987])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static wide-angle shot presents a scientific cross-section of a massive stratovolcano, rendered against a stark, featureless background, with a surreal, high-contrast digital appearance. At the base of the frame, a vast magma chamber shimmers with incandescent molten energy, its contents a churning, viscous mass of bright orange and yellow-white, serving as the primary light source. Rising from this fiery center, distinct layers of deep black solidified rock and porous volcanic rock rise, their formations delineated by sharp lines, casting long, dramatic shadows upward. A central conduit, visibly filled with pulsating, upward-moving magma, acts as a fiery artery connecting the chamber to the world above. The surface of the volcano's summit is visibly deformed, bulging upward into a strained dome under immense underground pressure, while a network of fine cracks spreads across the strained earth, some faintly glowing with a menacing red heat emanating from within. Every detail, from the crystalline texture of the rock to the subtle beams of superheated gas escaping from the cracks, is in impeccable, piercing focus, creating a powerful, clinical, and tense atmosphere of an impending catastrophic eruption.", + "seed": 13304, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_14/index_4987_seed_13304_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.025635309505742044, + "vlm_logprob": -0.025635309505742044, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.025635309505742044, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014620312489569187 + ], + "opposite_logprob_tokens": [ + -4.764620304107666 + ], + "logprob_reward": -0.014620312489569187, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.014620312489569187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18957798182964325, + 0.19332122802734375, + 0.08376717567443848, + 0.15362350642681122, + 0.07589633017778397, + 0.05083824694156647, + 0.0725073367357254, + 0.03959232196211815, + 0.10865769535303116, + 0.03221819922327995 + ], + "credit_map": [ + 0.034321852028369904, + 0.020620036870241165, + 0.017475100234150887, + 0.024732006713747978, + 0.02070833556354046, + 0.02221267856657505, + 0.022286608815193176, + 0.027221370488405228, + 0.029843106865882874, + 0.024487854912877083, + 0.02555941976606846, + 0.027524609118700027, + 0.020576149225234985, + 0.018967444077134132, + 0.027325186878442764, + 0.019037455320358276, + 0.015358446165919304, + 0.011030775494873524, + 0.007683614268898964, + 0.01314880046993494, + 0.007129068952053785, + 0.010225065052509308, + 0.00885207112878561, + 0.010339331813156605, + 0.024322092533111572, + 0.018302323296666145, + 0.016744384542107582, + 0.025036189705133438, + 0.014480591751635075, + 0.01607976108789444, + 0.01820792257785797, + 0.020450236275792122, + 0.010908020660281181, + 0.0069856299087405205, + 0.008175347000360489, + 0.00969606637954712, + 0.007882134057581425, + 0.00864685233682394, + 0.015692077577114105, + 0.007910200394690037, + 0.00760656176134944, + 0.004557354375720024, + 0.005055699963122606, + 0.0060561830177903175, + 0.006049889139831066, + 0.0051351371221244335, + 0.007884953171014786, + 0.008492469787597656, + 0.009115924127399921, + 0.009189178235828876, + 0.007769965101033449, + 0.01208802405744791, + 0.0083229448646307, + 0.007974619045853615, + 0.008026661351323128, + 0.010020023211836815, + 0.006094677373766899, + 0.004813875537365675, + 0.003657175460830331, + 0.006457615178078413, + 0.003679995657876134, + 0.004602519329637289, + 0.005692962557077408, + 0.004593500401824713, + 0.011913890019059181, + 0.012082258239388466, + 0.012675629928708076, + 0.016715157777071, + 0.017898226156830788, + 0.011840381659567356, + 0.011415223591029644, + 0.014116926118731499, + 0.004202575888484716, + 0.004257720895111561, + 0.004258202854543924, + 0.0035590040497481823, + 0.00451690936461091, + 0.0036726410035043955, + 0.0036692526191473007, + 0.0040818932466208935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.014620312489569187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035237669944763 + ], + "opposite_logprob_tokens": [ + -2.3503522872924805 + ], + "logprob_reward": -0.10035237669944763, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035237669944763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20646511018276215, + 0.1964762806892395, + 0.1786259263753891, + 0.07723338156938553, + 0.04993950575590134, + 0.05442255735397339, + 0.09067357331514359, + 0.06271862238645554, + 0.06019028276205063, + 0.023254770785570145 + ], + "credit_map": [ + 0.025204425677657127, + 0.025800541043281555, + 0.027192309498786926, + 0.025350181385874748, + 0.02117188461124897, + 0.022381596267223358, + 0.03041411004960537, + 0.0289500430226326, + 0.02720753662288189, + 0.02687476947903633, + 0.028209634125232697, + 0.02318044751882553, + 0.024458425119519234, + 0.01951093040406704, + 0.018634622916579247, + 0.028399888426065445, + 0.029352040961384773, + 0.019965754821896553, + 0.021591510623693466, + 0.020261134952306747, + 0.018672052770853043, + 0.01818845234811306, + 0.025359908118844032, + 0.02523505687713623, + 0.011000365950167179, + 0.010321362875401974, + 0.006573629565536976, + 0.013835222460329533, + 0.00845025759190321, + 0.006778338924050331, + 0.013573107309639454, + 0.006701089907437563, + 0.00745439762249589, + 0.0053482092916965485, + 0.004258765839040279, + 0.007932659238576889, + 0.0055694920010864735, + 0.0052164290100336075, + 0.00951877050101757, + 0.004640779923647642, + 0.0067551229149103165, + 0.006719513330608606, + 0.0041360086761415005, + 0.009158368222415447, + 0.007600496523082256, + 0.005178649909794331, + 0.0075682587921619415, + 0.007306134328246117, + 0.013489124365150928, + 0.008763606660068035, + 0.008095177821815014, + 0.011239206418395042, + 0.010398445650935173, + 0.010841895826160908, + 0.017188720405101776, + 0.010657384991645813, + 0.008820871822535992, + 0.00712274294346571, + 0.00592073006555438, + 0.007509162649512291, + 0.007595954462885857, + 0.007784837391227484, + 0.00922421645373106, + 0.008740101009607315, + 0.006814505439251661, + 0.007488744333386421, + 0.006441812962293625, + 0.009923291392624378, + 0.006680009420961142, + 0.006779165472835302, + 0.008567996323108673, + 0.007494750432670116, + 0.003806645516306162, + 0.0024125066120177507, + 0.001878676819615066, + 0.00360264559276402, + 0.002311547752469778, + 0.0026879985816776752, + 0.003201786894351244, + 0.0033529598731547594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.10035237669944763 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009310677996836603 + ], + "opposite_logprob_tokens": [ + -7.000931262969971 + ], + "logprob_reward": -0.0009310677996836603, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009310677996836603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19404472410678864, + 0.15286976099014282, + 0.09948799759149551, + 0.07075310498476028, + 0.05821312591433525, + 0.0738154873251915, + 0.07462958246469498, + 0.14486631751060486, + 0.07112827897071838, + 0.060191560536623 + ], + "credit_map": [ + 0.025478845462203026, + 0.02436467446386814, + 0.022719737142324448, + 0.02783099003136158, + 0.024919359013438225, + 0.018899042159318924, + 0.030445139855146408, + 0.019386935979127884, + 0.023514607921242714, + 0.0218181349337101, + 0.017424650490283966, + 0.022764354944229126, + 0.017994500696659088, + 0.015285695903003216, + 0.01857122592628002, + 0.015496588312089443, + 0.022728074342012405, + 0.010954238474369049, + 0.012537306174635887, + 0.012067065574228764, + 0.010047048330307007, + 0.01299243327230215, + 0.008793621324002743, + 0.009368205443024635, + 0.011979436501860619, + 0.007695356849581003, + 0.008172712288796902, + 0.008887218311429024, + 0.007783022243529558, + 0.007738683372735977, + 0.008716907352209091, + 0.009779764339327812, + 0.007471190299838781, + 0.00759846530854702, + 0.005477841477841139, + 0.01010878011584282, + 0.00806181225925684, + 0.005479892715811729, + 0.007974633947014809, + 0.006040511652827263, + 0.009090485982596874, + 0.008090957067906857, + 0.006586316041648388, + 0.011266152374446392, + 0.010186511091887951, + 0.008169530890882015, + 0.01091324258595705, + 0.009512287564575672, + 0.009728898294270039, + 0.006645811256021261, + 0.006496235728263855, + 0.010684527456760406, + 0.007948541082441807, + 0.00852940697222948, + 0.015611705370247364, + 0.008984457701444626, + 0.014075102284550667, + 0.02055823989212513, + 0.011146769858896732, + 0.02953292615711689, + 0.019783036783337593, + 0.014191549271345139, + 0.019529076293110847, + 0.01604962721467018, + 0.006897709798067808, + 0.008107234723865986, + 0.0073343743570148945, + 0.009957471862435341, + 0.01242666132748127, + 0.0076568229123950005, + 0.010332525707781315, + 0.008415480144321918, + 0.009601031430065632, + 0.0069289919920265675, + 0.005631696432828903, + 0.008153247646987438, + 0.009947863407433033, + 0.005722783040255308, + 0.006960682105273008, + 0.0072452714666724205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0009310677996836603 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763084227219224 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.00763084227219224, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763084227219224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10484357178211212, + 0.14261376857757568, + 0.1670781970024109, + 0.10013839602470398, + 0.08628358691930771, + 0.08479432016611099, + 0.06566669046878815, + 0.10540829598903656, + 0.058981359004974365, + 0.08419187366962433 + ], + "credit_map": [ + 0.014905964955687523, + 0.013639803975820541, + 0.011476204730570316, + 0.016091572120785713, + 0.01000827644020319, + 0.012369632720947266, + 0.01446615532040596, + 0.011885951273143291, + 0.016826516017317772, + 0.017117436975240707, + 0.01278893556445837, + 0.024951213970780373, + 0.016330284997820854, + 0.01216160412877798, + 0.024728281423449516, + 0.017709484323859215, + 0.025834061205387115, + 0.020830916240811348, + 0.01931268721818924, + 0.024273864924907684, + 0.017673859372735023, + 0.015355678275227547, + 0.02464495413005352, + 0.01915215700864792, + 0.014206880703568459, + 0.013937924057245255, + 0.011527922004461288, + 0.01687861979007721, + 0.009477958083152771, + 0.008612135425209999, + 0.014602197334170341, + 0.01089475303888321, + 0.009069892577826977, + 0.00876943115144968, + 0.008703294210135937, + 0.01399523951113224, + 0.008339304476976395, + 0.007701947819441557, + 0.0182008258998394, + 0.011503647081553936, + 0.015148268081247807, + 0.008762672543525696, + 0.007637220434844494, + 0.01090710237622261, + 0.008851232938468456, + 0.008461900986731052, + 0.014762758277356625, + 0.010263153351843357, + 0.007577656302601099, + 0.007463810499757528, + 0.006351013667881489, + 0.010839738883078098, + 0.007728928700089455, + 0.006681274622678757, + 0.012746153399348259, + 0.0062781088054180145, + 0.011137827299535275, + 0.01389987114816904, + 0.008942000567913055, + 0.021987007930874825, + 0.015111001208424568, + 0.010194564238190651, + 0.012607975862920284, + 0.011528035625815392, + 0.0055295866914093494, + 0.006291149649769068, + 0.005080141127109528, + 0.006557445041835308, + 0.009115973487496376, + 0.006376152858138084, + 0.011752955615520477, + 0.008277947083115578, + 0.014852066524326801, + 0.012598580680787563, + 0.008466489613056183, + 0.011808693408966064, + 0.007356186397373676, + 0.00897831842303276, + 0.010822745971381664, + 0.009308787994086742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.00763084227219224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004641948267817497 + ], + "opposite_logprob_tokens": [ + -5.379642009735107 + ], + "logprob_reward": -0.004641948267817497, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004641948267817497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.210819274187088, + 0.16311824321746826, + 0.1063249483704567, + 0.04642277583479881, + 0.12999187409877777, + 0.07004546374082565, + 0.06051907315850258, + 0.09080996364355087, + 0.06005880609154701, + 0.06188957765698433 + ], + "credit_map": [ + 0.0294649600982666, + 0.030203551054000854, + 0.025914954021573067, + 0.030451811850070953, + 0.021763933822512627, + 0.0227553378790617, + 0.023697247728705406, + 0.026567474007606506, + 0.027491021901369095, + 0.02148899994790554, + 0.015663938596844673, + 0.02734595723450184, + 0.01602630317211151, + 0.017502455040812492, + 0.020066067576408386, + 0.017533494159579277, + 0.017394378781318665, + 0.012859848327934742, + 0.012651114724576473, + 0.009593682363629341, + 0.010774927213788033, + 0.0118442727252841, + 0.014907089993357658, + 0.01629962958395481, + 0.005359055008739233, + 0.005075064487755299, + 0.004268016666173935, + 0.007273946888744831, + 0.005442908965051174, + 0.0043409401550889015, + 0.008553571067750454, + 0.006109272595494986, + 0.01694604568183422, + 0.01649620197713375, + 0.012560303322970867, + 0.02361293137073517, + 0.012012410908937454, + 0.01378029678016901, + 0.017811112105846405, + 0.01677258312702179, + 0.006784443277865648, + 0.007841590791940689, + 0.0075198691338300705, + 0.010507035069167614, + 0.010548283345997334, + 0.007580867037177086, + 0.01180005632340908, + 0.007463324349373579, + 0.006767713464796543, + 0.006092787720263004, + 0.005571102257817984, + 0.008992613293230534, + 0.007447218056768179, + 0.0070863766595721245, + 0.008201965130865574, + 0.01035929936915636, + 0.012108891271054745, + 0.012269517406821251, + 0.008135195821523666, + 0.012096005491912365, + 0.011556483805179596, + 0.008979229256510735, + 0.009708068333566189, + 0.015956569463014603, + 0.006344547029584646, + 0.006096863653510809, + 0.005915291607379913, + 0.007375575602054596, + 0.009189466945827007, + 0.005853861104696989, + 0.011704710312187672, + 0.007578486576676369, + 0.006376206874847412, + 0.007708477787673473, + 0.005177697632461786, + 0.010010440833866596, + 0.0061655351892113686, + 0.005701755173504353, + 0.009627800434827805, + 0.01112165953963995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.004641948267817497 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static wide-angle shot presents a scientific cross-section of a massive stratovolcano, rendered against a stark, featureless background, with a surreal, high-contrast digital appearance. At the base of the frame, a vast magma chamber shimmers with incandescent molten energy, its contents a churning, viscous mass of bright orange and yellow-white, serving as the primary light source. Rising from this fiery center, distinct layers of deep black solidified rock and porous volcanic rock rise, their formations delineated by sharp lines, casting long, dramatic shadows upward. A central conduit, visibly filled with pulsating, upward-moving magma, acts as a fiery artery connecting the chamber to the world above. The surface of the volcano's summit is visibly deformed, bulging upward into a strained dome under immense underground pressure, while a network of fine cracks spreads across the strained earth, some faintly glowing with a menacing red heat emanating from within. Every detail, from the crystalline texture of the rock to the subtle beams of superheated gas escaping from the cracks, is in impeccable, piercing focus, creating a powerful, clinical, and tense atmosphere of an impending catastrophic eruption.", + "seed": 13305, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_14/index_4987_seed_13305_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07590587476734072, + "vlm_logprob": -0.07590587476734072, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.07590587476734072, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1698429137468338 + ], + "opposite_logprob_tokens": [ + -1.9198429584503174 + ], + "logprob_reward": -0.1698429137468338, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1698429137468338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1648983657360077, + 0.0633939728140831, + 0.06554746627807617, + 0.04627044126391411, + 0.04704324156045914, + 0.10076682269573212, + 0.12925434112548828, + 0.1541762799024582, + 0.05922824889421463, + 0.16942080855369568 + ], + "credit_map": [ + 0.022930506616830826, + 0.03361552953720093, + 0.012978513725101948, + 0.010170859284698963, + 0.027291495352983475, + 0.02346421591937542, + 0.019921226426959038, + 0.014526020735502243, + 0.006391663569957018, + 0.012387662194669247, + 0.007374526001513004, + 0.0059648179449141026, + 0.006850789301097393, + 0.009978481568396091, + 0.0071611045859754086, + 0.007284930907189846, + 0.008478496223688126, + 0.009416891261935234, + 0.007394416257739067, + 0.006328072398900986, + 0.007927543483674526, + 0.009778033941984177, + 0.009628660045564175, + 0.006595353130251169, + 0.009769636206328869, + 0.007446387317031622, + 0.00424895528703928, + 0.0036498631816357374, + 0.0034167426638305187, + 0.007323505822569132, + 0.0035450742579996586, + 0.006870273966342211, + 0.005619253031909466, + 0.008162935264408588, + 0.006833287887275219, + 0.005850261077284813, + 0.004179093521088362, + 0.005629875231534243, + 0.00544182863086462, + 0.005326706450432539, + 0.0147892776876688, + 0.015271863900125027, + 0.016301341354846954, + 0.011558184400200844, + 0.013270998373627663, + 0.012699823826551437, + 0.008876892738044262, + 0.00799844041466713, + 0.01312330923974514, + 0.022048601880669594, + 0.011309780180454254, + 0.015463151037693024, + 0.02079096995294094, + 0.015248212963342667, + 0.012671100907027721, + 0.018599217757582664, + 0.01581740193068981, + 0.02096930332481861, + 0.023814216256141663, + 0.019267519935965538, + 0.01876324601471424, + 0.026950931176543236, + 0.008581114932894707, + 0.02001253142952919, + 0.006133455783128738, + 0.007391579449176788, + 0.006831741891801357, + 0.008025983348488808, + 0.007479677442461252, + 0.008304104208946228, + 0.008594237267971039, + 0.006467469502240419, + 0.014548399485647678, + 0.02650698833167553, + 0.019953899085521698, + 0.018790192902088165, + 0.025161076337099075, + 0.019837813451886177, + 0.016947198659181595, + 0.027675243094563484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.1698429137468338 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993655800819397 + ], + "opposite_logprob_tokens": [ + -1.8049365282058716 + ], + "logprob_reward": -0.17993655800819397, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993655800819397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10176561772823334, + 0.09563134610652924, + 0.09887345880270004, + 0.11278308182954788, + 0.09210092574357986, + 0.10459482669830322, + 0.0790724977850914, + 0.1351684182882309, + 0.06610704213380814, + 0.11390279233455658 + ], + "credit_map": [ + 0.01577652059495449, + 0.014649385586380959, + 0.010265138000249863, + 0.010540029034018517, + 0.014015903696417809, + 0.01673010364174843, + 0.010393989272415638, + 0.009394547902047634, + 0.016953209415078163, + 0.01900765672326088, + 0.01049073226749897, + 0.009201119653880596, + 0.007800820749253035, + 0.015681466087698936, + 0.007402355317026377, + 0.00909398403018713, + 0.012159495614469051, + 0.020697245374321938, + 0.008304898627102375, + 0.009697763249278069, + 0.01123877801001072, + 0.018124915659427643, + 0.00974626000970602, + 0.00890409667044878, + 0.014740468002855778, + 0.020435065031051636, + 0.012122120708227158, + 0.010497386567294598, + 0.010955116711556911, + 0.018410351127386093, + 0.010904544964432716, + 0.014718027785420418, + 0.012821344658732414, + 0.014673727564513683, + 0.013863642700016499, + 0.007823220454156399, + 0.009101410396397114, + 0.014440563507378101, + 0.008504319936037064, + 0.010872683487832546, + 0.01393045112490654, + 0.016963863745331764, + 0.01666184514760971, + 0.013775285333395004, + 0.012167845852673054, + 0.011843518353998661, + 0.010592465288937092, + 0.00865954253822565, + 0.007580806501209736, + 0.010141739621758461, + 0.012893595732748508, + 0.008972222916781902, + 0.011259729973971844, + 0.009629292413592339, + 0.007404885720461607, + 0.011190217919647694, + 0.014159956946969032, + 0.017789756879210472, + 0.02116643451154232, + 0.016986183822155, + 0.01388041116297245, + 0.020392635837197304, + 0.010527113452553749, + 0.02026592381298542, + 0.007520935032516718, + 0.009970944374799728, + 0.009164481423795223, + 0.00763221038505435, + 0.008660338819026947, + 0.010192553512752056, + 0.006157932803034782, + 0.006807640194892883, + 0.014380252920091152, + 0.014872095547616482, + 0.013015607371926308, + 0.012542393989861012, + 0.016650835052132607, + 0.013962360098958015, + 0.011214756406843662, + 0.017264481633901596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.17993655800819397 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001527215470559895 + ], + "opposite_logprob_tokens": [ + -6.501527309417725 + ], + "logprob_reward": -0.001527215470559895, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001527215470559895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15185745060443878, + 0.21982505917549133, + 0.11623164266347885, + 0.05529678240418434, + 0.06216318905353546, + 0.11860215663909912, + 0.07570973038673401, + 0.09939712285995483, + 0.05028869956731796, + 0.050628188997507095 + ], + "credit_map": [ + 0.021515825763344765, + 0.02148520201444626, + 0.023686692118644714, + 0.01565682329237461, + 0.016066409647464752, + 0.020209550857543945, + 0.01671404018998146, + 0.016522904857993126, + 0.035418055951595306, + 0.04396510124206543, + 0.027884256094694138, + 0.017735984176397324, + 0.018415382131934166, + 0.023974958807229996, + 0.01912640407681465, + 0.033304907381534576, + 0.014912259764969349, + 0.026693567633628845, + 0.008780407719314098, + 0.00929730199277401, + 0.01122212689369917, + 0.02181859314441681, + 0.012533609755337238, + 0.010973765514791012, + 0.009711511433124542, + 0.010190817527472973, + 0.00581224774941802, + 0.004515995737165213, + 0.0075170681811869144, + 0.007818760350346565, + 0.005190808326005936, + 0.004539568908512592, + 0.006862185429781675, + 0.008304408751428127, + 0.006316179875284433, + 0.0074128820560872555, + 0.007727848831564188, + 0.007035191170871258, + 0.009817851707339287, + 0.008686638437211514, + 0.014339569956064224, + 0.018696440383791924, + 0.019313568249344826, + 0.01827160455286503, + 0.013360614888370037, + 0.013318636454641819, + 0.009731942787766457, + 0.011569764465093613, + 0.008706818334758282, + 0.010586783289909363, + 0.008533096872270107, + 0.007880757562816143, + 0.010358300060033798, + 0.008560512214899063, + 0.008490016683936119, + 0.012593444436788559, + 0.010257950983941555, + 0.014723141677677631, + 0.020653391256928444, + 0.013158907182514668, + 0.009172169491648674, + 0.011914524249732494, + 0.007287360727787018, + 0.012229669839143753, + 0.006527965422719717, + 0.006833015009760857, + 0.005580488126724958, + 0.005212992429733276, + 0.007531309500336647, + 0.008050384931266308, + 0.004560275934636593, + 0.005992257967591286, + 0.005278218071907759, + 0.005609904881566763, + 0.0065866680815815926, + 0.006510497070848942, + 0.005334610119462013, + 0.007171496283262968, + 0.0037857491988688707, + 0.01035104040056467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.001527215470559895 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585810765624046 + ], + "opposite_logprob_tokens": [ + -3.895585775375366 + ], + "logprob_reward": -0.020585810765624046, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585810765624046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10462266206741333, + 0.10935689508914948, + 0.12166769802570343, + 0.09013835340738297, + 0.10956046730279922, + 0.19863180816173553, + 0.1390591412782669, + 0.034825436770915985, + 0.05724816396832466, + 0.034889426082372665 + ], + "credit_map": [ + 0.013637768104672432, + 0.018726598471403122, + 0.01523295883089304, + 0.008882435970008373, + 0.010474210605025291, + 0.013344328850507736, + 0.009732880629599094, + 0.014591475948691368, + 0.014097146689891815, + 0.034058429300785065, + 0.008106064051389694, + 0.008958005346357822, + 0.007406744174659252, + 0.019275881350040436, + 0.00864274613559246, + 0.00881187990307808, + 0.017919212579727173, + 0.017936764284968376, + 0.012897534295916557, + 0.011020257137715816, + 0.011890829540789127, + 0.01495390571653843, + 0.016925260424613953, + 0.01812393218278885, + 0.016220582649111748, + 0.01754647120833397, + 0.008192922919988632, + 0.0083976611495018, + 0.011521623469889164, + 0.011727535165846348, + 0.010768585838377476, + 0.005762971006333828, + 0.01300530694425106, + 0.01573151908814907, + 0.012967913411557674, + 0.014533765614032745, + 0.013917774893343449, + 0.011756771244108677, + 0.014962210319936275, + 0.012685205787420273, + 0.02362373098731041, + 0.031968481838703156, + 0.03655001148581505, + 0.030866168439388275, + 0.020091507583856583, + 0.02415645681321621, + 0.012311571277678013, + 0.019063876941800117, + 0.016938187181949615, + 0.018733803182840347, + 0.022050995379686356, + 0.019358323886990547, + 0.017108358442783356, + 0.016940098255872726, + 0.010855963453650475, + 0.017073428258299828, + 0.003748138900846243, + 0.005065310746431351, + 0.005851258523762226, + 0.004535786807537079, + 0.004141641315072775, + 0.004310031421482563, + 0.0032979752868413925, + 0.003875289810821414, + 0.005746443755924702, + 0.007563435472548008, + 0.005355719476938248, + 0.006823855452239513, + 0.01079468708485365, + 0.0072807432152330875, + 0.005330459214746952, + 0.008352817967534065, + 0.0028965536039322615, + 0.004019906744360924, + 0.004974681418389082, + 0.004543235059827566, + 0.0037842176388949156, + 0.005595484282821417, + 0.003081338480114937, + 0.005994007922708988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.020585810765624046 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636875845491886 + ], + "opposite_logprob_tokens": [ + -4.882637023925781 + ], + "logprob_reward": -0.007636875845491886, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636875845491886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11891979724168777, + 0.12239593267440796, + 0.08076400309801102, + 0.061921410262584686, + 0.05555232614278793, + 0.09413013607263565, + 0.1850915402173996, + 0.1494045853614807, + 0.06735090911388397, + 0.06446928530931473 + ], + "credit_map": [ + 0.015705348923802376, + 0.021157149225473404, + 0.011812552809715271, + 0.00928263459354639, + 0.016128050163388252, + 0.014775155112147331, + 0.014946818351745605, + 0.015112090855836868, + 0.016505105420947075, + 0.023151040077209473, + 0.01605277881026268, + 0.010524715296924114, + 0.00947954598814249, + 0.0151998745277524, + 0.012475507333874702, + 0.01900736801326275, + 0.010943969711661339, + 0.013507519848644733, + 0.008233275264501572, + 0.008349036797881126, + 0.008902854286134243, + 0.013957750052213669, + 0.008465009741485119, + 0.008404592052102089, + 0.01017562672495842, + 0.011678598821163177, + 0.007682712748646736, + 0.00481400266289711, + 0.005177921615540981, + 0.011495028622448444, + 0.0053153010085225105, + 0.00558222271502018, + 0.006249844096601009, + 0.008310750126838684, + 0.006911781150847673, + 0.006753878202289343, + 0.006431028246879578, + 0.0067292992025613785, + 0.00794210471212864, + 0.006223641335964203, + 0.012142607010900974, + 0.01112905889749527, + 0.013669266365468502, + 0.013501618057489395, + 0.010234556160867214, + 0.012043986469507217, + 0.008272211067378521, + 0.013136823661625385, + 0.01942088082432747, + 0.02363273873925209, + 0.024874024093151093, + 0.025329435244202614, + 0.02127726376056671, + 0.02483983337879181, + 0.016079550608992577, + 0.029637819156050682, + 0.018030263483524323, + 0.020907949656248093, + 0.031074702739715576, + 0.020853158086538315, + 0.013400963507592678, + 0.017440106719732285, + 0.009941629134118557, + 0.017755815759301186, + 0.00802894402295351, + 0.010292308405041695, + 0.007397559005767107, + 0.006912401877343655, + 0.011672214604914188, + 0.009578105993568897, + 0.006518649403005838, + 0.006950723007321358, + 0.006715521216392517, + 0.008102623745799065, + 0.00867245253175497, + 0.00761403189972043, + 0.009663369506597519, + 0.008347700349986553, + 0.004787818528711796, + 0.010565764270722866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.007636875845491886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static wide-angle shot presents a scientific cross-section of a massive stratovolcano, rendered against a stark, featureless background, with a surreal, high-contrast digital appearance. At the base of the frame, a vast magma chamber shimmers with incandescent molten energy, its contents a churning, viscous mass of bright orange and yellow-white, serving as the primary light source. Rising from this fiery center, distinct layers of deep black solidified rock and porous volcanic rock rise, their formations delineated by sharp lines, casting long, dramatic shadows upward. A central conduit, visibly filled with pulsating, upward-moving magma, acts as a fiery artery connecting the chamber to the world above. The surface of the volcano's summit is visibly deformed, bulging upward into a strained dome under immense underground pressure, while a network of fine cracks spreads across the strained earth, some faintly glowing with a menacing red heat emanating from within. Every detail, from the crystalline texture of the rock to the subtle beams of superheated gas escaping from the cracks, is in impeccable, piercing focus, creating a powerful, clinical, and tense atmosphere of an impending catastrophic eruption.", + "seed": 13306, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_14/index_4987_seed_13306_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06621407247148454, + "vlm_logprob": -0.06621407247148454, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.06621407247148454, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12672212719917297 + ], + "opposite_logprob_tokens": [ + -2.2517220973968506 + ], + "logprob_reward": -0.12672212719917297, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12672212719917297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055551331490278244, + 0.1466977894306183, + 0.04309037700295448, + 0.035765379667282104, + 0.08830003440380096, + 0.31327563524246216, + 0.1641223132610321, + 0.04504840448498726, + 0.06843721866607666, + 0.039711516350507736 + ], + "credit_map": [ + 0.004634334705770016, + 0.004981894511729479, + 0.01151445135474205, + 0.006757454480975866, + 0.006893297657370567, + 0.004958972334861755, + 0.00941463652998209, + 0.0063962931744754314, + 0.01332732755690813, + 0.012440359219908714, + 0.0340309776365757, + 0.01483768317848444, + 0.015821125358343124, + 0.01673423871397972, + 0.018809504806995392, + 0.020696565508842468, + 0.005433219019323587, + 0.0053971135057508945, + 0.005644806660711765, + 0.0039042001590132713, + 0.005839404184371233, + 0.005558664910495281, + 0.00703289033845067, + 0.004280075430870056, + 0.003454208606854081, + 0.00418392289429903, + 0.006440706085413694, + 0.004094287287443876, + 0.004601430147886276, + 0.0033230616245418787, + 0.005628878716379404, + 0.004038882441818714, + 0.007089320570230484, + 0.00800708681344986, + 0.011947546154260635, + 0.014419149607419968, + 0.010646242648363113, + 0.01277525257319212, + 0.011867905966937542, + 0.011547520756721497, + 0.014909304678440094, + 0.02349734678864479, + 0.038574326783418655, + 0.06891150772571564, + 0.034774962812662125, + 0.018841393291950226, + 0.08292294293642044, + 0.03084384649991989, + 0.015420328825712204, + 0.016809770837426186, + 0.028827592730522156, + 0.022297650575637817, + 0.015656517818570137, + 0.02764415554702282, + 0.019871024414896965, + 0.017595279961824417, + 0.00507425656542182, + 0.004910309333354235, + 0.011748346500098705, + 0.004790480248630047, + 0.003602567594498396, + 0.005840549245476723, + 0.0052884905599057674, + 0.0037934028077870607, + 0.003950110636651516, + 0.005640777293592691, + 0.010975785553455353, + 0.013898713514208794, + 0.007408455945551395, + 0.004020977299660444, + 0.01626311056315899, + 0.006279288791120052, + 0.0026688179932534695, + 0.004104832652956247, + 0.006046553608030081, + 0.003712206380441785, + 0.006656482815742493, + 0.0034097374882549047, + 0.008620196022093296, + 0.00449268938973546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.12672212719917297 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799643337726593 + ], + "opposite_logprob_tokens": [ + -1.804964303970337 + ], + "logprob_reward": -0.1799643337726593, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799643337726593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06958240270614624, + 0.07899116724729538, + 0.08270388841629028, + 0.05224861577153206, + 0.10412611067295074, + 0.12297631800174713, + 0.19176442921161652, + 0.22989268600940704, + 0.043807629495859146, + 0.02390671893954277 + ], + "credit_map": [ + 0.00898097176104784, + 0.0075156837701797485, + 0.009083634242415428, + 0.010792366228997707, + 0.009100610390305519, + 0.00668316800147295, + 0.009137159213423729, + 0.008288812823593616, + 0.006222320720553398, + 0.009069838561117649, + 0.010363185778260231, + 0.007904020138084888, + 0.011158625595271587, + 0.011662920005619526, + 0.010936971753835678, + 0.011673294007778168, + 0.010549875907599926, + 0.0063761197961866856, + 0.015679989010095596, + 0.009256142191588879, + 0.008990411646664143, + 0.013984412886202335, + 0.008038587868213654, + 0.009828356094658375, + 0.004729267209768295, + 0.006001729052513838, + 0.009661737829446793, + 0.006218457594513893, + 0.007008832413703203, + 0.004355982411652803, + 0.008552652783691883, + 0.0057199569419026375, + 0.009457306936383247, + 0.013305563479661942, + 0.014675267040729523, + 0.015262175351381302, + 0.012988686561584473, + 0.017465371638536453, + 0.009962312877178192, + 0.01100943237543106, + 0.007028333842754364, + 0.010372494347393513, + 0.017439287155866623, + 0.02519000694155693, + 0.013623486272990704, + 0.008733638562262058, + 0.027619006112217903, + 0.01297006942331791, + 0.01811976544559002, + 0.013954772613942623, + 0.04547866806387901, + 0.020598597824573517, + 0.017721349373459816, + 0.033203575760126114, + 0.022016700357198715, + 0.020671023055911064, + 0.03061414510011673, + 0.018419941887259483, + 0.05986814200878143, + 0.027185553684830666, + 0.01894429139792919, + 0.034779828041791916, + 0.019729794934391975, + 0.020351000130176544, + 0.002534864703193307, + 0.004241538234055042, + 0.005310590844601393, + 0.009381997399032116, + 0.004888620227575302, + 0.0024217639584094286, + 0.010880308225750923, + 0.004147946834564209, + 0.0019667292945086956, + 0.0022287967149168253, + 0.0037121037021279335, + 0.0036812678445130587, + 0.0026358128525316715, + 0.0022361185401678085, + 0.004796485882252455, + 0.0026494076009839773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.1799643337726593 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015280486550182104 + ], + "opposite_logprob_tokens": [ + -6.501528263092041 + ], + "logprob_reward": -0.0015280486550182104, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015280486550182104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15876775979995728, + 0.11735773831605911, + 0.04783934727311134, + 0.03908347710967064, + 0.12443394958972931, + 0.2807840406894684, + 0.045983243733644485, + 0.04739660769701004, + 0.08593949675559998, + 0.05241433531045914 + ], + "credit_map": [ + 0.01560854073613882, + 0.010057523846626282, + 0.04723632335662842, + 0.017837030813097954, + 0.015826677903532982, + 0.011114425025880337, + 0.026849249377846718, + 0.014237998984754086, + 0.014073694124817848, + 0.011571523733437061, + 0.027308199554681778, + 0.013591351918876171, + 0.011769124306738377, + 0.01473724190145731, + 0.011424100957810879, + 0.012882507406175137, + 0.004898186773061752, + 0.004975651390850544, + 0.007469470612704754, + 0.004795084707438946, + 0.007517578545957804, + 0.004838508553802967, + 0.008038894273340702, + 0.005305971018970013, + 0.0034751438070088625, + 0.005452215205878019, + 0.0055034710094332695, + 0.005510503891855478, + 0.005214002449065447, + 0.003169226460158825, + 0.00657916534692049, + 0.004179749637842178, + 0.010070202872157097, + 0.011740483343601227, + 0.016735300421714783, + 0.02096380665898323, + 0.01454942487180233, + 0.017029667273163795, + 0.017054064199328423, + 0.016291003674268723, + 0.013884710147976875, + 0.021841825917363167, + 0.03235715627670288, + 0.060965120792388916, + 0.03130399435758591, + 0.015987152233719826, + 0.07506556808948517, + 0.0293784961104393, + 0.004747128114104271, + 0.003885718295350671, + 0.012249569408595562, + 0.005446729715913534, + 0.0037326731253415346, + 0.0056622214615345, + 0.004975020885467529, + 0.005284180864691734, + 0.005670296959578991, + 0.003960711881518364, + 0.014699635095894337, + 0.004150772001594305, + 0.0029525170102715492, + 0.00666454341262579, + 0.003949062433093786, + 0.0053490702994167805, + 0.0046682232059538364, + 0.006901846267282963, + 0.012363137677311897, + 0.016596030443906784, + 0.009773951955139637, + 0.0051857163198292255, + 0.022410951554775238, + 0.008039642125368118, + 0.004086779430508614, + 0.005739306565374136, + 0.008856208063662052, + 0.006521062459796667, + 0.006892007775604725, + 0.005069865845143795, + 0.00893706176429987, + 0.0063120415434241295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0015280486550182104 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206347987055779 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206347987055779, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206347987055779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0752720832824707, + 0.14003898203372955, + 0.03462390974164009, + 0.027271278202533722, + 0.20748861134052277, + 0.17965689301490784, + 0.09607865661382675, + 0.11488734930753708, + 0.06679891794919968, + 0.05788330361247063 + ], + "credit_map": [ + 0.006748768500983715, + 0.0063433717004954815, + 0.01639828272163868, + 0.009002982638776302, + 0.010254533030092716, + 0.007518248166888952, + 0.009660044685006142, + 0.009345851838588715, + 0.013873323798179626, + 0.013636462390422821, + 0.028556283563375473, + 0.014712943695485592, + 0.015201016329228878, + 0.016264230012893677, + 0.017772208899259567, + 0.020022518932819366, + 0.0037562809884548187, + 0.0045523010194301605, + 0.006737600080668926, + 0.003323320299386978, + 0.004553664941340685, + 0.0034535154700279236, + 0.003973029088228941, + 0.004274201113730669, + 0.002642425475642085, + 0.0038065381813794374, + 0.004327025264501572, + 0.0029032472521066666, + 0.004348109941929579, + 0.0032775215804576874, + 0.0030227969400584698, + 0.002943613100796938, + 0.01486198790371418, + 0.021264495328068733, + 0.026722446084022522, + 0.035385146737098694, + 0.02534233219921589, + 0.030390921980142593, + 0.026444371789693832, + 0.027076907455921173, + 0.009816676378250122, + 0.014178868383169174, + 0.020069563761353493, + 0.03602723032236099, + 0.020422600209712982, + 0.012513347901403904, + 0.04512789100408554, + 0.021500715985894203, + 0.0078009734861552715, + 0.0114588038995862, + 0.010930887423455715, + 0.015503846108913422, + 0.01095966249704361, + 0.013531441800296307, + 0.015006315894424915, + 0.010886725969612598, + 0.009555181488394737, + 0.012853337451815605, + 0.014452089555561543, + 0.016052542254328728, + 0.013500443659722805, + 0.018544647842645645, + 0.015294612385332584, + 0.014634489081799984, + 0.003725379705429077, + 0.005802681669592857, + 0.008386703208088875, + 0.012568728066980839, + 0.0076894136145710945, + 0.004934364929795265, + 0.01735505275428295, + 0.006336595397442579, + 0.006362522952258587, + 0.007937014102935791, + 0.009672076441347599, + 0.005462672561407089, + 0.008686196990311146, + 0.006015930790454149, + 0.007459778804332018, + 0.006287110038101673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.014206347987055779 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649504743516445 + ], + "opposite_logprob_tokens": [ + -4.7586493492126465 + ], + "logprob_reward": -0.008649504743516445, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649504743516445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11490104347467422, + 0.05394311994314194, + 0.06656402349472046, + 0.05181827023625374, + 0.08051103353500366, + 0.229155495762825, + 0.10082381963729858, + 0.08580218255519867, + 0.09201463311910629, + 0.12446633726358414 + ], + "credit_map": [ + 0.011920968070626259, + 0.013595388270914555, + 0.015139936469495296, + 0.02203349582850933, + 0.01130246464163065, + 0.009838695637881756, + 0.01704888977110386, + 0.01402120292186737, + 0.008052504621446133, + 0.004659201949834824, + 0.008227542042732239, + 0.006317815277725458, + 0.006001729983836412, + 0.005029117688536644, + 0.007916847243905067, + 0.007738358806818724, + 0.011459493078291416, + 0.007744372822344303, + 0.010757142677903175, + 0.006826275959610939, + 0.007670877501368523, + 0.008083976805210114, + 0.007077061105519533, + 0.006944822613149881, + 0.006071883719414473, + 0.006281349807977676, + 0.008349226787686348, + 0.005908535327762365, + 0.005552308633923531, + 0.006061102729290724, + 0.0070311687886714935, + 0.006562698632478714, + 0.008480130694806576, + 0.00841674767434597, + 0.010196682065725327, + 0.011507541872560978, + 0.010640825144946575, + 0.01404271088540554, + 0.00876628328114748, + 0.008460109122097492, + 0.013950636610388756, + 0.01882198452949524, + 0.02533012442290783, + 0.050738926976919174, + 0.02327198162674904, + 0.015584699809551239, + 0.05706347897648811, + 0.024393675848841667, + 0.010123245418071747, + 0.009286210872232914, + 0.02006455697119236, + 0.012273344211280346, + 0.008541794493794441, + 0.016477523371577263, + 0.011985099874436855, + 0.012072048150002956, + 0.007666511926800013, + 0.00806708075106144, + 0.015076159499585629, + 0.01054130308330059, + 0.009278868325054646, + 0.010654761455953121, + 0.016101529821753502, + 0.008415967226028442, + 0.005757089238613844, + 0.007985003292560577, + 0.010363124310970306, + 0.018838541582226753, + 0.011556196957826614, + 0.005536813288927078, + 0.023481858894228935, + 0.008496013469994068, + 0.009326563216745853, + 0.012797195464372635, + 0.026340944692492485, + 0.015662575140595436, + 0.012126894667744637, + 0.011646797880530357, + 0.02419178932905197, + 0.012373572215437889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.008649504743516445 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static wide-angle shot presents a scientific cross-section of a massive stratovolcano, rendered against a stark, featureless background, with a surreal, high-contrast digital appearance. At the base of the frame, a vast magma chamber shimmers with incandescent molten energy, its contents a churning, viscous mass of bright orange and yellow-white, serving as the primary light source. Rising from this fiery center, distinct layers of deep black solidified rock and porous volcanic rock rise, their formations delineated by sharp lines, casting long, dramatic shadows upward. A central conduit, visibly filled with pulsating, upward-moving magma, acts as a fiery artery connecting the chamber to the world above. The surface of the volcano's summit is visibly deformed, bulging upward into a strained dome under immense underground pressure, while a network of fine cracks spreads across the strained earth, some faintly glowing with a menacing red heat emanating from within. Every detail, from the crystalline texture of the rock to the subtle beams of superheated gas escaping from the cracks, is in impeccable, piercing focus, creating a powerful, clinical, and tense atmosphere of an impending catastrophic eruption.", + "seed": 13307, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_14/index_4987_seed_13307_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04861524135340005, + "vlm_logprob": -0.04861524135340005, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.04861524135340005, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06352519243955612 + ], + "opposite_logprob_tokens": [ + -2.9385251998901367 + ], + "logprob_reward": -0.06352519243955612, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06352519243955612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15393175184726715, + 0.13098552823066711, + 0.10368898510932922, + 0.06525513529777527, + 0.07193661481142044, + 0.05156022682785988, + 0.08254803717136383, + 0.14065931737422943, + 0.1200440302491188, + 0.07939037680625916 + ], + "credit_map": [ + 0.02527615800499916, + 0.011739701963961124, + 0.02621256187558174, + 0.01833159476518631, + 0.015464242547750473, + 0.022036058828234673, + 0.021202752366662025, + 0.01366867870092392, + 0.016531113535165787, + 0.011160393245518208, + 0.02492002211511135, + 0.015840629115700722, + 0.01499162707477808, + 0.018168248236179352, + 0.02103309892117977, + 0.008340389467775822, + 0.011974465101957321, + 0.007939544506371021, + 0.016659488901495934, + 0.012834367342293262, + 0.011663500219583511, + 0.014288037084043026, + 0.01928005740046501, + 0.009049510583281517, + 0.009123505093157291, + 0.006605017464607954, + 0.01247011311352253, + 0.006935258861631155, + 0.00589601369574666, + 0.01029957178980112, + 0.007500980980694294, + 0.006424664054065943, + 0.008152314461767673, + 0.005746070295572281, + 0.010777083225548267, + 0.009832306765019894, + 0.007899545133113861, + 0.009923365898430347, + 0.010058571584522724, + 0.009547356516122818, + 0.007179670035839081, + 0.0052832625806331635, + 0.005770951975136995, + 0.0064851222559809685, + 0.006232682149857283, + 0.006632894277572632, + 0.009083116427063942, + 0.004892521537840366, + 0.00789345521479845, + 0.008220096118748188, + 0.011810468509793282, + 0.01170398760586977, + 0.01082556415349245, + 0.0116222333163023, + 0.012342671863734722, + 0.008129552006721497, + 0.01246653962880373, + 0.01288686040788889, + 0.029595669358968735, + 0.01783108524978161, + 0.013705393299460411, + 0.022187426686286926, + 0.015720678493380547, + 0.01626564934849739, + 0.01519690826535225, + 0.013492914848029613, + 0.016452563926577568, + 0.015052598901093006, + 0.013338173739612103, + 0.017769906669855118, + 0.016977813094854355, + 0.011763143353164196, + 0.010157874785363674, + 0.008069394156336784, + 0.016135333105921745, + 0.009199989959597588, + 0.006909879390150309, + 0.012638184241950512, + 0.007854772731661797, + 0.00842494610697031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.06352519243955612 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16041186451911926 + ], + "opposite_logprob_tokens": [ + -1.9104118347167969 + ], + "logprob_reward": -0.16041186451911926, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16041186451911926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16716907918453217, + 0.11482919007539749, + 0.07044786959886551, + 0.07485806941986084, + 0.12204609811306, + 0.08546280115842819, + 0.113349549472332, + 0.08587248623371124, + 0.057789310812950134, + 0.10817551612854004 + ], + "credit_map": [ + 0.028343450278043747, + 0.011897060088813305, + 0.025826554745435715, + 0.01902247965335846, + 0.017460232600569725, + 0.025679470971226692, + 0.02382955513894558, + 0.015110285952687263, + 0.015165664255619049, + 0.009166444651782513, + 0.014058209955692291, + 0.012683036737143993, + 0.014735234901309013, + 0.015540585853159428, + 0.020677179098129272, + 0.012802841141819954, + 0.011087176389992237, + 0.00555815314874053, + 0.009204418398439884, + 0.010734356008470058, + 0.007794157601892948, + 0.010619943961501122, + 0.009094848297536373, + 0.006354817189276218, + 0.007165446877479553, + 0.006622083485126495, + 0.008716431446373463, + 0.011368535459041595, + 0.007960769347846508, + 0.01034080795943737, + 0.009132392704486847, + 0.013551612384617329, + 0.018044402822852135, + 0.009898018091917038, + 0.0153899434953928, + 0.0114876264706254, + 0.014928719960153103, + 0.017443086951971054, + 0.021975219249725342, + 0.012879088521003723, + 0.011500461027026176, + 0.009115941822528839, + 0.015451309271156788, + 0.008341610431671143, + 0.01385758351534605, + 0.013034875504672527, + 0.00757712172344327, + 0.006583909038454294, + 0.011971416883170605, + 0.008886315859854221, + 0.015387623570859432, + 0.0157562717795372, + 0.013389627449214458, + 0.015131982043385506, + 0.021316232159733772, + 0.011510093696415424, + 0.007845772430300713, + 0.011062384583055973, + 0.009034084156155586, + 0.010301974602043629, + 0.011628357693552971, + 0.010568162426352501, + 0.015447497367858887, + 0.009984264150261879, + 0.0065081059001386166, + 0.0041801100596785545, + 0.014796906150877476, + 0.005006345454603434, + 0.0054041072726249695, + 0.011040924116969109, + 0.005020469892770052, + 0.005832342896610498, + 0.016936222091317177, + 0.00932997278869152, + 0.02306787297129631, + 0.008563295938074589, + 0.00830242969095707, + 0.019832903519272804, + 0.00938921794295311, + 0.012753605842590332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.16041186451911926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019521719077602029 + ], + "opposite_logprob_tokens": [ + -6.251952171325684 + ], + "logprob_reward": -0.0019521719077602029, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019521719077602029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19493432343006134, + 0.14629296958446503, + 0.07941614091396332, + 0.08661624044179916, + 0.07373721152544022, + 0.16561612486839294, + 0.04230416193604469, + 0.06661517918109894, + 0.07411164045333862, + 0.07035602629184723 + ], + "credit_map": [ + 0.03344867751002312, + 0.01570865511894226, + 0.028624193742871284, + 0.022979585453867912, + 0.02195177972316742, + 0.029647190123796463, + 0.02428116649389267, + 0.018293090164661407, + 0.020247843116521835, + 0.011973239481449127, + 0.03415224701166153, + 0.015730641782283783, + 0.012239164672791958, + 0.025647753849625587, + 0.012010252103209496, + 0.014291832223534584, + 0.008693859912455082, + 0.007430728059262037, + 0.0073065790347754955, + 0.014298127964138985, + 0.009023504331707954, + 0.009617152623832226, + 0.016553666442632675, + 0.006492515094578266, + 0.009291939437389374, + 0.011078821495175362, + 0.010160569101572037, + 0.011238185688853264, + 0.012034947052598, + 0.012032137252390385, + 0.008857925422489643, + 0.011921722441911697, + 0.00970772746950388, + 0.008870746940374374, + 0.01029499527066946, + 0.008015118539333344, + 0.009725423529744148, + 0.010423008352518082, + 0.007353088352829218, + 0.009347102604806423, + 0.01906762458384037, + 0.019676178693771362, + 0.023299813270568848, + 0.017511704936623573, + 0.021535471081733704, + 0.025231938809156418, + 0.023304369300603867, + 0.015989026054739952, + 0.004268741700798273, + 0.003686940763145685, + 0.008638528175652027, + 0.004173234570771456, + 0.0051001301035285, + 0.006454077549278736, + 0.005290864035487175, + 0.004691645968705416, + 0.005477267783135176, + 0.005960528738796711, + 0.012411953881382942, + 0.009153710678219795, + 0.0050012641586363316, + 0.009738323278725147, + 0.010065528564155102, + 0.008806603029370308, + 0.008038341999053955, + 0.007933820597827435, + 0.007722829002887011, + 0.010164440609514713, + 0.010245400480926037, + 0.010573514737188816, + 0.012494385242462158, + 0.006938903592526913, + 0.005611092783510685, + 0.008039364591240883, + 0.009666298516094685, + 0.010145825333893299, + 0.008501618169248104, + 0.010039014741778374, + 0.01060901302844286, + 0.00774379400536418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0019521719077602029 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542063370347023 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542063370347023, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542063370347023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1810898333787918, + 0.18155720829963684, + 0.08557901531457901, + 0.11171642690896988, + 0.09354138374328613, + 0.07148533314466476, + 0.037577126175165176, + 0.0964377224445343, + 0.05285745486617088, + 0.08815855532884598 + ], + "credit_map": [ + 0.02919618971645832, + 0.01376599632203579, + 0.03389172628521919, + 0.017434168606996536, + 0.021102488040924072, + 0.028623782098293304, + 0.015759317204356194, + 0.02131616324186325, + 0.02353338897228241, + 0.015109270811080933, + 0.038789052516222, + 0.017353927716612816, + 0.023316510021686554, + 0.029524503275752068, + 0.019869273528456688, + 0.014061286114156246, + 0.01269163191318512, + 0.005117792636156082, + 0.015211135149002075, + 0.00850654672831297, + 0.008424445055425167, + 0.012767234817147255, + 0.014475448988378048, + 0.008384786546230316, + 0.011856984347105026, + 0.017054976895451546, + 0.014258380979299545, + 0.010113592259585857, + 0.018186133354902267, + 0.015407491475343704, + 0.009502085857093334, + 0.015336773358285427, + 0.01193819660693407, + 0.008872087113559246, + 0.011053654365241528, + 0.00842228438705206, + 0.01350446417927742, + 0.013172976672649384, + 0.01541880052536726, + 0.011158914305269718, + 0.010330830700695515, + 0.00872772466391325, + 0.009515192359685898, + 0.005210739094763994, + 0.011510917916893959, + 0.009871724992990494, + 0.008684242144227028, + 0.007633960340172052, + 0.003610965795814991, + 0.002800531452521682, + 0.007063557859510183, + 0.004008183721452951, + 0.004288234282284975, + 0.005931889172643423, + 0.006158350966870785, + 0.003715411527082324, + 0.007026512175798416, + 0.0101175457239151, + 0.017937155440449715, + 0.013744269497692585, + 0.011544314213097095, + 0.015316138975322247, + 0.011002824641764164, + 0.009748962707817554, + 0.006804763805121183, + 0.007243054918944836, + 0.006774847395718098, + 0.006236772984266281, + 0.006358713377267122, + 0.00731642497703433, + 0.00741176400333643, + 0.0047111124731600285, + 0.006333270575851202, + 0.009489881806075573, + 0.011873014271259308, + 0.013104282319545746, + 0.010041186586022377, + 0.012690738774836063, + 0.0160240288823843, + 0.00860215537250042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.012542063370347023 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static wide-angle shot of a scientific cross-section of a massive stratovolcano?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background stark and featureless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast magma chamber visible at the base containing churning bright orange and yellow-white molten energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the volcano's summit visibly deformed into a strained dome with a network of fine cracks spreading across the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with a surreal, high-contrast digital appearance in impeccable, piercing focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046449145302176476 + ], + "opposite_logprob_tokens": [ + -5.379644870758057 + ], + "logprob_reward": -0.0046449145302176476, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046449145302176476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16068124771118164, + 0.11987893283367157, + 0.04344336315989494, + 0.108932726085186, + 0.11996862292289734, + 0.1181354746222496, + 0.12929315865039825, + 0.08508169651031494, + 0.06631889194250107, + 0.04826592653989792 + ], + "credit_map": [ + 0.023243630304932594, + 0.01318688690662384, + 0.015056688338518143, + 0.018636006861925125, + 0.02180967479944229, + 0.021007781848311424, + 0.02985510043799877, + 0.017885467037558556, + 0.013537505641579628, + 0.009496593847870827, + 0.018591413274407387, + 0.010458854958415031, + 0.013440161012113094, + 0.017895454540848732, + 0.0208269115537405, + 0.01563202776014805, + 0.006363733671605587, + 0.0038597865495830774, + 0.008286027237772942, + 0.006875932682305574, + 0.003269841894507408, + 0.0070857880637049675, + 0.004155698698014021, + 0.003546550404280424, + 0.013999362476170063, + 0.008226734586060047, + 0.00855847354978323, + 0.009233410470187664, + 0.014401118271052837, + 0.01578627899289131, + 0.02244218811392784, + 0.016285160556435585, + 0.01942472718656063, + 0.018616223707795143, + 0.01054065115749836, + 0.009446147829294205, + 0.019744668155908585, + 0.017421506345272064, + 0.013200635090470314, + 0.01157405786216259, + 0.015850065276026726, + 0.01257316768169403, + 0.010672440752387047, + 0.013858417049050331, + 0.0162692628800869, + 0.016475962474942207, + 0.020326664671301842, + 0.012109477072954178, + 0.009512688964605331, + 0.008283916860818863, + 0.018223406746983528, + 0.019141322001814842, + 0.01552222203463316, + 0.01801959052681923, + 0.02640780806541443, + 0.0141821950674057, + 0.008395806886255741, + 0.009331912733614445, + 0.01359486673027277, + 0.010346612893044949, + 0.009757242165505886, + 0.011271081864833832, + 0.012242138385772705, + 0.010142029263079166, + 0.008540560491383076, + 0.0072072041220963, + 0.011507335118949413, + 0.006726729683578014, + 0.00802435353398323, + 0.011348347179591656, + 0.0062170131132006645, + 0.006747343111783266, + 0.004474436864256859, + 0.005252435337752104, + 0.006889629177749157, + 0.006296134553849697, + 0.005242374259978533, + 0.007114655338227749, + 0.008663474582135677, + 0.004332783631980419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0046449145302176476 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_15/index_tensor([3330, 3330])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_15/index_tensor([3330, 3330])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2960beee710389ce5152a9d8a22580127965cafc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_15/index_tensor([3330, 3330])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extremely wide shot, perfectly still, captures a breathtakingly grand scene from a distant vantage point atop a gentle grassy knoll. In the foreground, a figure in a dark olive green field jacket and utility trousers stands in focused concentration. Steadily and deliberately, they raise a professional black SLR camera equipped with a powerful telephoto lens, bringing the viewfinder to their right eye. In the center, a magnificent neoclassical mansion dominates the landscape, its white marble facade and towering columns glowing warmly in the golden morning light of the setting sun. The mansion's multi-tiered windows gleam like gold, reflecting the day's final rays, and are surrounded by manicured emerald lawns and intricate hedge gardens. The photographer's body is poised, his left hand holding the long lens steady as his right index finger gently presses the shutter button—a small yet decisive movement that captures the majestic and serene atmosphere of this vast estate beneath a clear, soft sky. The long, soft shadows cast across the rolling hills and the warm, saturated color palette create a sense of timeless grandeur and quiet reverence.", + "seed": 7890, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_15/index_3330_seed_7890_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.22600656785070897, + "vlm_logprob": -0.22600656785070897, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.22600656785070897, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15217356383800507 + ], + "opposite_logprob_tokens": [ + -2.1521735191345215 + ], + "logprob_reward": -0.15217356383800507, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.15217356383800507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.338453471660614, + 0.03696506842970848, + 0.01843836158514023, + 0.03687126561999321, + 0.017599526792764664, + 0.019343797117471695, + 0.020687950775027275, + 0.09995785355567932, + 0.3043884336948395, + 0.10729429870843887 + ], + "credit_map": [ + 0.03133777156472206, + 0.04718717560172081, + 0.022458981722593307, + 0.061966490000486374, + 0.053572069853544235, + 0.04035673290491104, + 0.05748504400253296, + 0.02408919483423233, + 0.003700984176248312, + 0.005015629809349775, + 0.0026026659179478884, + 0.005108326207846403, + 0.006769725587219, + 0.00602692412212491, + 0.0042777094058692455, + 0.0034631050657480955, + 0.0019088039407506585, + 0.0024361489340662956, + 0.0015495424158871174, + 0.0028813753742724657, + 0.0029411264695227146, + 0.0029392919968813658, + 0.0019291837234050035, + 0.0018528882646933198, + 0.004419669974595308, + 0.005010490771383047, + 0.0028298464603722095, + 0.006108541507273912, + 0.004802198149263859, + 0.005081792362034321, + 0.006053255870938301, + 0.0025654728524386883, + 0.0020544459111988544, + 0.002432481152936816, + 0.0017139731207862496, + 0.003289717249572277, + 0.002629440277814865, + 0.002962245838716626, + 0.0012060687877237797, + 0.0013111549196764827, + 0.00146481825504452, + 0.002392582129687071, + 0.0012538748560473323, + 0.002639204263687134, + 0.00540302786976099, + 0.003134168451651931, + 0.0018053798703476787, + 0.001250740489922464, + 0.0030096531845629215, + 0.0031141163781285286, + 0.0015572027768939734, + 0.002166355261579156, + 0.004327245056629181, + 0.00269124424085021, + 0.002508000936359167, + 0.0013141336385160685, + 0.013223769143223763, + 0.019684607163071632, + 0.009086507372558117, + 0.012740061618387699, + 0.015959041193127632, + 0.010607542470097542, + 0.010515274479985237, + 0.008141052909195423, + 0.036251250654459, + 0.06169283390045166, + 0.02581881172955036, + 0.03274083137512207, + 0.05708463862538338, + 0.038375046104192734, + 0.03343724459409714, + 0.01898777484893799, + 0.013126895762979984, + 0.023737119510769844, + 0.008239339105784893, + 0.01033762190490961, + 0.02233116887509823, + 0.013315673917531967, + 0.008283264935016632, + 0.007923212833702564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.15217356383800507 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983846306800842 + ], + "opposite_logprob_tokens": [ + -0.5233846306800842 + ], + "logprob_reward": -0.8983846306800842, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983846306800842, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04420852288603783, + 0.01566370762884617, + 0.011381606571376324, + 0.01218269020318985, + 0.009769421070814133, + 0.01618882268667221, + 0.021433241665363312, + 0.4324658215045929, + 0.18904288113117218, + 0.24766325950622559 + ], + "credit_map": [ + 0.002196573419496417, + 0.004823573864996433, + 0.0044423844665288925, + 0.009428009390830994, + 0.008315211161971092, + 0.006906360853463411, + 0.00523453438654542, + 0.0028618748765438795, + 0.0014926299918442965, + 0.0020437645725905895, + 0.0017937900265678763, + 0.003053057938814163, + 0.0017903934931382537, + 0.0019231654005125165, + 0.0018374172504991293, + 0.0017294897697865963, + 0.0011079463874921203, + 0.0018018336268141866, + 0.001315456349402666, + 0.0020513187628239393, + 0.0014215130358934402, + 0.0014759341720491648, + 0.0011164980242028832, + 0.0010911065619438887, + 0.0009641245123930275, + 0.0014362435322254896, + 0.0018819591496139765, + 0.002868332201614976, + 0.0011950680054724216, + 0.0012271457817405462, + 0.0011042205151170492, + 0.001505595981143415, + 0.0010619439417496324, + 0.0013466516975313425, + 0.0010634019272401929, + 0.001495044562034309, + 0.0016399560263380408, + 0.001470879651606083, + 0.0009171828860417008, + 0.0007743601454421878, + 0.001208106055855751, + 0.0014661455061286688, + 0.002675988944247365, + 0.004613695200532675, + 0.0010615804931148887, + 0.0014634104445576668, + 0.0020977617241442204, + 0.0016021346673369408, + 0.00158341892529279, + 0.0021983084734529257, + 0.0031300147529691458, + 0.005089020822197199, + 0.0025145208928734064, + 0.0018383520655333996, + 0.0027813834603875875, + 0.002298219595104456, + 0.051594674587249756, + 0.08750960230827332, + 0.03437555953860283, + 0.042767349630594254, + 0.08676662296056747, + 0.0545947439968586, + 0.04747077822685242, + 0.02738649956882, + 0.02231377176940441, + 0.03770097717642784, + 0.015153707005083561, + 0.019847529008984566, + 0.037801653146743774, + 0.025427352637052536, + 0.019400255754590034, + 0.01139764953404665, + 0.0304510910063982, + 0.05655406042933464, + 0.017341090366244316, + 0.023087825626134872, + 0.05207440257072449, + 0.03170429542660713, + 0.019767988473176956, + 0.016682490706443787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.8983846306800842 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637290023267269 + ], + "opposite_logprob_tokens": [ + -3.6513729095458984 + ], + "logprob_reward": -0.02637290023267269, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637290023267269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30810871720314026, + 0.09045364707708359, + 0.07243240624666214, + 0.02520953118801117, + 0.09736710786819458, + 0.037984006106853485, + 0.015890857204794884, + 0.13371828198432922, + 0.1206577867269516, + 0.09817764908075333 + ], + "credit_map": [ + 0.02435755543410778, + 0.03478662297129631, + 0.03506196662783623, + 0.04762091487646103, + 0.04695183038711548, + 0.04802195727825165, + 0.037781354039907455, + 0.03352654352784157, + 0.0066695366986095905, + 0.008347077295184135, + 0.007199374493211508, + 0.021832603961229324, + 0.013639714568853378, + 0.012423520907759666, + 0.013471686281263828, + 0.00687014264985919, + 0.007699430454522371, + 0.01045615691691637, + 0.006635184399783611, + 0.014638983644545078, + 0.011426222510635853, + 0.007534901611506939, + 0.00799874309450388, + 0.00604279525578022, + 0.003072593826800585, + 0.0033756354823708534, + 0.0022393730469048023, + 0.0037524241488426924, + 0.004684848245233297, + 0.0034371144138276577, + 0.003111619269475341, + 0.0015359246172010899, + 0.010733583010733128, + 0.010269717313349247, + 0.01153660286217928, + 0.021052997559309006, + 0.011179203167557716, + 0.013855598866939545, + 0.010016195476055145, + 0.008723210543394089, + 0.0034850924275815487, + 0.0037504693027585745, + 0.004148658365011215, + 0.010599296540021896, + 0.003611838212236762, + 0.004133367910981178, + 0.004988092929124832, + 0.003267193678766489, + 0.0014865936245769262, + 0.0015732282772660255, + 0.001486981869675219, + 0.0036423527635633945, + 0.002292540157213807, + 0.0018433575751259923, + 0.0022855079732835293, + 0.0012802969431504607, + 0.015731962397694588, + 0.026014871895313263, + 0.008322094567120075, + 0.016614144667983055, + 0.02471327781677246, + 0.01734360307455063, + 0.015991058200597763, + 0.008987268432974815, + 0.014543771743774414, + 0.024352172389626503, + 0.011125077493488789, + 0.013253271579742432, + 0.022689953446388245, + 0.01431095227599144, + 0.012593057937920094, + 0.007789540570229292, + 0.011857647448778152, + 0.022634070366621017, + 0.006220138631761074, + 0.010892299003899097, + 0.020721009001135826, + 0.012305616401135921, + 0.007433079183101654, + 0.006113796029239893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.02637290023267269 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029802024364471436 + ], + "opposite_logprob_tokens": [ + -3.529802083969116 + ], + "logprob_reward": -0.029802024364471436, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029802024364471436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41100791096687317, + 0.08549953252077103, + 0.07592514902353287, + 0.015547524206340313, + 0.04180086776614189, + 0.029508622363209724, + 0.01767002046108246, + 0.09908957779407501, + 0.05619749054312706, + 0.16775326430797577 + ], + "credit_map": [ + 0.031901925802230835, + 0.06324388086795807, + 0.04267406836152077, + 0.04093630611896515, + 0.07489877939224243, + 0.06697875261306763, + 0.05514395982027054, + 0.035230252891778946, + 0.006725971587002277, + 0.010454830713570118, + 0.009039098396897316, + 0.009259458631277084, + 0.016123617067933083, + 0.016669228672981262, + 0.01032550074160099, + 0.006901824846863747, + 0.005507929716259241, + 0.01195642352104187, + 0.005080231465399265, + 0.0063063884153962135, + 0.01962381973862648, + 0.015103633515536785, + 0.008650355041027069, + 0.0036963673774152994, + 0.0015480323927477002, + 0.002274641301482916, + 0.00168359256349504, + 0.002254038816317916, + 0.002368454821407795, + 0.0019895348232239485, + 0.0020990618504583836, + 0.0013301674043759704, + 0.003642079886049032, + 0.004603230860084295, + 0.003360430710017681, + 0.00441108038648963, + 0.009339090436697006, + 0.008502405136823654, + 0.005551944021135569, + 0.002390606328845024, + 0.0029560260009020567, + 0.0036412121262401342, + 0.002783498726785183, + 0.0035159222315996885, + 0.006124725099653006, + 0.004818189423531294, + 0.003833158640190959, + 0.001835889765061438, + 0.0018474350217729807, + 0.002428995445370674, + 0.0013982069212943316, + 0.0020019561052322388, + 0.003990213852375746, + 0.0028971051797270775, + 0.001971842022612691, + 0.0011342670768499374, + 0.012758470140397549, + 0.019999807700514793, + 0.006525792181491852, + 0.010734887793660164, + 0.017916450276970863, + 0.013100073672831059, + 0.01061095017939806, + 0.0074431500397622585, + 0.006517874542623758, + 0.010934626683592796, + 0.005794655065983534, + 0.006829280871897936, + 0.009955080226063728, + 0.006342199165374041, + 0.006347633898258209, + 0.0034761414863169193, + 0.019921211525797844, + 0.039399951696395874, + 0.01132740918546915, + 0.016795067116618156, + 0.03577785566449165, + 0.021056117489933968, + 0.01281485054641962, + 0.010660809464752674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.029802024364471436 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023299720138311386 + ], + "opposite_logprob_tokens": [ + -3.7732996940612793 + ], + "logprob_reward": -0.023299720138311386, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023299720138311386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1669480949640274, + 0.04229061305522919, + 0.02713157795369625, + 0.021120864897966385, + 0.02192816138267517, + 0.01766245998442173, + 0.021097656339406967, + 0.07879593968391418, + 0.4116339087486267, + 0.1913907378911972 + ], + "credit_map": [ + 0.015844587236642838, + 0.018485374748706818, + 0.023378225043416023, + 0.02170788124203682, + 0.022419581189751625, + 0.024638479575514793, + 0.017693299800157547, + 0.022780654951930046, + 0.0037378226406872272, + 0.004324550274759531, + 0.0055017005652189255, + 0.008095452561974525, + 0.004550380632281303, + 0.005817414727061987, + 0.004322828724980354, + 0.0059404619969427586, + 0.0019879296887665987, + 0.003802873659878969, + 0.0024508386850357056, + 0.003730094525963068, + 0.005820847116410732, + 0.00486783916130662, + 0.0022011040709912777, + 0.0022700524423271418, + 0.0022555349860340357, + 0.0018949155928567052, + 0.0029352137353271246, + 0.004043211229145527, + 0.003179738763719797, + 0.003420160850510001, + 0.0016990891890600324, + 0.001693001133389771, + 0.0015313095645979047, + 0.002633408410474658, + 0.0014553177170455456, + 0.0022939457558095455, + 0.005908516235649586, + 0.004683936946094036, + 0.0023474630434066057, + 0.0010742624290287495, + 0.0020180747378617525, + 0.002550137462094426, + 0.0019023423083126545, + 0.0031675102654844522, + 0.0033582455944269896, + 0.0019016065634787083, + 0.0014853845350444317, + 0.0012791596818715334, + 0.0015413040528073907, + 0.001988509204238653, + 0.0020766991656273603, + 0.0039329552091658115, + 0.004282753448933363, + 0.002935549942776561, + 0.002688869833946228, + 0.001651014550589025, + 0.007970286533236504, + 0.011713799089193344, + 0.007119348272681236, + 0.015396708622574806, + 0.010496624745428562, + 0.007792445831000805, + 0.009740378707647324, + 0.008566344156861305, + 0.04473543539643288, + 0.07944326847791672, + 0.03599291667342186, + 0.0465213879942894, + 0.08127355575561523, + 0.054093606770038605, + 0.04322509095072746, + 0.02634865790605545, + 0.024386698380112648, + 0.0443563349545002, + 0.012058372609317303, + 0.021385300904512405, + 0.039126500487327576, + 0.023783931508660316, + 0.014827802777290344, + 0.011465799994766712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.023299720138311386 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely wide shot, perfectly still, captures a breathtakingly grand scene from a distant vantage point atop a gentle grassy knoll. In the foreground, a figure in a dark olive green field jacket and utility trousers stands in focused concentration. Steadily and deliberately, they raise a professional black SLR camera equipped with a powerful telephoto lens, bringing the viewfinder to their right eye. In the center, a magnificent neoclassical mansion dominates the landscape, its white marble facade and towering columns glowing warmly in the golden morning light of the setting sun. The mansion's multi-tiered windows gleam like gold, reflecting the day's final rays, and are surrounded by manicured emerald lawns and intricate hedge gardens. The photographer's body is poised, his left hand holding the long lens steady as his right index finger gently presses the shutter button—a small yet decisive movement that captures the majestic and serene atmosphere of this vast estate beneath a clear, soft sky. The long, soft shadows cast across the rolling hills and the warm, saturated color palette create a sense of timeless grandeur and quiet reverence.", + "seed": 7891, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_15/index_3330_seed_7891_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.20755752441473305, + "vlm_logprob": -0.20755752441473305, + "vlm_yes_no_margin": 2.799999988079071, + "vlm_token_credit": -0.20755752441473305, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.36830997467041016 + ], + "opposite_logprob_tokens": [ + -1.2433099746704102 + ], + "logprob_reward": -0.36830997467041016, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.36830997467041016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45992156863212585, + 0.16674086451530457, + 0.049739934504032135, + 0.0499567985534668, + 0.02563757449388504, + 0.020977718755602837, + 0.060579705983400345, + 0.03350730240345001, + 0.10804911702871323, + 0.024889379739761353 + ], + "credit_map": [ + 0.06585843116044998, + 0.061366673558950424, + 0.06487947702407837, + 0.036194562911987305, + 0.08475742489099503, + 0.06292606890201569, + 0.050522446632385254, + 0.0334165096282959, + 0.030080508440732956, + 0.03318484500050545, + 0.024740176275372505, + 0.015273015014827251, + 0.01762908510863781, + 0.023603452369570732, + 0.00944958534091711, + 0.012780197896063328, + 0.007092263549566269, + 0.009804981760680676, + 0.006674485746771097, + 0.005038103554397821, + 0.006225443910807371, + 0.006768540479242802, + 0.0032477935310453176, + 0.0048883226700127125, + 0.007330327294766903, + 0.009355069138109684, + 0.005077526438981295, + 0.004712575115263462, + 0.0060185836628079414, + 0.006802446208894253, + 0.004637852311134338, + 0.006022420711815357, + 0.00339516787789762, + 0.003123518778011203, + 0.00428178533911705, + 0.0021690099965780973, + 0.0031951775308698416, + 0.004335474688559771, + 0.0023570596240460873, + 0.0027803813572973013, + 0.002741813426837325, + 0.0032054355833679438, + 0.002669129054993391, + 0.0018969194497913122, + 0.0031624340917915106, + 0.0032660348806530237, + 0.0017527771415188909, + 0.002283174777403474, + 0.008232547901570797, + 0.011007601395249367, + 0.00849139504134655, + 0.005461954977363348, + 0.008959152735769749, + 0.007771031465381384, + 0.005610841326415539, + 0.005045183468610048, + 0.003958779387176037, + 0.00576041080057621, + 0.00414592819288373, + 0.0027472455985844135, + 0.00579639058560133, + 0.004410644061863422, + 0.0034232100006192923, + 0.0032646909821778536, + 0.018992014229297638, + 0.024014106020331383, + 0.012082988396286964, + 0.00872462335973978, + 0.015217216685414314, + 0.01599409431219101, + 0.006050232797861099, + 0.006973847281187773, + 0.004540327005088329, + 0.005583421792834997, + 0.003165579866617918, + 0.0018433828372508287, + 0.003067433601245284, + 0.004011872224509716, + 0.001259363954886794, + 0.0014179982244968414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.36830997467041016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487110733985901 + ], + "opposite_logprob_tokens": [ + -1.2237110137939453 + ], + "logprob_reward": -0.3487110733985901, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3487110733985901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3824717104434967, + 0.07291723042726517, + 0.07596734911203384, + 0.046564981341362, + 0.052096862345933914, + 0.042195189744234085, + 0.04666569456458092, + 0.14086154103279114, + 0.10273624956607819, + 0.03752313181757927 + ], + "credit_map": [ + 0.048751648515462875, + 0.05378018692135811, + 0.044965218752622604, + 0.03496203199028969, + 0.07530149817466736, + 0.048431213945150375, + 0.04124431312084198, + 0.0350356251001358, + 0.011672534048557281, + 0.010034315288066864, + 0.010276410728693008, + 0.006523962132632732, + 0.010393591597676277, + 0.010328290984034538, + 0.006694594398140907, + 0.0069935377687215805, + 0.012974724173545837, + 0.014559541828930378, + 0.008097132667899132, + 0.008139618672430515, + 0.010557342320680618, + 0.010559001937508583, + 0.005277504213154316, + 0.005802495405077934, + 0.006565535441040993, + 0.008051585406064987, + 0.005516284145414829, + 0.003836915595456958, + 0.007968273013830185, + 0.006321579217910767, + 0.004294501151889563, + 0.004010312259197235, + 0.007840086705982685, + 0.010678627528250217, + 0.006351975724101067, + 0.00518285296857357, + 0.006123502738773823, + 0.007264825515449047, + 0.0038627164904028177, + 0.004792277701199055, + 0.005630106199532747, + 0.004347400739789009, + 0.006116011179983616, + 0.0040285843424499035, + 0.006289883516728878, + 0.005189318675547838, + 0.004825494717806578, + 0.0057683964259922504, + 0.0064268046990036964, + 0.009147653356194496, + 0.006154800299555063, + 0.004155789501965046, + 0.005556831136345863, + 0.006903964560478926, + 0.003752125659957528, + 0.0045677307061851025, + 0.011847808957099915, + 0.016606628894805908, + 0.011497925035655499, + 0.011738905683159828, + 0.035748060792684555, + 0.016943303868174553, + 0.021334489807486534, + 0.01514443103224039, + 0.014064641669392586, + 0.013932544738054276, + 0.0103457635268569, + 0.00967163871973753, + 0.016680944710969925, + 0.0125955268740654, + 0.013423382304608822, + 0.012021820992231369, + 0.005519013851881027, + 0.0060835937038064, + 0.004251983016729355, + 0.003925351891666651, + 0.005745879374444485, + 0.005223223939538002, + 0.0034932552371174097, + 0.0032808342948555946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.3487110733985901 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134717345237732 + ], + "opposite_logprob_tokens": [ + -1.313471794128418 + ], + "logprob_reward": -0.3134717345237732, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134717345237732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26781001687049866, + 0.3106549084186554, + 0.10005860030651093, + 0.049957700073719025, + 0.09463411569595337, + 0.033468980342149734, + 0.02165190689265728, + 0.018357565626502037, + 0.013875499367713928, + 0.08953066915273666 + ], + "credit_map": [ + 0.042004551738500595, + 0.04721333459019661, + 0.03562761843204498, + 0.019814971834421158, + 0.04109639301896095, + 0.040507521480321884, + 0.02132655680179596, + 0.020219065248966217, + 0.054033197462558746, + 0.07065831124782562, + 0.03390854224562645, + 0.024929676204919815, + 0.03675389289855957, + 0.04924233257770538, + 0.016497552394866943, + 0.024631403386592865, + 0.017201928421854973, + 0.024199742823839188, + 0.011126036755740643, + 0.007902546785771847, + 0.011649781838059425, + 0.015384344384074211, + 0.0055434503592550755, + 0.0070507763884961605, + 0.008530463092029095, + 0.01169387437403202, + 0.006280736532062292, + 0.003339264076203108, + 0.005533748771995306, + 0.008579069748520851, + 0.002747396705672145, + 0.003253149101510644, + 0.016732240095734596, + 0.024343304336071014, + 0.01024051383137703, + 0.006872603669762611, + 0.010880251415073872, + 0.016241993755102158, + 0.0044196899980306625, + 0.004903527908027172, + 0.005161012057214975, + 0.00798137579113245, + 0.004428142681717873, + 0.002420096192508936, + 0.00368582084774971, + 0.005590234883129597, + 0.0020592526998370886, + 0.0021430477499961853, + 0.002642920706421137, + 0.003410837147384882, + 0.0023520668037235737, + 0.001833515358157456, + 0.004102243576198816, + 0.0029139211401343346, + 0.002298858715221286, + 0.002097545890137553, + 0.0027201042976230383, + 0.0037068522069603205, + 0.0021702053491026163, + 0.0016067222459241748, + 0.0033981024753302336, + 0.0023633542004972696, + 0.001317909685894847, + 0.001074316562153399, + 0.001501038670539856, + 0.0013329378562048078, + 0.001400271779857576, + 0.0015133656561374664, + 0.0027843446005135775, + 0.0019946631509810686, + 0.001622504903934896, + 0.0017263740301132202, + 0.015197820961475372, + 0.022099260240793228, + 0.00888850074261427, + 0.006442571058869362, + 0.011962435208261013, + 0.01465423684567213, + 0.004885033238679171, + 0.005400815512984991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.3134717345237732 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003196133067831397 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.003196133067831397, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003196133067831397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39251261949539185, + 0.16713231801986694, + 0.07174108922481537, + 0.029469504952430725, + 0.0495762899518013, + 0.03394484147429466, + 0.03687921166419983, + 0.06194135919213295, + 0.10210026800632477, + 0.0547025091946125 + ], + "credit_map": [ + 0.062256015837192535, + 0.06370019167661667, + 0.05347071960568428, + 0.02594427391886711, + 0.06266210228204727, + 0.06392966210842133, + 0.027080930769443512, + 0.033468712121248245, + 0.02443947270512581, + 0.029066307470202446, + 0.021806461736559868, + 0.01151720155030489, + 0.024094637483358383, + 0.02491806447505951, + 0.018235182389616966, + 0.013054993003606796, + 0.008910750970244408, + 0.008698313497006893, + 0.010184426791965961, + 0.005641819443553686, + 0.013375378213822842, + 0.010359263978898525, + 0.010212032124400139, + 0.004359101410955191, + 0.003986863885074854, + 0.006733080837875605, + 0.002723456360399723, + 0.002698230789974332, + 0.004581262823194265, + 0.004526924341917038, + 0.002091094618663192, + 0.0021285905968397856, + 0.006888398434966803, + 0.005960698239505291, + 0.009138948284089565, + 0.003778469981625676, + 0.00768614187836647, + 0.008071749471127987, + 0.004235171712934971, + 0.0038167103193700314, + 0.00394195131957531, + 0.00300635420717299, + 0.005035943351686001, + 0.0029631718061864376, + 0.006035766564309597, + 0.004231174476444721, + 0.005722030531615019, + 0.003008450847119093, + 0.003765786299481988, + 0.004173211753368378, + 0.003971600905060768, + 0.0037175852339714766, + 0.006277165841311216, + 0.006018359679728746, + 0.003798535792157054, + 0.005156968254595995, + 0.005256730131804943, + 0.008128839544951916, + 0.005711269099265337, + 0.005600039381533861, + 0.016771702095866203, + 0.008105849847197533, + 0.008528361096978188, + 0.003838572883978486, + 0.015986016020178795, + 0.020335406064987183, + 0.011075953021645546, + 0.008094574324786663, + 0.016204342246055603, + 0.014806605875492096, + 0.007607075851410627, + 0.00799030065536499, + 0.008084796369075775, + 0.01197725161910057, + 0.005518669728189707, + 0.00393596151843667, + 0.008430222049355507, + 0.00834517739713192, + 0.004008009564131498, + 0.004402423743158579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.003196133067831397 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098706413060427 + ], + "opposite_logprob_tokens": [ + -5.504098892211914 + ], + "logprob_reward": -0.004098706413060427, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098706413060427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2651265561580658, + 0.15840303897857666, + 0.07573253661394119, + 0.07789789885282516, + 0.08631228655576706, + 0.0791187733411789, + 0.08167657256126404, + 0.04994070529937744, + 0.07280153036117554, + 0.05299007520079613 + ], + "credit_map": [ + 0.04336318373680115, + 0.042563267052173615, + 0.028352033346891403, + 0.015105519443750381, + 0.03951046243309975, + 0.04139796644449234, + 0.029489755630493164, + 0.02534436620771885, + 0.018165167421102524, + 0.022821635007858276, + 0.01812826283276081, + 0.012315860949456692, + 0.0282097440212965, + 0.01936444826424122, + 0.021007124334573746, + 0.01839078776538372, + 0.009290444664657116, + 0.009335861541330814, + 0.008842704817652702, + 0.007336197886615992, + 0.011678148061037064, + 0.008351530879735947, + 0.011810054071247578, + 0.009087602607905865, + 0.009755361825227737, + 0.01531138177961111, + 0.0075857629999518394, + 0.006872354540973902, + 0.013250758871436119, + 0.010291830636560917, + 0.007689211051911116, + 0.007141239941120148, + 0.0069137681275606155, + 0.00928127858787775, + 0.009467339143157005, + 0.009737025015056133, + 0.015410424210131168, + 0.010169673711061478, + 0.01341849472373724, + 0.011914287693798542, + 0.009474902413785458, + 0.011374068446457386, + 0.009293665178120136, + 0.006366928108036518, + 0.014737630262970924, + 0.010928181000053883, + 0.008195033296942711, + 0.008748362772166729, + 0.013195134699344635, + 0.01769578829407692, + 0.008470846340060234, + 0.005953268613666296, + 0.010370834730565548, + 0.012705427594482899, + 0.006631351541727781, + 0.006653918884694576, + 0.00513847591355443, + 0.007530042435973883, + 0.004173958208411932, + 0.004648230038583279, + 0.009064310230314732, + 0.006757584400475025, + 0.007612834684550762, + 0.005015271715819836, + 0.010665803216397762, + 0.014529832638800144, + 0.008832587860524654, + 0.007137990090996027, + 0.011663193814456463, + 0.010208859108388424, + 0.00511512765660882, + 0.004648132715374231, + 0.00567838316783309, + 0.007847905158996582, + 0.00523767713457346, + 0.005098157562315464, + 0.00958274956792593, + 0.006408641114830971, + 0.007067494094371796, + 0.006069068796932697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.004098706413060427 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely wide shot, perfectly still, captures a breathtakingly grand scene from a distant vantage point atop a gentle grassy knoll. In the foreground, a figure in a dark olive green field jacket and utility trousers stands in focused concentration. Steadily and deliberately, they raise a professional black SLR camera equipped with a powerful telephoto lens, bringing the viewfinder to their right eye. In the center, a magnificent neoclassical mansion dominates the landscape, its white marble facade and towering columns glowing warmly in the golden morning light of the setting sun. The mansion's multi-tiered windows gleam like gold, reflecting the day's final rays, and are surrounded by manicured emerald lawns and intricate hedge gardens. The photographer's body is poised, his left hand holding the long lens steady as his right index finger gently presses the shutter button—a small yet decisive movement that captures the majestic and serene atmosphere of this vast estate beneath a clear, soft sky. The long, soft shadows cast across the rolling hills and the warm, saturated color palette create a sense of timeless grandeur and quiet reverence.", + "seed": 7892, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_15/index_3330_seed_7892_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1187952577834949, + "vlm_logprob": -0.1187952577834949, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.1187952577834949, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0879543349146843 + ], + "opposite_logprob_tokens": [ + -2.71295428276062 + ], + "logprob_reward": -0.0879543349146843, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0879543349146843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15527047216892242, + 0.4348433017730713, + 0.19275051355361938, + 0.056099388748407364, + 0.02986188791692257, + 0.03352166339755058, + 0.021176502108573914, + 0.022859811782836914, + 0.022760367020964622, + 0.030856087803840637 + ], + "credit_map": [ + 0.0160861574113369, + 0.0344085767865181, + 0.01582416705787182, + 0.020410042256116867, + 0.02287125587463379, + 0.01662573404610157, + 0.015037729404866695, + 0.014006799086928368, + 0.04212433844804764, + 0.08784251660108566, + 0.040867503732442856, + 0.04093121737241745, + 0.0715922862291336, + 0.032898664474487305, + 0.06495730578899384, + 0.05362943932414055, + 0.034000273793935776, + 0.026988638564944267, + 0.01991909369826317, + 0.01531805470585823, + 0.029796771705150604, + 0.01854313164949417, + 0.02105959877371788, + 0.027124954387545586, + 0.0056632873602211475, + 0.010823382064700127, + 0.0053860959596931934, + 0.00550219789147377, + 0.008349490351974964, + 0.005805746652185917, + 0.0077269421890378, + 0.006842251401394606, + 0.003408577758818865, + 0.005039283540099859, + 0.0034409938380122185, + 0.0033810955937951803, + 0.0045694890432059765, + 0.003940057475119829, + 0.0031373193487524986, + 0.002945073414593935, + 0.0033204841893166304, + 0.00658292043954134, + 0.0036552848760038614, + 0.003958556801080704, + 0.0053071496076881886, + 0.0032846685498952866, + 0.0039360057562589645, + 0.003476593876257539, + 0.002631667535752058, + 0.0021346842404454947, + 0.0032131194602698088, + 0.0032054975163191557, + 0.002275989856570959, + 0.0024358786176890135, + 0.0035509399604052305, + 0.0017287248047068715, + 0.0029587855096906424, + 0.002923537977039814, + 0.002589830895885825, + 0.002869801130145788, + 0.0025840045418590307, + 0.002430310472846031, + 0.0036730077117681503, + 0.002830532845109701, + 0.002713691908866167, + 0.0024428905453532934, + 0.003084036521613598, + 0.0040837242268025875, + 0.002556420862674713, + 0.0029373702127486467, + 0.002798529574647546, + 0.0021437040995806456, + 0.003696887753903866, + 0.005995830520987511, + 0.003933408763259649, + 0.003138454630970955, + 0.004209061153233051, + 0.002548155840486288, + 0.0048218038864433765, + 0.0025124859530478716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0879543349146843 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31351348757743835 + ], + "opposite_logprob_tokens": [ + -1.3135135173797607 + ], + "logprob_reward": -0.31351348757743835, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31351348757743835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36624765396118164, + 0.2104707658290863, + 0.15370284020900726, + 0.022963110357522964, + 0.09761568158864975, + 0.02895701862871647, + 0.022190628573298454, + 0.05695342645049095, + 0.03167314454913139, + 0.009225700981914997 + ], + "credit_map": [ + 0.03478270396590233, + 0.07419601082801819, + 0.03383191302418709, + 0.036386191844940186, + 0.0614173449575901, + 0.027105873450636864, + 0.05567894130945206, + 0.042848702520132065, + 0.020186709240078926, + 0.04351894557476044, + 0.021185167133808136, + 0.02069488726556301, + 0.034865789115428925, + 0.017035311087965965, + 0.02886362001299858, + 0.024120356887578964, + 0.014075708575546741, + 0.03517298027873039, + 0.014466147869825363, + 0.015783434733748436, + 0.024665173143148422, + 0.012299773283302784, + 0.0201779305934906, + 0.017061689868569374, + 0.0028140575159341097, + 0.0047355834394693375, + 0.002585224574431777, + 0.0031472884584218264, + 0.0030009497422724962, + 0.0024912997614592314, + 0.0018264459213241935, + 0.002362262224778533, + 0.009730269201099873, + 0.020883332937955856, + 0.009740401059389114, + 0.009980903938412666, + 0.016111036762595177, + 0.008842570707201958, + 0.012220362201333046, + 0.010106815956532955, + 0.002686647931113839, + 0.005798785947263241, + 0.002605714835226536, + 0.003560019424185157, + 0.004891630262136459, + 0.0025351073127239943, + 0.003776107681915164, + 0.003103007562458515, + 0.002617845544591546, + 0.004120000172406435, + 0.0025537493638694286, + 0.0034926969092339277, + 0.0026900293305516243, + 0.0023798267357051373, + 0.0019127465784549713, + 0.0024237376637756824, + 0.005301500670611858, + 0.010880794376134872, + 0.005469076335430145, + 0.005875241011381149, + 0.010801051743328571, + 0.0046001216396689415, + 0.00888913869857788, + 0.005136511288583279, + 0.003700061235576868, + 0.00561140663921833, + 0.0036818855442106724, + 0.00320760952308774, + 0.004631771240383387, + 0.003733552759513259, + 0.0032791951671242714, + 0.003827662207186222, + 0.0010280549759045243, + 0.0015865142922848463, + 0.0010292917722836137, + 0.001163296983577311, + 0.0012305636191740632, + 0.0008981286082416773, + 0.0012803676072508097, + 0.0010094840545207262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.31351348757743835 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799178421497345 + ], + "opposite_logprob_tokens": [ + -1.804917812347412 + ], + "logprob_reward": -0.1799178421497345, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799178421497345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18102861940860748, + 0.0636635571718216, + 0.5028172731399536, + 0.05903402343392372, + 0.049019813537597656, + 0.05004461482167244, + 0.04895328730344772, + 0.008949373848736286, + 0.013782497495412827, + 0.022706935182213783 + ], + "credit_map": [ + 0.01877603866159916, + 0.026517486199736595, + 0.02052505686879158, + 0.024480244144797325, + 0.025370704010128975, + 0.014907436445355415, + 0.028616908937692642, + 0.021834732964634895, + 0.006196110509335995, + 0.013424091972410679, + 0.006256650201976299, + 0.006433017551898956, + 0.009902246296405792, + 0.005053211469203234, + 0.00876010861247778, + 0.007638117298483849, + 0.04770645126700401, + 0.11734805256128311, + 0.047471094876527786, + 0.052614726126194, + 0.07913931459188461, + 0.042191002517938614, + 0.06226763501763344, + 0.05407900735735893, + 0.00820169597864151, + 0.009959261864423752, + 0.0076371836476027966, + 0.009062960743904114, + 0.004717821720987558, + 0.00726340664550662, + 0.00562372338026762, + 0.0065679727122187614, + 0.005943373776972294, + 0.00616485858336091, + 0.0065847402438521385, + 0.010466140694916248, + 0.00362825533375144, + 0.004649237729609013, + 0.005168841686099768, + 0.006414366420358419, + 0.004231484606862068, + 0.009595190174877644, + 0.004802870564162731, + 0.007290961220860481, + 0.007811222691088915, + 0.0036810480523854494, + 0.007066330872476101, + 0.005565505940467119, + 0.006285942159593105, + 0.0077389986254274845, + 0.006423776037991047, + 0.009947097860276699, + 0.0042786854319274426, + 0.00503535894677043, + 0.0039054860826581717, + 0.005337940528988838, + 0.00113580038305372, + 0.0018093392718583345, + 0.0010084818350151181, + 0.0012533681001514196, + 0.0011080813128501177, + 0.0010191532783210278, + 0.0007148300646804273, + 0.0009003193699754775, + 0.0018498179269954562, + 0.002030886011198163, + 0.001842600991949439, + 0.002160967793315649, + 0.0013329277280718088, + 0.0014955287333577871, + 0.00133104482665658, + 0.0017387229017913342, + 0.002605005633085966, + 0.003176738740876317, + 0.002695351606234908, + 0.003113911021500826, + 0.003192297415807843, + 0.0021700954530388117, + 0.003615487599745393, + 0.0021380477119237185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.1799178421497345 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015132891712710261 + ], + "opposite_logprob_tokens": [ + -6.501513481140137 + ], + "logprob_reward": -0.0015132891712710261, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015132891712710261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19657406210899353, + 0.17874202132225037, + 0.32772117853164673, + 0.05057959258556366, + 0.1301148384809494, + 0.017165984958410263, + 0.017106104642152786, + 0.032121457159519196, + 0.022499151527881622, + 0.027375532314181328 + ], + "credit_map": [ + 0.018446650356054306, + 0.03627534210681915, + 0.01875503547489643, + 0.0190776027739048, + 0.03036750666797161, + 0.014167875982820988, + 0.03167794272303581, + 0.027806097641587257, + 0.017807666212320328, + 0.03300238400697708, + 0.018379757180809975, + 0.01670168712735176, + 0.027354784309864044, + 0.013427807949483395, + 0.02788619138300419, + 0.024181751534342766, + 0.03069869615137577, + 0.07522772252559662, + 0.031326137483119965, + 0.035595010966062546, + 0.051662627607584, + 0.028065554797649384, + 0.037932027131319046, + 0.03721342608332634, + 0.005603879224509001, + 0.010965021327137947, + 0.004603316076099873, + 0.005072842352092266, + 0.008695916272699833, + 0.003906693775206804, + 0.0058986428193748, + 0.005833286326378584, + 0.012641526758670807, + 0.02404548041522503, + 0.012724817730486393, + 0.012030843645334244, + 0.021355392411351204, + 0.010795717127621174, + 0.020218240097165108, + 0.01630282960832119, + 0.0020191280636936426, + 0.0021553633268922567, + 0.0019944829400628805, + 0.0020498696248978376, + 0.0017825135728344321, + 0.002102469326928258, + 0.0023574421647936106, + 0.002704715123400092, + 0.0020376055035740137, + 0.0019298515981063247, + 0.0021514305844902992, + 0.0022665869910269976, + 0.0016482025384902954, + 0.002335361437872052, + 0.0019214109051972628, + 0.00281565566547215, + 0.0029899664223194122, + 0.006770452484488487, + 0.0032942728139460087, + 0.0034353488590568304, + 0.0059439511969685555, + 0.002942143939435482, + 0.003937311004847288, + 0.002808012766763568, + 0.002900980645790696, + 0.0030118473805487156, + 0.0028847840148955584, + 0.003306484315544367, + 0.0022331690415740013, + 0.002292908960953355, + 0.0029520101379603148, + 0.0029169695917516947, + 0.0034301290288567543, + 0.005051034968346357, + 0.0032481742091476917, + 0.003050809493288398, + 0.0036220429465174675, + 0.003234428120777011, + 0.00290398090146482, + 0.002834932878613472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0015132891712710261 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077335104346275 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077335104346275, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077335104346275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29897409677505493, + 0.2324492484331131, + 0.06765519082546234, + 0.030972322449088097, + 0.19838985800743103, + 0.04131404682993889, + 0.030806273221969604, + 0.060745496302843094, + 0.018621273338794708, + 0.02007213979959488 + ], + "credit_map": [ + 0.031679943203926086, + 0.0416254848241806, + 0.03230960667133331, + 0.03281356021761894, + 0.04006484895944595, + 0.026806900277733803, + 0.05557128041982651, + 0.03810247406363487, + 0.023243604227900505, + 0.044616058468818665, + 0.02264726348221302, + 0.022264137864112854, + 0.03753839060664177, + 0.016887955367565155, + 0.036278482526540756, + 0.028973359614610672, + 0.008878280408680439, + 0.011333662085235119, + 0.006343081593513489, + 0.00788203813135624, + 0.014488586224615574, + 0.005481208208948374, + 0.007691610604524612, + 0.005556723568588495, + 0.003423232352361083, + 0.005213426426053047, + 0.0027740218210965395, + 0.003592470893636346, + 0.004880843684077263, + 0.0026432236190885305, + 0.004550311714410782, + 0.003894790541380644, + 0.02115250751376152, + 0.031891435384750366, + 0.01976984739303589, + 0.018826084211468697, + 0.029783962294459343, + 0.017982929944992065, + 0.03195697069168091, + 0.027026120573282242, + 0.004866159521043301, + 0.0062334551475942135, + 0.004135332070291042, + 0.003937970846891403, + 0.006205217447131872, + 0.0036581524182111025, + 0.006375899072736502, + 0.005901859141886234, + 0.003958535380661488, + 0.0050066495314240456, + 0.003989789634943008, + 0.003927441313862801, + 0.0028235919307917356, + 0.0039881146512925625, + 0.002967295702546835, + 0.004144852515310049, + 0.005823021288961172, + 0.01175749208778143, + 0.005927611142396927, + 0.005993496626615524, + 0.01095615979284048, + 0.004736804869025946, + 0.009128948673605919, + 0.00642196461558342, + 0.0020624243188649416, + 0.0036041438579559326, + 0.0019596912898123264, + 0.0018548545194789767, + 0.00316431256942451, + 0.0018292346503585577, + 0.0019724899902939796, + 0.002174121793359518, + 0.0025702957063913345, + 0.002959580859169364, + 0.0022953033912926912, + 0.002319198800250888, + 0.00287735927850008, + 0.002333336975425482, + 0.002449692226946354, + 0.002267372328788042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.011077335104346275 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely wide shot, perfectly still, captures a breathtakingly grand scene from a distant vantage point atop a gentle grassy knoll. In the foreground, a figure in a dark olive green field jacket and utility trousers stands in focused concentration. Steadily and deliberately, they raise a professional black SLR camera equipped with a powerful telephoto lens, bringing the viewfinder to their right eye. In the center, a magnificent neoclassical mansion dominates the landscape, its white marble facade and towering columns glowing warmly in the golden morning light of the setting sun. The mansion's multi-tiered windows gleam like gold, reflecting the day's final rays, and are surrounded by manicured emerald lawns and intricate hedge gardens. The photographer's body is poised, his left hand holding the long lens steady as his right index finger gently presses the shutter button—a small yet decisive movement that captures the majestic and serene atmosphere of this vast estate beneath a clear, soft sky. The long, soft shadows cast across the rolling hills and the warm, saturated color palette create a sense of timeless grandeur and quiet reverence.", + "seed": 7893, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_15/index_3330_seed_7893_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16567018581554294, + "vlm_logprob": -0.16567018581554294, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.16567018581554294, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22641178965568542 + ], + "opposite_logprob_tokens": [ + -1.7264118194580078 + ], + "logprob_reward": -0.22641178965568542, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.22641178965568542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39007699489593506, + 0.13128425180912018, + 0.0838569700717926, + 0.04124927520751953, + 0.07750365138053894, + 0.03039928525686264, + 0.10370460897684097, + 0.05716974288225174, + 0.025734908878803253, + 0.0590202696621418 + ], + "credit_map": [ + 0.045191433280706406, + 0.026724912226200104, + 0.05760680511593819, + 0.04640211537480354, + 0.06113731488585472, + 0.06439898163080215, + 0.04968327283859253, + 0.03893217071890831, + 0.015922971069812775, + 0.010595102794468403, + 0.02686634473502636, + 0.019546614959836006, + 0.014126216061413288, + 0.01787607930600643, + 0.012005780823528767, + 0.014345148578286171, + 0.012707949616014957, + 0.008200875483453274, + 0.015551503747701645, + 0.011666066013276577, + 0.009259483776986599, + 0.011695701628923416, + 0.006502974778413773, + 0.008272419683635235, + 0.007604618091136217, + 0.0031254158820956945, + 0.006532942410558462, + 0.005693142302334309, + 0.004852048587054014, + 0.006526162847876549, + 0.0030917932745069265, + 0.003823155304417014, + 0.011272392235696316, + 0.007635790854692459, + 0.012396348640322685, + 0.011233841069042683, + 0.009246339090168476, + 0.00890395138412714, + 0.009149674326181412, + 0.00766531378030777, + 0.005992611404508352, + 0.003211900359019637, + 0.0021294932812452316, + 0.0034859704319387674, + 0.0037765565793961287, + 0.0068665253929793835, + 0.0024974667467176914, + 0.0024387624580413103, + 0.014717977494001389, + 0.007248989772051573, + 0.010741706937551498, + 0.013934427872300148, + 0.012963504530489445, + 0.018034618347883224, + 0.014104501344263554, + 0.011958889663219452, + 0.008191999979317188, + 0.0062683383002877235, + 0.005720039363950491, + 0.006943251471966505, + 0.007337164133787155, + 0.01249887328594923, + 0.005708508659154177, + 0.004501573741436005, + 0.004119647201150656, + 0.0025182506069540977, + 0.0030156439170241356, + 0.0034289725590497255, + 0.0035400833003222942, + 0.0039750621654093266, + 0.003124401904642582, + 0.002012847922742367, + 0.009328855201601982, + 0.006759889889508486, + 0.007188247051090002, + 0.006712904665619135, + 0.007040846161544323, + 0.01379903219640255, + 0.0035923896357417107, + 0.004598104394972324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.22641178965568542 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.474375456571579 + ], + "opposite_logprob_tokens": [ + -0.9743754863739014 + ], + "logprob_reward": -0.474375456571579, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.474375456571579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24467310309410095, + 0.1177234873175621, + 0.0672934353351593, + 0.052503280341625214, + 0.07636580616235733, + 0.029255148023366928, + 0.06520043313503265, + 0.03561917692422867, + 0.09718547016382217, + 0.21418064832687378 + ], + "credit_map": [ + 0.03441430628299713, + 0.03201377019286156, + 0.016740480437874794, + 0.03996432200074196, + 0.02728988043963909, + 0.04405505582690239, + 0.02969796024262905, + 0.02049732580780983, + 0.015609033405780792, + 0.012596947140991688, + 0.009736355394124985, + 0.014802847057580948, + 0.0140149537473917, + 0.029350552707910538, + 0.010843917727470398, + 0.010768888518214226, + 0.008810420520603657, + 0.006333358585834503, + 0.010415068827569485, + 0.0108555369079113, + 0.006313262041658163, + 0.012696433812379837, + 0.005860111676156521, + 0.006009237375110388, + 0.007419590372592211, + 0.004381612874567509, + 0.010636428371071815, + 0.00836265180259943, + 0.004472396336495876, + 0.008036662824451923, + 0.004275892395526171, + 0.004918043967336416, + 0.011520551517605782, + 0.008682912215590477, + 0.007769736927002668, + 0.0069877817295491695, + 0.007936730980873108, + 0.022110963240265846, + 0.004840651527047157, + 0.006516479887068272, + 0.003870041575282812, + 0.0023593204095959663, + 0.0022556118201464415, + 0.005150743760168552, + 0.003444134956225753, + 0.003969001118093729, + 0.003588308347389102, + 0.004617984406650066, + 0.008856897242367268, + 0.006624349392950535, + 0.00797698087990284, + 0.009930387139320374, + 0.007699053734540939, + 0.013461164198815823, + 0.006287582218647003, + 0.00436401879414916, + 0.0063645923510193825, + 0.003486428176984191, + 0.002522828057408333, + 0.005247810855507851, + 0.003397034713998437, + 0.006806644145399332, + 0.004649871960282326, + 0.0031439654994755983, + 0.012514900416135788, + 0.010807867161929607, + 0.010885007679462433, + 0.010347855277359486, + 0.010071204043924809, + 0.021649029105901718, + 0.009977311827242374, + 0.010932298377156258, + 0.03681211173534393, + 0.021680820733308792, + 0.022366484627127647, + 0.025705618783831596, + 0.0227457657456398, + 0.057207755744457245, + 0.013345440849661827, + 0.014316647313535213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.474375456571579 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297174543142319 + ], + "opposite_logprob_tokens": [ + -2.237971782684326 + ], + "logprob_reward": -0.11297174543142319, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297174543142319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45151618123054504, + 0.09198030829429626, + 0.03445444628596306, + 0.046812962740659714, + 0.08269426971673965, + 0.053930360823869705, + 0.06922271102666855, + 0.030203428119421005, + 0.05900569632649422, + 0.0801795944571495 + ], + "credit_map": [ + 0.05076931044459343, + 0.03952302038669586, + 0.037163302302360535, + 0.0650353729724884, + 0.06677448004484177, + 0.06318794935941696, + 0.0794868990778923, + 0.04957584664225578, + 0.012220199219882488, + 0.008048657327890396, + 0.009388278238475323, + 0.014731135219335556, + 0.01245252974331379, + 0.008976048789918423, + 0.015487831085920334, + 0.01067562960088253, + 0.004738988820463419, + 0.005017288960516453, + 0.0033000719267874956, + 0.005601003300398588, + 0.0035273474641144276, + 0.006026984192430973, + 0.0030343851540237665, + 0.0032083746045827866, + 0.0071629928424954414, + 0.004396470263600349, + 0.007988442666828632, + 0.0070428745821118355, + 0.004598311614245176, + 0.005598061718046665, + 0.005198006518185139, + 0.004827804863452911, + 0.011807974427938461, + 0.00846464280039072, + 0.012889728881418705, + 0.012790423817932606, + 0.008478691801428795, + 0.00881581287831068, + 0.01120750606060028, + 0.008239486254751682, + 0.007668924052268267, + 0.004273831844329834, + 0.009088050574064255, + 0.008532208390533924, + 0.0060762097127735615, + 0.004901174921542406, + 0.007571275345981121, + 0.005818684119731188, + 0.010502051562070847, + 0.006682765670120716, + 0.012126780115067959, + 0.008361222222447395, + 0.006290951278060675, + 0.012148202396929264, + 0.006196130067110062, + 0.0069146063178777695, + 0.0037356431130319834, + 0.002796152839437127, + 0.004420531447976828, + 0.0054051694460213184, + 0.003645265242084861, + 0.003968500066548586, + 0.0034771570935845375, + 0.002755008405074477, + 0.006637575104832649, + 0.004500305280089378, + 0.005702978931367397, + 0.0081197340041399, + 0.0071468898095190525, + 0.015382222831249237, + 0.005418690852820873, + 0.006097298115491867, + 0.012487778440117836, + 0.007214241661131382, + 0.011591004207730293, + 0.013241080567240715, + 0.008388002403080463, + 0.013495816849172115, + 0.007204976864159107, + 0.006556689739227295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.11297174543142319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948933307081461 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005948933307081461, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948933307081461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24967525899410248, + 0.14598962664604187, + 0.09115961194038391, + 0.04021812975406647, + 0.09685476124286652, + 0.07796622067689896, + 0.037475764751434326, + 0.02310439944267273, + 0.04246603325009346, + 0.1950901746749878 + ], + "credit_map": [ + 0.03766607865691185, + 0.030269021168351173, + 0.03562874719500542, + 0.03449704498052597, + 0.023365182802081108, + 0.03655829653143883, + 0.02367687225341797, + 0.028014013543725014, + 0.021222583949565887, + 0.0196334607899189, + 0.011905895546078682, + 0.02350809797644615, + 0.014645716175436974, + 0.022698253393173218, + 0.01725301332771778, + 0.015122596174478531, + 0.012771357782185078, + 0.010538887232542038, + 0.01040984969586134, + 0.014768954366445541, + 0.010275201871991158, + 0.012257068417966366, + 0.010706772096455097, + 0.009431520476937294, + 0.006407839711755514, + 0.005871592555195093, + 0.0025608674623072147, + 0.005325138568878174, + 0.0045980545692145824, + 0.009058207273483276, + 0.0034435505513101816, + 0.002952876966446638, + 0.013300551101565361, + 0.013143416494131088, + 0.01539680641144514, + 0.010134751908481121, + 0.009487075731158257, + 0.01865496300160885, + 0.007577594369649887, + 0.009159602224826813, + 0.011664184741675854, + 0.008023490197956562, + 0.012683786451816559, + 0.00982846599072218, + 0.007483290042728186, + 0.012082083150744438, + 0.008743965066969395, + 0.007456954102963209, + 0.005159090738743544, + 0.004445353057235479, + 0.0061698188073933125, + 0.004053894430398941, + 0.003844822756946087, + 0.006296789273619652, + 0.003964120987802744, + 0.0035418751649558544, + 0.0033648302778601646, + 0.002049405360594392, + 0.002091709291562438, + 0.003349454142153263, + 0.0028557346668094397, + 0.003543853061273694, + 0.0032488692086189985, + 0.0026005429681390524, + 0.005259629338979721, + 0.005427486263215542, + 0.004307101480662823, + 0.005670721177011728, + 0.0045210071839392185, + 0.00889801885932684, + 0.0051428996957838535, + 0.0032391666900366545, + 0.026332460343837738, + 0.019446099177002907, + 0.024260252714157104, + 0.04162866994738579, + 0.015394105575978756, + 0.029094241559505463, + 0.019833819940686226, + 0.019100524485111237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.005948933307081461 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extremely wide shot taken from a distant vantage point atop a grassy knoll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perfectly still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a figure visible wearing a dark olive green field jacket and utility trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a magnificent neoclassical mansion with a white marble facade and towering columns visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the mansion illuminated by warm golden light with multi-tiered windows gleaming like gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643004111945629 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643004111945629, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643004111945629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2387741208076477, + 0.07101615518331528, + 0.06789964437484741, + 0.03728658705949783, + 0.06666560471057892, + 0.029084058478474617, + 0.03625623881816864, + 0.021071573719382286, + 0.13595066964626312, + 0.2959953844547272 + ], + "credit_map": [ + 0.032001860439777374, + 0.0240319836884737, + 0.020559245720505714, + 0.031991321593523026, + 0.026306932792067528, + 0.056105826050043106, + 0.023887433111667633, + 0.023889515548944473, + 0.011796437203884125, + 0.008174063637852669, + 0.005642533767968416, + 0.013491539284586906, + 0.006768498104065657, + 0.01204688474535942, + 0.007382058072835207, + 0.005714135710150003, + 0.009511268697679043, + 0.007281627506017685, + 0.00482948636636138, + 0.0119412150233984, + 0.007042103447020054, + 0.01642295904457569, + 0.005935871507972479, + 0.00493510439991951, + 0.006124974228441715, + 0.005335016641765833, + 0.002640185412019491, + 0.00489344447851181, + 0.004552309401333332, + 0.006332761142402887, + 0.004249140620231628, + 0.003158751642331481, + 0.00884449016302824, + 0.0063810632564127445, + 0.0075250547379255295, + 0.009794740006327629, + 0.00725529994815588, + 0.01542737614363432, + 0.006012653466314077, + 0.005424922797828913, + 0.004055886995047331, + 0.0033502678852528334, + 0.004245089832693338, + 0.004028507508337498, + 0.0028911971021443605, + 0.0059690289199352264, + 0.0022321308497339487, + 0.0023119489196687937, + 0.005931873340159655, + 0.0038390900008380413, + 0.006694837007671595, + 0.004247703589498997, + 0.0032125331927090883, + 0.006073672790080309, + 0.002797340741381049, + 0.003459187224507332, + 0.0033325732219964266, + 0.0021805560681968927, + 0.0019461936317384243, + 0.002920945640653372, + 0.0020148230250924826, + 0.0033786222338676453, + 0.0030430417973548174, + 0.0022548146080225706, + 0.022235743701457977, + 0.012767271138727665, + 0.013729201629757881, + 0.020161516964435577, + 0.012644151225686073, + 0.035747021436691284, + 0.009737793356180191, + 0.008927968330681324, + 0.041889335960149765, + 0.05125850439071655, + 0.019986556842923164, + 0.058970410376787186, + 0.03146415203809738, + 0.036356594413518906, + 0.03182690963149071, + 0.024242904037237167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.008643004111945629 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_2/index_tensor([2370, 2370])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_2/index_tensor([2370, 2370])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b2f1aeec3dd205a279fa9ccc905748019431c7be --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_2/index_tensor([2370, 2370])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tense, extreme close-up shot, captured in the style of competitive cinematography, follows the subject, a man with a resolute face and a sharp jawline, wearing a dark gray, moisture-wicking athletic T-shirt. His expression of concentration and a slight frown, as if he were in the thick of a decisive tournament. Under the cool, sharp sidelight of a desk lamp, his steady, powerful hands swiftly manipulate a minimalist black blood glucose meter. A slight motion blur captures the precise moment as he swiftly and decisively inserts a test strip into the side slot. Then, the camera's sharp focus shifts to the illuminated digital display, while his face and surroundings appear slightly out of focus. The numbers on the screen are jumping and changing rapidly and silently. Every flash is as gripping as a countdown, casting a faint blue light on his focused face and the skin on his knuckles that have turned white due to exertion. The whole picture is filled with a high degree of tension and concentration, sublimating a daily health monitoring behavior into a silent duel with oneself and racing against time.", + "seed": 24408, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_2/index_2370_seed_24408_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021961135789752007, + "vlm_logprob": -0.021961135789752007, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.021961135789752007, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012018056586384773 + ], + "opposite_logprob_tokens": [ + -5.012018203735352 + ], + "logprob_reward": -0.012018056586384773, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012018056586384773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.330135703086853, + 0.10354606062173843, + 0.06255912035703659, + 0.05355849489569664, + 0.1328345239162445, + 0.09539701789617538, + 0.059081386774778366, + 0.0644812285900116, + 0.045051656663417816, + 0.05335480347275734 + ], + "credit_map": [ + 0.03301987797021866, + 0.026327937841415405, + 0.054422736167907715, + 0.06474856287240982, + 0.03529418259859085, + 0.03991643339395523, + 0.04043172672390938, + 0.03597426414489746, + 0.010172608308494091, + 0.010893799364566803, + 0.012976942583918571, + 0.012133831158280373, + 0.015981201082468033, + 0.017257437109947205, + 0.013081653974950314, + 0.011048588901758194, + 0.008130495436489582, + 0.006177778821438551, + 0.009897965006530285, + 0.007495087571442127, + 0.007105612196028233, + 0.00742725282907486, + 0.010171068832278252, + 0.006153861992061138, + 0.007871279492974281, + 0.00540077593177557, + 0.0069837807677686214, + 0.005713461432605982, + 0.0075742220506072044, + 0.005863513331860304, + 0.009090308099985123, + 0.005061150994151831, + 0.011371869593858719, + 0.023195402696728706, + 0.016480345278978348, + 0.031659796833992004, + 0.01673608459532261, + 0.007387102581560612, + 0.014673183672130108, + 0.01133074052631855, + 0.015231418423354626, + 0.009112728759646416, + 0.01583433523774147, + 0.007824759930372238, + 0.010572127997875214, + 0.0136148976162076, + 0.013596533797681332, + 0.009610224515199661, + 0.004633490927517414, + 0.006161392200738192, + 0.008135013282299042, + 0.01243141945451498, + 0.00920393131673336, + 0.006936429068446159, + 0.007771121338009834, + 0.0038085877895355225, + 0.004179633222520351, + 0.0031989142298698425, + 0.01566203497350216, + 0.007746095769107342, + 0.006616887636482716, + 0.00680293096229434, + 0.015990279614925385, + 0.0042844489216804504, + 0.0038364941719919443, + 0.003120242850854993, + 0.005483339074999094, + 0.006436674855649471, + 0.005605960264801979, + 0.00395617401227355, + 0.010241677984595299, + 0.0063710943795740604, + 0.004085034132003784, + 0.00609263451769948, + 0.0067216260358691216, + 0.007427260745316744, + 0.007736591622233391, + 0.006234561558812857, + 0.009877309203147888, + 0.005179785192012787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.012018056586384773 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982828952372074 + ], + "opposite_logprob_tokens": [ + -3.5298283100128174 + ], + "logprob_reward": -0.02982828952372074, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982828952372074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1897399127483368, + 0.06464516371488571, + 0.05748376622796059, + 0.08438713103532791, + 0.07408161461353302, + 0.06550924479961395, + 0.044883597642183304, + 0.056800372898578644, + 0.03126269951462746, + 0.33120644092559814 + ], + "credit_map": [ + 0.020155301317572594, + 0.01687709242105484, + 0.02732858620584011, + 0.037264734506607056, + 0.023593638092279434, + 0.017156055197119713, + 0.022828472778201103, + 0.02453603595495224, + 0.007493008393794298, + 0.008849995210766792, + 0.008370314724743366, + 0.006996811833232641, + 0.010874244384467602, + 0.008228601887822151, + 0.006595221348106861, + 0.007236965931952, + 0.007575932424515486, + 0.005622999742627144, + 0.0069648995995521545, + 0.0071593658067286015, + 0.006312802899628878, + 0.004864159040153027, + 0.011994196102023125, + 0.006989407353103161, + 0.01529689785093069, + 0.00891120731830597, + 0.00878401380032301, + 0.009058676660060883, + 0.01489652693271637, + 0.009382798336446285, + 0.011539780534803867, + 0.006517226807773113, + 0.006717825774103403, + 0.00752996513620019, + 0.011699819006025791, + 0.013022849336266518, + 0.012961072847247124, + 0.007843575440347195, + 0.00862326193600893, + 0.005683249793946743, + 0.006560531444847584, + 0.006649625021964312, + 0.008373865857720375, + 0.006977467332035303, + 0.011936073191463947, + 0.00860021635890007, + 0.008674859069287777, + 0.007736607920378447, + 0.003624913515523076, + 0.003671071259304881, + 0.00465625524520874, + 0.006372875068336725, + 0.007753504440188408, + 0.004782745614647865, + 0.010175371542572975, + 0.003846863517537713, + 0.003244817955419421, + 0.0032354481518268585, + 0.0144069017842412, + 0.005675637628883123, + 0.004717530217021704, + 0.0056494553573429585, + 0.016084257513284683, + 0.003786320798099041, + 0.003190248040482402, + 0.0029342977795749903, + 0.0032196827232837677, + 0.003718145890161395, + 0.004239705856889486, + 0.0034919362515211105, + 0.0062524485401809216, + 0.004216233734041452, + 0.025683343410491943, + 0.024001726880669594, + 0.05027323216199875, + 0.0438968762755394, + 0.046987276524305344, + 0.025531740859150887, + 0.08034185320138931, + 0.03449040651321411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.02982828952372074 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744596175849438 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.006744596175849438, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744596175849438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1322324126958847, + 0.044688791036605835, + 0.2625499963760376, + 0.13175520300865173, + 0.17300671339035034, + 0.07612377405166626, + 0.07856801897287369, + 0.04072978347539902, + 0.023876840248703957, + 0.03646845743060112 + ], + "credit_map": [ + 0.014964032918214798, + 0.011312378570437431, + 0.020221246406435966, + 0.025753462687134743, + 0.011746353469789028, + 0.01731419749557972, + 0.016115443781018257, + 0.014805302023887634, + 0.004857389256358147, + 0.005181761458516121, + 0.005166918504983187, + 0.006567021366208792, + 0.00499000633135438, + 0.006616074126213789, + 0.005896592512726784, + 0.005413034465163946, + 0.02864476665854454, + 0.02026284672319889, + 0.030308235436677933, + 0.0464872308075428, + 0.021533627063035965, + 0.02731119468808174, + 0.05623911693692207, + 0.0317629911005497, + 0.02076922170817852, + 0.011964276432991028, + 0.01828000880777836, + 0.012683556415140629, + 0.018493851646780968, + 0.012228475883603096, + 0.025501592084765434, + 0.011834243312478065, + 0.013926944695413113, + 0.02507185749709606, + 0.021446499973535538, + 0.04191715642809868, + 0.021700985729694366, + 0.011152906343340874, + 0.02072448469698429, + 0.01706589013338089, + 0.010200082324445248, + 0.008919020183384418, + 0.010313304141163826, + 0.006641769781708717, + 0.01137065514922142, + 0.00770175876095891, + 0.010949922725558281, + 0.010027275420725346, + 0.005879613570868969, + 0.004057778976857662, + 0.010889246128499508, + 0.009362575598061085, + 0.011121407151222229, + 0.009084039367735386, + 0.019806431606411934, + 0.008366935886442661, + 0.003018379909917712, + 0.003341947216540575, + 0.008224398829042912, + 0.005270257126539946, + 0.004249621648341417, + 0.004545030649751425, + 0.008398396894335747, + 0.0036817574873566628, + 0.0016857912996783853, + 0.0018681290093809366, + 0.0038509401492774487, + 0.003767680376768112, + 0.0023249106016010046, + 0.0022241640836000443, + 0.005048195365816355, + 0.0031070318073034286, + 0.0028248357120901346, + 0.003309039631858468, + 0.004366969224065542, + 0.004939179867506027, + 0.005535825155675411, + 0.005451418925076723, + 0.006941747386008501, + 0.003099443158134818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.006744596175849438 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048677146434783936 + ], + "opposite_logprob_tokens": [ + -3.0486772060394287 + ], + "logprob_reward": -0.048677146434783936, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048677146434783936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09720605611801147, + 0.06889722496271133, + 0.07537850737571716, + 0.11385621130466461, + 0.1394020915031433, + 0.06174192950129509, + 0.1495668888092041, + 0.06458549201488495, + 0.11464272439479828, + 0.11472291499376297 + ], + "credit_map": [ + 0.013430085964500904, + 0.00942823477089405, + 0.01403043232858181, + 0.013041459955275059, + 0.01022875402122736, + 0.011317072436213493, + 0.013241112232208252, + 0.012488891370594501, + 0.007167826406657696, + 0.006916357669979334, + 0.008341335691511631, + 0.012654850259423256, + 0.00811975821852684, + 0.007714608684182167, + 0.008464274927973747, + 0.009518202394247055, + 0.009738213382661343, + 0.005642547272145748, + 0.009825760498642921, + 0.012006823904812336, + 0.005261223763227463, + 0.004920983221381903, + 0.01859743706882, + 0.009385515004396439, + 0.019758315756917, + 0.015126931481063366, + 0.012363743968307972, + 0.012539834715425968, + 0.01905185915529728, + 0.010272163897752762, + 0.014666043221950531, + 0.010077313520014286, + 0.013114795088768005, + 0.021758198738098145, + 0.023440934717655182, + 0.027676021680235863, + 0.016380149871110916, + 0.009451614692807198, + 0.015104045160114765, + 0.012476311065256596, + 0.006559533532708883, + 0.00844648014754057, + 0.0052014603279531, + 0.0057070450857281685, + 0.013200449757277966, + 0.005891588982194662, + 0.007635393179953098, + 0.009099974296987057, + 0.01364466268569231, + 0.01041555404663086, + 0.02675566077232361, + 0.023014146834611893, + 0.023016545921564102, + 0.012365074828267097, + 0.023604633286595345, + 0.01675059273838997, + 0.004189955070614815, + 0.00386450975202024, + 0.013367896899580956, + 0.005903749726712704, + 0.008684143424034119, + 0.007432463113218546, + 0.015452684834599495, + 0.005690090358257294, + 0.007046593353152275, + 0.006611523684114218, + 0.018245115876197815, + 0.023764679208397865, + 0.012787648476660252, + 0.0061364867724478245, + 0.03051474504172802, + 0.009535931050777435, + 0.007757712621241808, + 0.01127276849001646, + 0.016419045627117157, + 0.016090210527181625, + 0.017320862039923668, + 0.013421837240457535, + 0.023704160004854202, + 0.008736306801438332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.048677146434783936 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537590228021145 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537590228021145, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537590228021145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057034432888031006, + 0.07465315610170364, + 0.09784944355487823, + 0.34665077924728394, + 0.09547059237957001, + 0.157409206032753, + 0.07854577153921127, + 0.02509969286620617, + 0.034109216183423996, + 0.033177681267261505 + ], + "credit_map": [ + 0.011817020364105701, + 0.005208609160035849, + 0.009327064268290997, + 0.007404276169836521, + 0.005612120032310486, + 0.0043074581772089005, + 0.00625374261289835, + 0.007104149088263512, + 0.00870504416525364, + 0.007480854168534279, + 0.010277939960360527, + 0.011748387478291988, + 0.008009377866983414, + 0.007274889387190342, + 0.012471289373934269, + 0.008685378357768059, + 0.011235496960580349, + 0.006956346333026886, + 0.010476792231202126, + 0.017146944999694824, + 0.008544402197003365, + 0.005947167985141277, + 0.02598983608186245, + 0.011552461422979832, + 0.061901018023490906, + 0.035153526812791824, + 0.04552128165960312, + 0.04695463180541992, + 0.0608157254755497, + 0.029452122747898102, + 0.04156261309981346, + 0.025289878249168396, + 0.009526006877422333, + 0.011979128234088421, + 0.014353922568261623, + 0.016498960554599762, + 0.013689626939594746, + 0.008774335496127605, + 0.0136911291629076, + 0.006957479752600193, + 0.015296027064323425, + 0.019690899178385735, + 0.01896192692220211, + 0.018132060766220093, + 0.025480350479483604, + 0.013615631498396397, + 0.021739132702350616, + 0.02449316345155239, + 0.006640248466283083, + 0.006864922121167183, + 0.009562071412801743, + 0.01466357335448265, + 0.011850657872855663, + 0.008611098863184452, + 0.015108928084373474, + 0.0052442667074501514, + 0.002161884680390358, + 0.0029994023498147726, + 0.0034656859934329987, + 0.002849290147423744, + 0.0027456474490463734, + 0.004136905074119568, + 0.004339751321822405, + 0.002401128178462386, + 0.0024419038090854883, + 0.0025273654609918594, + 0.005878453608602285, + 0.005457000806927681, + 0.005111932288855314, + 0.002762231742963195, + 0.006961929611861706, + 0.0029683976899832487, + 0.0032214652746915817, + 0.0030860428232699633, + 0.004715615883469582, + 0.005302751902490854, + 0.003586464561522007, + 0.0038833024445921183, + 0.00558182829990983, + 0.0038002098444849253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.012537590228021145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tense, extreme close-up shot, captured in the style of competitive cinematography, follows the subject, a man with a resolute face and a sharp jawline, wearing a dark gray, moisture-wicking athletic T-shirt. His expression of concentration and a slight frown, as if he were in the thick of a decisive tournament. Under the cool, sharp sidelight of a desk lamp, his steady, powerful hands swiftly manipulate a minimalist black blood glucose meter. A slight motion blur captures the precise moment as he swiftly and decisively inserts a test strip into the side slot. Then, the camera's sharp focus shifts to the illuminated digital display, while his face and surroundings appear slightly out of focus. The numbers on the screen are jumping and changing rapidly and silently. Every flash is as gripping as a countdown, casting a faint blue light on his focused face and the skin on his knuckles that have turned white due to exertion. The whole picture is filled with a high degree of tension and concentration, sublimating a daily health monitoring behavior into a silent duel with oneself and racing against time.", + "seed": 24409, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_2/index_2370_seed_24409_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09343708101660013, + "vlm_logprob": -0.09343708101660013, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.09343708101660013, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.267393559217453 + ], + "opposite_logprob_tokens": [ + -1.5173935890197754 + ], + "logprob_reward": -0.267393559217453, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.267393559217453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20353087782859802, + 0.0380328968167305, + 0.057674143463373184, + 0.026770539581775665, + 0.05760766193270683, + 0.08187276870012283, + 0.2880055904388428, + 0.0970490500330925, + 0.08044971525669098, + 0.06900675594806671 + ], + "credit_map": [ + 0.019758472219109535, + 0.01778881996870041, + 0.023057159036397934, + 0.022498229518532753, + 0.037790220230817795, + 0.01986639015376568, + 0.03719215840101242, + 0.025579428300261497, + 0.0033879189286381006, + 0.004106343258172274, + 0.0034843378234654665, + 0.0021993345580995083, + 0.005084699485450983, + 0.006282694637775421, + 0.005989484954625368, + 0.007498085964471102, + 0.005428990814834833, + 0.007098136469721794, + 0.006046294700354338, + 0.008017892017960548, + 0.007124027702957392, + 0.007723389193415642, + 0.011862519197165966, + 0.004372896160930395, + 0.0028288126923143864, + 0.003283456899225712, + 0.0033710270654410124, + 0.00427214615046978, + 0.0036743255332112312, + 0.0037269042804837227, + 0.0030427747406065464, + 0.0025710933841764927, + 0.004655122756958008, + 0.005370541010051966, + 0.004456007853150368, + 0.006615149322897196, + 0.015768617391586304, + 0.008798426017165184, + 0.004911345895379782, + 0.00703245447948575, + 0.005424536764621735, + 0.012628800235688686, + 0.010100573301315308, + 0.011045696213841438, + 0.008946833200752735, + 0.012068474665284157, + 0.011257406324148178, + 0.01040045265108347, + 0.012813184410333633, + 0.017232375219464302, + 0.025868941098451614, + 0.014459501951932907, + 0.02249801903963089, + 0.10410542786121368, + 0.02649693749845028, + 0.06453119218349457, + 0.01023165974766016, + 0.02199839986860752, + 0.012233499437570572, + 0.017701921984553337, + 0.007478139828890562, + 0.011005870997905731, + 0.008231264539062977, + 0.008168294094502926, + 0.009839900769293308, + 0.013591921888291836, + 0.0074399844743311405, + 0.014434177428483963, + 0.007938334718346596, + 0.011301100254058838, + 0.006233337335288525, + 0.009670953266322613, + 0.00791507214307785, + 0.010345958173274994, + 0.01184912584722042, + 0.011710270307958126, + 0.006116720847785473, + 0.011410005390644073, + 0.005332622677087784, + 0.004326978232711554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.267393559217453 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07905983179807663 + ], + "opposite_logprob_tokens": [ + -2.5790598392486572 + ], + "logprob_reward": -0.07905983179807663, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07905983179807663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11900293081998825, + 0.12122754752635956, + 0.17772646248340607, + 0.12890546023845673, + 0.06992853432893753, + 0.22779729962348938, + 0.049884114414453506, + 0.04311782121658325, + 0.0245707668364048, + 0.03783904388546944 + ], + "credit_map": [ + 0.015231283381581306, + 0.013554020784795284, + 0.018965208902955055, + 0.011652852408587933, + 0.02006450854241848, + 0.015716515481472015, + 0.01248136255890131, + 0.011337186209857464, + 0.011292550712823868, + 0.013962797820568085, + 0.012393186800181866, + 0.008526572026312351, + 0.012926858849823475, + 0.01745215617120266, + 0.020596273243427277, + 0.024077141657471657, + 0.016754917800426483, + 0.0158443171530962, + 0.0296935997903347, + 0.021054891869425774, + 0.017469633370637894, + 0.023840028792619705, + 0.034762486815452576, + 0.018306594341993332, + 0.010444414801895618, + 0.018715940415859222, + 0.012384224683046341, + 0.023915285244584084, + 0.02534841187298298, + 0.01938684657216072, + 0.012020280584692955, + 0.006690065376460552, + 0.006847035605460405, + 0.00833424087613821, + 0.004939435515552759, + 0.0065474542789161205, + 0.0190015509724617, + 0.010842404328286648, + 0.0060651181265711784, + 0.0073512946255505085, + 0.0173649825155735, + 0.03217168152332306, + 0.034732554107904434, + 0.03395826369524002, + 0.0184040404856205, + 0.014792425557971, + 0.055783938616514206, + 0.0205894336104393, + 0.0023866950068622828, + 0.0030939884018152952, + 0.005229543894529343, + 0.0032179460395127535, + 0.003952398896217346, + 0.015369793400168419, + 0.0050611975602805614, + 0.011572552844882011, + 0.0035579779651015997, + 0.006793771870434284, + 0.005374053958803415, + 0.007334291003644466, + 0.0035111510660499334, + 0.005562799051403999, + 0.004678376019001007, + 0.006305398419499397, + 0.0025466512888669968, + 0.0030716722831130028, + 0.0025345277972519398, + 0.00455439044162631, + 0.001868577441200614, + 0.0033181060571223497, + 0.003016076982021332, + 0.0036607645452022552, + 0.005188017152249813, + 0.004970429930835962, + 0.004218334797769785, + 0.004638337995857, + 0.0032657880801707506, + 0.005710729397833347, + 0.004509612452238798, + 0.005337791983038187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.07905983179807663 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812231495976448 + ], + "opposite_logprob_tokens": [ + -3.2881224155426025 + ], + "logprob_reward": -0.03812231495976448, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812231495976448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13282135128974915, + 0.18433569371700287, + 0.10473912209272385, + 0.14833702147006989, + 0.13334700465202332, + 0.05604640021920204, + 0.08077620714902878, + 0.06536675989627838, + 0.06662250310182571, + 0.027607955038547516 + ], + "credit_map": [ + 0.01623372547328472, + 0.015667129307985306, + 0.02105553448200226, + 0.01480154786258936, + 0.016617393121123314, + 0.023967955261468887, + 0.009866433218121529, + 0.014611635357141495, + 0.0403066948056221, + 0.017483754083514214, + 0.021721629425883293, + 0.012007639743387699, + 0.014685600996017456, + 0.026997145265340805, + 0.03330578655004501, + 0.01782744564116001, + 0.009179250337183475, + 0.007760613691061735, + 0.019933046773076057, + 0.01696341671049595, + 0.01040706131607294, + 0.013778261840343475, + 0.013869984075427055, + 0.012847491540014744, + 0.01726214773952961, + 0.01798067055642605, + 0.021456260234117508, + 0.024039601907134056, + 0.01755194365978241, + 0.019215228036046028, + 0.019026927649974823, + 0.01180424727499485, + 0.010848300531506538, + 0.011016696691513062, + 0.011541621759533882, + 0.01441909372806549, + 0.03725602850317955, + 0.019928501918911934, + 0.011369328945875168, + 0.016967440024018288, + 0.005194688681513071, + 0.008206960745155811, + 0.005271999631077051, + 0.006312909070402384, + 0.0075027295388281345, + 0.006998033262789249, + 0.008019262924790382, + 0.008539813570678234, + 0.005432522855699062, + 0.0087837353348732, + 0.007207914255559444, + 0.004580353852361441, + 0.007071956526488066, + 0.027410969138145447, + 0.007084136363118887, + 0.01320461742579937, + 0.0046786414459347725, + 0.009005515836179256, + 0.007965872064232826, + 0.011039718054234982, + 0.006171654909849167, + 0.007507619448006153, + 0.008752061054110527, + 0.010245678946375847, + 0.0058049908839166164, + 0.007748421747237444, + 0.007920019328594208, + 0.009816688485443592, + 0.005731930956244469, + 0.009995860047638416, + 0.00805600080639124, + 0.011548594571650028, + 0.003301221877336502, + 0.003948620520532131, + 0.0030599578749388456, + 0.0055185286328196526, + 0.001988536212593317, + 0.004751088563352823, + 0.0023738646414130926, + 0.002666136482730508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.03812231495976448 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006516307592392 + ], + "opposite_logprob_tokens": [ + -2.6950652599334717 + ], + "logprob_reward": -0.07006516307592392, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006516307592392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13661685585975647, + 0.059969644993543625, + 0.058554068207740784, + 0.06869836896657944, + 0.11597506701946259, + 0.12994013726711273, + 0.22785590589046478, + 0.1091139167547226, + 0.06061637029051781, + 0.032659661024808884 + ], + "credit_map": [ + 0.01890435628592968, + 0.01469135470688343, + 0.023286068812012672, + 0.015735842287540436, + 0.015291125513613224, + 0.024346860125660896, + 0.010791209526360035, + 0.013570043258368969, + 0.006962658371776342, + 0.00546649657189846, + 0.0061760712414979935, + 0.005862316116690636, + 0.007210118696093559, + 0.008698297664523125, + 0.009545652195811272, + 0.0100480318069458, + 0.008529840968549252, + 0.007467078045010567, + 0.007522855885326862, + 0.006882695015519857, + 0.007748639211058617, + 0.008076262660324574, + 0.008458893746137619, + 0.0038678026758134365, + 0.007501410786062479, + 0.006590953096747398, + 0.00936767179518938, + 0.014312942512333393, + 0.007579726632684469, + 0.010212825611233711, + 0.0066335746087133884, + 0.006499257404357195, + 0.012221497483551502, + 0.011085704900324345, + 0.009209816344082355, + 0.013305509462952614, + 0.02985670045018196, + 0.01817712001502514, + 0.00950150191783905, + 0.01261721272021532, + 0.012127337045967579, + 0.011500214226543903, + 0.026799647137522697, + 0.01131844986230135, + 0.020511994138360023, + 0.01380801573395729, + 0.016842614859342575, + 0.017031876370310783, + 0.012939376756548882, + 0.016071679070591927, + 0.017526257783174515, + 0.011180107481777668, + 0.01855323277413845, + 0.08373356610536575, + 0.020142799243330956, + 0.04770888388156891, + 0.007552096154540777, + 0.01574118249118328, + 0.012680919840931892, + 0.012268532998859882, + 0.010151018388569355, + 0.011021507903933525, + 0.017394285649061203, + 0.022304363548755646, + 0.005759024526923895, + 0.007759640924632549, + 0.006908408831804991, + 0.007766354363411665, + 0.0061080483719706535, + 0.00873646978288889, + 0.007802440319210291, + 0.009775982238352299, + 0.0032805101945996284, + 0.00404232507571578, + 0.004647241439670324, + 0.006740384269505739, + 0.0028109580744057894, + 0.0059730662032961845, + 0.0028791800141334534, + 0.0022859973832964897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.07006516307592392 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544536031782627 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544536031782627, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544536031782627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14011622965335846, + 0.05438854545354843, + 0.204080268740654, + 0.023404041305184364, + 0.045785050839185715, + 0.07029794156551361, + 0.28802812099456787, + 0.03201161324977875, + 0.06891603767871857, + 0.07297220081090927 + ], + "credit_map": [ + 0.01714332215487957, + 0.014335166662931442, + 0.021332932636141777, + 0.017241565510630608, + 0.016217615455389023, + 0.019473455846309662, + 0.0185997374355793, + 0.01577242650091648, + 0.006947204936295748, + 0.007785130757838488, + 0.004670613445341587, + 0.003086341079324484, + 0.005951384548097849, + 0.007765451446175575, + 0.009089682251214981, + 0.009092736057937145, + 0.022703763097524643, + 0.022425392642617226, + 0.02695251628756523, + 0.021981505677103996, + 0.021300705149769783, + 0.027652284130454063, + 0.043956682085990906, + 0.01710738055408001, + 0.0022255978547036648, + 0.003181090345606208, + 0.003304117824882269, + 0.0047669666819274426, + 0.002620290732011199, + 0.0028524307999759912, + 0.0021709545981138945, + 0.0022825913038104773, + 0.003736173966899514, + 0.005174448247998953, + 0.004382758866995573, + 0.006094731390476227, + 0.009791339747607708, + 0.005910302046686411, + 0.004684970714151859, + 0.0060103219002485275, + 0.0055747684091329575, + 0.006525292061269283, + 0.008796772919595242, + 0.007351330481469631, + 0.010575253516435623, + 0.009817800484597683, + 0.0149369016289711, + 0.006719815544784069, + 0.013907278887927532, + 0.01310284249484539, + 0.022929413244128227, + 0.014701205305755138, + 0.02325914241373539, + 0.10880555212497711, + 0.025552673265337944, + 0.0657699927687645, + 0.003457232378423214, + 0.004728326573967934, + 0.0044240569695830345, + 0.004696467891335487, + 0.0025682486593723297, + 0.003607154358178377, + 0.003975267522037029, + 0.0045548551715910435, + 0.006094117183238268, + 0.00955860037356615, + 0.007581026758998632, + 0.010468849912285805, + 0.006836785934865475, + 0.009922679513692856, + 0.006819469388574362, + 0.011634504422545433, + 0.00994016882032156, + 0.009497769176959991, + 0.01446413341909647, + 0.009535718709230423, + 0.004843132570385933, + 0.012748685665428638, + 0.006550130434334278, + 0.00539244944229722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.012544536031782627 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tense, extreme close-up shot, captured in the style of competitive cinematography, follows the subject, a man with a resolute face and a sharp jawline, wearing a dark gray, moisture-wicking athletic T-shirt. His expression of concentration and a slight frown, as if he were in the thick of a decisive tournament. Under the cool, sharp sidelight of a desk lamp, his steady, powerful hands swiftly manipulate a minimalist black blood glucose meter. A slight motion blur captures the precise moment as he swiftly and decisively inserts a test strip into the side slot. Then, the camera's sharp focus shifts to the illuminated digital display, while his face and surroundings appear slightly out of focus. The numbers on the screen are jumping and changing rapidly and silently. Every flash is as gripping as a countdown, casting a faint blue light on his focused face and the skin on his knuckles that have turned white due to exertion. The whole picture is filled with a high degree of tension and concentration, sublimating a daily health monitoring behavior into a silent duel with oneself and racing against time.", + "seed": 24410, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_2/index_2370_seed_24410_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03256603553891182, + "vlm_logprob": -0.03256603553891182, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.03256603553891182, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013502226211130619 + ], + "opposite_logprob_tokens": [ + -5.01350212097168 + ], + "logprob_reward": -0.013502226211130619, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.013502226211130619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22496706247329712, + 0.09537474066019058, + 0.045425690710544586, + 0.02749716304242611, + 0.3570500612258911, + 0.03521500155329704, + 0.06311282515525818, + 0.0497237853705883, + 0.07658813893795013, + 0.02504555694758892 + ], + "credit_map": [ + 0.014608818106353283, + 0.010601780377328396, + 0.025117643177509308, + 0.009580020792782307, + 0.010318854823708534, + 0.08101105690002441, + 0.014342743903398514, + 0.05938614904880524, + 0.01258919294923544, + 0.00820227898657322, + 0.020244672894477844, + 0.014770392328500748, + 0.011961085721850395, + 0.0093842763453722, + 0.01041630282998085, + 0.007806534878909588, + 0.005264855455607176, + 0.004812616389244795, + 0.011532132513821125, + 0.006189572624862194, + 0.00506541458889842, + 0.0054215616546571255, + 0.004484322853386402, + 0.002655213698744774, + 0.0035835530143231153, + 0.003299441421404481, + 0.004785279277712107, + 0.0025762575678527355, + 0.0033544590696692467, + 0.002157848794013262, + 0.005230923183262348, + 0.0025094004813581705, + 0.022192925214767456, + 0.019271431490778923, + 0.011699806898832321, + 0.025865014642477036, + 0.14136900007724762, + 0.024789871647953987, + 0.015941748395562172, + 0.09592026472091675, + 0.0031993575394153595, + 0.0028186074923723936, + 0.002916697645559907, + 0.007319110445678234, + 0.0024170184042304754, + 0.005819510202854872, + 0.005923409480601549, + 0.004801290109753609, + 0.0061901407316327095, + 0.00645176088437438, + 0.005444901529699564, + 0.008970741182565689, + 0.006293649785220623, + 0.012061462737619877, + 0.009356525726616383, + 0.008343640714883804, + 0.004972672089934349, + 0.00491455988958478, + 0.005052273627370596, + 0.0060469359159469604, + 0.004226536490023136, + 0.010912483558058739, + 0.005652124993503094, + 0.007946198806166649, + 0.007837310433387756, + 0.007442529778927565, + 0.006534152198582888, + 0.008539184927940369, + 0.009057632647454739, + 0.012418561615049839, + 0.013587906025350094, + 0.011170858517289162, + 0.0023940291721373796, + 0.0026379614137113094, + 0.0023945742286741734, + 0.003719918429851532, + 0.002669826615601778, + 0.00283609377220273, + 0.004230888094753027, + 0.00416226452216506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.013502226211130619 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790700763463974 + ], + "opposite_logprob_tokens": [ + -2.5790700912475586 + ], + "logprob_reward": -0.0790700763463974, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790700763463974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18422119319438934, + 0.06552662700414658, + 0.08976571261882782, + 0.03487993776798248, + 0.2869700491428375, + 0.045309726148843765, + 0.08296079188585281, + 0.07213717699050903, + 0.05137587711215019, + 0.08685295283794403 + ], + "credit_map": [ + 0.015033106319606304, + 0.009805737994611263, + 0.02251674421131611, + 0.00825305376201868, + 0.008063306100666523, + 0.061573319137096405, + 0.01291289459913969, + 0.04606301710009575, + 0.00849697832018137, + 0.0075008864514529705, + 0.00933110248297453, + 0.008734222501516342, + 0.01006025169044733, + 0.006365523673593998, + 0.008937747217714787, + 0.00609990768134594, + 0.009111291728913784, + 0.008926705457270145, + 0.016547508537769318, + 0.01078199502080679, + 0.010858085937798023, + 0.011621017009019852, + 0.014493758790194988, + 0.007425336167216301, + 0.0029130466282367706, + 0.0033444571308791637, + 0.004133522976189852, + 0.0037861447781324387, + 0.00728355860337615, + 0.0021782268304377794, + 0.008849617093801498, + 0.0023913634940981865, + 0.020717553794384003, + 0.026859527453780174, + 0.020323878154158592, + 0.02681698091328144, + 0.090436190366745, + 0.028483176603913307, + 0.018970375880599022, + 0.05436236038804054, + 0.004421868361532688, + 0.006178170442581177, + 0.002852392615750432, + 0.006494992412626743, + 0.005171258933842182, + 0.008089398965239525, + 0.0056619830429553986, + 0.00643965695053339, + 0.008551204577088356, + 0.00906121451407671, + 0.006886943243443966, + 0.011479592882096767, + 0.008161528967320919, + 0.014744442887604237, + 0.012546908110380173, + 0.01152894925326109, + 0.0075263213366270065, + 0.007444875314831734, + 0.0080238813534379, + 0.009213756769895554, + 0.005862186662852764, + 0.014538495801389217, + 0.008299698121845722, + 0.011227958835661411, + 0.005920421797782183, + 0.005565992556512356, + 0.0046815588138997555, + 0.006350575014948845, + 0.004120105877518654, + 0.0074982005171477795, + 0.009426330216228962, + 0.007812686264514923, + 0.0064058504067361355, + 0.0075362687930464745, + 0.011896908283233643, + 0.012321810238063335, + 0.006527133751660585, + 0.016878843307495117, + 0.011053628288209438, + 0.014232506044209003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.0790700763463974 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088770814239979 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011088770814239979, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088770814239979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.084388867020607, + 0.04751792922616005, + 0.06144294515252113, + 0.04138680174946785, + 0.16095511615276337, + 0.10470530390739441, + 0.10022284835577011, + 0.1790982335805893, + 0.11939918249845505, + 0.10088275372982025 + ], + "credit_map": [ + 0.009963290765881538, + 0.005191234406083822, + 0.007279976736754179, + 0.006600709166377783, + 0.005293890833854675, + 0.02140030451118946, + 0.008672311902046204, + 0.019987158477306366, + 0.00643711956217885, + 0.00776243582367897, + 0.006700486876070499, + 0.01018163189291954, + 0.004035166930407286, + 0.004569326061755419, + 0.003623462747782469, + 0.004208301194012165, + 0.006681717932224274, + 0.007623748853802681, + 0.012491893023252487, + 0.007816673256456852, + 0.009220346808433533, + 0.00795697420835495, + 0.0060657006688416, + 0.003585897386074066, + 0.00407804548740387, + 0.00479446304962039, + 0.004803038202226162, + 0.004189386032521725, + 0.009045325219631195, + 0.002632725052535534, + 0.00793569628149271, + 0.003908122889697552, + 0.015622220933437347, + 0.017103692516684532, + 0.012546258978545666, + 0.014537738636136055, + 0.04672108218073845, + 0.012785738334059715, + 0.010707615874707699, + 0.030930787324905396, + 0.012322155758738518, + 0.01011658925563097, + 0.008699573576450348, + 0.02048666775226593, + 0.007995137944817543, + 0.014197159558534622, + 0.01479377131909132, + 0.016094252467155457, + 0.011081301607191563, + 0.010222177021205425, + 0.00834363792091608, + 0.0138814402744174, + 0.009743737056851387, + 0.018214695155620575, + 0.015960393473505974, + 0.012775471433997154, + 0.016409702599048615, + 0.011649863794445992, + 0.008554385043680668, + 0.032025739550590515, + 0.011747585609555244, + 0.01492485124617815, + 0.050867710262537, + 0.03291841223835945, + 0.013566307723522186, + 0.012394316494464874, + 0.011247779242694378, + 0.013938468880951405, + 0.01064271666109562, + 0.01836196705698967, + 0.021368777379393578, + 0.017878863960504532, + 0.007728529162704945, + 0.008982295170426369, + 0.011796213686466217, + 0.01736423745751381, + 0.007984449155628681, + 0.014536351896822453, + 0.01631004549562931, + 0.01618064194917679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.011088770814239979 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043100737035274506 + ], + "opposite_logprob_tokens": [ + -3.1681008338928223 + ], + "logprob_reward": -0.043100737035274506, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043100737035274506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14389969408512115, + 0.10475171357393265, + 0.07302426546812057, + 0.04470498487353325, + 0.25019317865371704, + 0.10877715796232224, + 0.06834638118743896, + 0.08281715214252472, + 0.034694235771894455, + 0.08879121392965317 + ], + "credit_map": [ + 0.014614556916058064, + 0.008167915046215057, + 0.0187132116407156, + 0.00905581098049879, + 0.008459823206067085, + 0.04410368576645851, + 0.011050492525100708, + 0.02973419427871704, + 0.016997165977954865, + 0.018276013433933258, + 0.012059478089213371, + 0.016884764656424522, + 0.012291789054870605, + 0.010624360293149948, + 0.008507026359438896, + 0.009111114777624607, + 0.012502768076956272, + 0.0039337254129350185, + 0.021923717111349106, + 0.008069724775850773, + 0.007438467815518379, + 0.005854727700352669, + 0.005932052154093981, + 0.007369083818048239, + 0.004837001673877239, + 0.005663215648382902, + 0.007857690565288067, + 0.004109114408493042, + 0.006114363670349121, + 0.004543090704828501, + 0.005908477120101452, + 0.005672028753906488, + 0.018302682787179947, + 0.02354484051465988, + 0.011844942346215248, + 0.019185520708560944, + 0.08303571492433548, + 0.018031083047389984, + 0.013633150607347488, + 0.06261523813009262, + 0.014547801576554775, + 0.01363407913595438, + 0.007101651281118393, + 0.016740521416068077, + 0.012051234021782875, + 0.0157026257365942, + 0.01572379842400551, + 0.013275445438921452, + 0.007609155960381031, + 0.00720930565148592, + 0.010234920307993889, + 0.007166149560362101, + 0.0056878384202718735, + 0.00995851494371891, + 0.009481175802648067, + 0.010999318212270737, + 0.009519866667687893, + 0.0070548695512115955, + 0.005705371033400297, + 0.0140271857380867, + 0.005082307383418083, + 0.009994395077228546, + 0.018425090238451958, + 0.013008072972297668, + 0.004666314925998449, + 0.0031218044459819794, + 0.004445598926395178, + 0.004715930204838514, + 0.0038861283101141453, + 0.0038417079485952854, + 0.005122880917042494, + 0.004893871955573559, + 0.007567047141492367, + 0.009410299360752106, + 0.013021060265600681, + 0.012191088870167732, + 0.008882810361683369, + 0.010893868282437325, + 0.013899214565753937, + 0.012925824150443077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.043100737035274506 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016068367287516594 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068367287516594, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068367287516594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08292955905199051, + 0.05549554154276848, + 0.12880070507526398, + 0.034016210585832596, + 0.24742192029953003, + 0.0899345874786377, + 0.1220482811331749, + 0.1556362360715866, + 0.06990890204906464, + 0.013808085583150387 + ], + "credit_map": [ + 0.004486681427806616, + 0.0031191366724669933, + 0.010482597164809704, + 0.0040365965105593204, + 0.0031490055844187737, + 0.03232693299651146, + 0.002861327026039362, + 0.022467270493507385, + 0.007578122429549694, + 0.007169114891439676, + 0.006436505820602179, + 0.0066684638150036335, + 0.008278518915176392, + 0.006024118047207594, + 0.007600185461342335, + 0.005740510765463114, + 0.0281063225120306, + 0.007949784398078918, + 0.05100734159350395, + 0.010759871453046799, + 0.008321464993059635, + 0.009244469925761223, + 0.006145603954792023, + 0.007265827152878046, + 0.003143729642033577, + 0.0017920699901878834, + 0.0035221974831074476, + 0.0028975761961191893, + 0.010500112548470497, + 0.001900352188386023, + 0.00822493713349104, + 0.0020352317951619625, + 0.01631884090602398, + 0.009590099565684795, + 0.012191131711006165, + 0.019054101780056953, + 0.09449028223752975, + 0.018323732540011406, + 0.010453141294419765, + 0.06700058281421661, + 0.008597581647336483, + 0.010577309876680374, + 0.00872963760048151, + 0.013324784114956856, + 0.00760024506598711, + 0.019919734448194504, + 0.00959695689380169, + 0.011588335037231445, + 0.01253349706530571, + 0.014108383096754551, + 0.01342763565480709, + 0.018381431698799133, + 0.011426159180700779, + 0.019953714683651924, + 0.013898833654820919, + 0.01831861585378647, + 0.01459317933768034, + 0.014052641578018665, + 0.015200426802039146, + 0.02270394377410412, + 0.011017357930541039, + 0.024873578920960426, + 0.02893258072435856, + 0.02426251955330372, + 0.007737041916698217, + 0.00712297298014164, + 0.006355798337608576, + 0.008446825668215752, + 0.007416098844259977, + 0.010577122680842876, + 0.011960066854953766, + 0.010292976163327694, + 0.0008402729290537536, + 0.0010744620813056827, + 0.0014467232394963503, + 0.002475735964253545, + 0.0011414416367188096, + 0.002694811439141631, + 0.002189117018133402, + 0.001945520518347621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 2 + ], + "selected_reward": -0.016068367287516594 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tense, extreme close-up shot, captured in the style of competitive cinematography, follows the subject, a man with a resolute face and a sharp jawline, wearing a dark gray, moisture-wicking athletic T-shirt. His expression of concentration and a slight frown, as if he were in the thick of a decisive tournament. Under the cool, sharp sidelight of a desk lamp, his steady, powerful hands swiftly manipulate a minimalist black blood glucose meter. A slight motion blur captures the precise moment as he swiftly and decisively inserts a test strip into the side slot. Then, the camera's sharp focus shifts to the illuminated digital display, while his face and surroundings appear slightly out of focus. The numbers on the screen are jumping and changing rapidly and silently. Every flash is as gripping as a countdown, casting a faint blue light on his focused face and the skin on his knuckles that have turned white due to exertion. The whole picture is filled with a high degree of tension and concentration, sublimating a daily health monitoring behavior into a silent duel with oneself and racing against time.", + "seed": 24411, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_2/index_2370_seed_24411_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10299895815551281, + "vlm_logprob": -0.10299895815551281, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.10299895815551281, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4015319049358368 + ], + "opposite_logprob_tokens": [ + -1.1515319347381592 + ], + "logprob_reward": -0.4015319049358368, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.4015319049358368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16600804030895233, + 0.1474974900484085, + 0.1551346331834793, + 0.07048887759447098, + 0.05951050668954849, + 0.09939450770616531, + 0.08469437807798386, + 0.033230844885110855, + 0.03285975754261017, + 0.15118099749088287 + ], + "credit_map": [ + 0.019047360867261887, + 0.019646326079964638, + 0.03195994719862938, + 0.014553125947713852, + 0.025979574769735336, + 0.020909521728754044, + 0.018669068813323975, + 0.0152431083843112, + 0.030839743092656136, + 0.04032840579748154, + 0.011267753317952156, + 0.009668697603046894, + 0.026239288970828056, + 0.007507036905735731, + 0.01361704058945179, + 0.008029528893530369, + 0.008802324533462524, + 0.005421512294560671, + 0.046496737748384476, + 0.0156992357224226, + 0.054686617106199265, + 0.008278008550405502, + 0.009671385399997234, + 0.006078819744288921, + 0.010425091721117496, + 0.006193609442561865, + 0.010114372707903385, + 0.006515604443848133, + 0.008612039498984814, + 0.009230914525687695, + 0.009034362621605396, + 0.010362877510488033, + 0.0049825976602733135, + 0.005189171992242336, + 0.013663730584084988, + 0.005956599023193121, + 0.006348418537527323, + 0.011383887380361557, + 0.006686152890324593, + 0.005299951881170273, + 0.0072066932916641235, + 0.006838180124759674, + 0.01729673705995083, + 0.017357762902975082, + 0.00780341774225235, + 0.016923068091273308, + 0.0176510289311409, + 0.008317624218761921, + 0.007084468845278025, + 0.011639676988124847, + 0.009162459522485733, + 0.012318278662860394, + 0.011817782185971737, + 0.009855254553258419, + 0.012115485966205597, + 0.010700971819460392, + 0.002010364318266511, + 0.002299913205206394, + 0.003879243740811944, + 0.0053457277826964855, + 0.00371652957983315, + 0.007134160492569208, + 0.0055465055629611015, + 0.0032984043937176466, + 0.0027465166058391333, + 0.0028200524393469095, + 0.0033945178147405386, + 0.005921629723161459, + 0.003399129491299391, + 0.003521948354318738, + 0.007951982319355011, + 0.0031039821915328503, + 0.010132640600204468, + 0.007016263902187347, + 0.029695820063352585, + 0.011130448430776596, + 0.04548968747258186, + 0.0192670039832592, + 0.014444777742028236, + 0.014004349708557129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.4015319049358368 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05502312257885933 + ], + "opposite_logprob_tokens": [ + -2.930023193359375 + ], + "logprob_reward": -0.05502312257885933, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05502312257885933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056139107793569565, + 0.09292034804821014, + 0.3263700008392334, + 0.09902287274599075, + 0.030049994587898254, + 0.04816019535064697, + 0.04082384333014488, + 0.020021548494696617, + 0.07027584314346313, + 0.21621626615524292 + ], + "credit_map": [ + 0.006895489525049925, + 0.0055691152811050415, + 0.013294266536831856, + 0.005889675114303827, + 0.006386500783264637, + 0.006089457776397467, + 0.00587709154933691, + 0.006137508433312178, + 0.013703860342502594, + 0.0181355569511652, + 0.010474789887666702, + 0.011334933340549469, + 0.009541940875351429, + 0.0074425796046853065, + 0.011014224961400032, + 0.011272463947534561, + 0.01772223971784115, + 0.011335794813930988, + 0.09777597337961197, + 0.031070055440068245, + 0.1200859546661377, + 0.01500389538705349, + 0.022062193602323532, + 0.011313890106976032, + 0.009349015541374683, + 0.0067472816444933414, + 0.01282823272049427, + 0.018474655225872993, + 0.014309592545032501, + 0.010325444862246513, + 0.019663646817207336, + 0.007325000129640102, + 0.0032775092404335737, + 0.002353836316615343, + 0.007262731436640024, + 0.003753876546397805, + 0.002329684793949127, + 0.004800508730113506, + 0.0039954702369868755, + 0.002276376821100712, + 0.005793765652924776, + 0.004656484816223383, + 0.00752475718036294, + 0.006626683287322521, + 0.006050292402505875, + 0.004654696211218834, + 0.009300017729401588, + 0.0035534952767193317, + 0.003465376328676939, + 0.003069013822823763, + 0.007817440666258335, + 0.005991192068904638, + 0.010172855108976364, + 0.0030710569117218256, + 0.004704700317233801, + 0.002532203681766987, + 0.001693671802058816, + 0.00232902355492115, + 0.0031513310968875885, + 0.0030555797275155783, + 0.001680161221884191, + 0.0031890259124338627, + 0.0029978605452924967, + 0.0019248920725658536, + 0.007969625294208527, + 0.004477100446820259, + 0.01285022497177124, + 0.01096627302467823, + 0.01184860896319151, + 0.006216541398316622, + 0.01001670304685831, + 0.005930758081376553, + 0.012741089798510075, + 0.007074100896716118, + 0.03558111563324928, + 0.012684483081102371, + 0.0871676430106163, + 0.017232688143849373, + 0.013762010261416435, + 0.029973121359944344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.05502312257885933 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552423402667046 + ], + "opposite_logprob_tokens": [ + -4.387552261352539 + ], + "logprob_reward": -0.012552423402667046, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552423402667046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037016093730926514, + 0.11691299825906754, + 0.375540554523468, + 0.030892327427864075, + 0.027701741084456444, + 0.07433781027793884, + 0.030505340546369553, + 0.04977339506149292, + 0.11303111165761948, + 0.1442885845899582 + ], + "credit_map": [ + 0.0046187518164515495, + 0.003676283173263073, + 0.004927642643451691, + 0.0031688741873949766, + 0.005988394841551781, + 0.006454325281083584, + 0.003929519560188055, + 0.004252304323017597, + 0.025572394952178, + 0.033831145614385605, + 0.009454506449401379, + 0.008654811419546604, + 0.014289233833551407, + 0.006313633639365435, + 0.009849228896200657, + 0.008948050439357758, + 0.019637735560536385, + 0.012097550556063652, + 0.11056046932935715, + 0.03309965506196022, + 0.14738601446151733, + 0.017240453511476517, + 0.022643106058239937, + 0.012875586748123169, + 0.0029089576564729214, + 0.0033417954109609127, + 0.005011807661503553, + 0.0036036630626767874, + 0.004184218123555183, + 0.003897130722180009, + 0.00475301081314683, + 0.003191745840013027, + 0.002009514020755887, + 0.0019088286207988858, + 0.006865205708891153, + 0.0027887574397027493, + 0.0022893936838954687, + 0.0051459986716508865, + 0.003394699189811945, + 0.003299343166872859, + 0.008127737790346146, + 0.006410390604287386, + 0.009728937409818172, + 0.01666911691427231, + 0.006302802823483944, + 0.006121051963418722, + 0.015196003019809723, + 0.005781771615147591, + 0.0024070923682302237, + 0.0015975707210600376, + 0.004809993784874678, + 0.005707148928195238, + 0.00566022889688611, + 0.0030041036661714315, + 0.004748220089823008, + 0.0025709832552820444, + 0.004040142055600882, + 0.003973717801272869, + 0.007693632040172815, + 0.009027870371937752, + 0.004129273351281881, + 0.006210985593497753, + 0.009730196557939053, + 0.004967578686773777, + 0.010277656838297844, + 0.006915026344358921, + 0.014369980432093143, + 0.021064510568976402, + 0.015507693402469158, + 0.01077002938836813, + 0.023137155920267105, + 0.01098906621336937, + 0.008177139796316624, + 0.0047983694821596146, + 0.025763167068362236, + 0.009613662958145142, + 0.05666247382760048, + 0.010843269526958466, + 0.009950215928256512, + 0.01848028041422367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.012552423402667046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029815100133419037 + ], + "opposite_logprob_tokens": [ + -3.529815196990967 + ], + "logprob_reward": -0.029815100133419037, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029815100133419037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035472359508275986, + 0.057086966931819916, + 0.4850694537162781, + 0.01402648538351059, + 0.01605456881225109, + 0.049899447709321976, + 0.028390804305672646, + 0.025602171197533607, + 0.07064209133386612, + 0.21775560081005096 + ], + "credit_map": [ + 0.0043217213824391365, + 0.003424357855692506, + 0.006896595004945993, + 0.003991449251770973, + 0.004959918558597565, + 0.004482480231672525, + 0.0037940063048154116, + 0.003601829754188657, + 0.01085734274238348, + 0.014630270190536976, + 0.005125660449266434, + 0.004800340626388788, + 0.006158086005598307, + 0.003739048959687352, + 0.005604397505521774, + 0.00617181695997715, + 0.02456900104880333, + 0.015420257113873959, + 0.14447475969791412, + 0.04348047822713852, + 0.1900547593832016, + 0.021577859297394753, + 0.0290034431964159, + 0.01648888923227787, + 0.0015179337933659554, + 0.001367267221212387, + 0.0020131028722971678, + 0.0017259935848414898, + 0.0015233242884278297, + 0.0016593370819464326, + 0.003073364030569792, + 0.0011461625108495355, + 0.0017637504497542977, + 0.0015921565936878324, + 0.0022319480776786804, + 0.0019928726833313704, + 0.0015733204782009125, + 0.0027265669777989388, + 0.001768818823620677, + 0.0024051342625170946, + 0.0046553052961826324, + 0.0048576039262115955, + 0.009047169238328934, + 0.009334508329629898, + 0.004644136410206556, + 0.004164581652730703, + 0.009344571270048618, + 0.003851570887491107, + 0.00298698665574193, + 0.0033641033805906773, + 0.003983779344707727, + 0.003417993662878871, + 0.006777804344892502, + 0.002827594755217433, + 0.003224059008061886, + 0.0018084823386743665, + 0.0016226681182160974, + 0.0018565303180366755, + 0.0033842301927506924, + 0.004165332764387131, + 0.0025120796635746956, + 0.005903678480535746, + 0.003973997198045254, + 0.0021836543455719948, + 0.005847658030688763, + 0.004333212040364742, + 0.009011098183691502, + 0.013599327765405178, + 0.009278962388634682, + 0.006772502791136503, + 0.014592808671295643, + 0.007206527050584555, + 0.010135205462574959, + 0.007093495223671198, + 0.017705725505948067, + 0.0170209389179945, + 0.06985693424940109, + 0.012387050315737724, + 0.02105569653213024, + 0.06250055134296417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.029815100133419037 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a man with a resolute face and sharp jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark gray moisture-wicking athletic T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man manipulating a minimalist black blood glucose meter with his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man insert a test strip into the side slot of the device?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus shift from the man's hands to the illuminated digital display while blurring the face and surroundings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072239726781845 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072239726781845, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072239726781845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022031519562005997, + 0.02183327078819275, + 0.11283551901578903, + 0.012000657618045807, + 0.023374047130346298, + 0.0641627162694931, + 0.019788602367043495, + 0.0151356291025877, + 0.18456952273845673, + 0.5242685079574585 + ], + "credit_map": [ + 0.002981089288368821, + 0.0027758837677538395, + 0.003703815396875143, + 0.002317867474630475, + 0.0029246900230646133, + 0.002978370990604162, + 0.002507520141080022, + 0.0018422825960442424, + 0.0037958475295454264, + 0.0048432969488203526, + 0.0019913623109459877, + 0.0017220821464434266, + 0.0038893986493349075, + 0.001327984151430428, + 0.002033373573794961, + 0.0022299252450466156, + 0.006009430158883333, + 0.0035011409781873226, + 0.03403739258646965, + 0.009755056351423264, + 0.043845728039741516, + 0.005156360566616058, + 0.006440266035497189, + 0.004090141039341688, + 0.0017133515793830156, + 0.0009769852040335536, + 0.0019709928892552853, + 0.0012746022548526525, + 0.0011562717845663428, + 0.0011911258334293962, + 0.0023196241818368435, + 0.0013977043563500047, + 0.002354079158976674, + 0.0017721593612805009, + 0.006029640324413776, + 0.002323405584320426, + 0.0023507368750870228, + 0.0038075412157922983, + 0.0022892269771546125, + 0.0024472572840750217, + 0.005422451067715883, + 0.005808526650071144, + 0.009167983196675777, + 0.010926802642643452, + 0.006644311361014843, + 0.008412927389144897, + 0.012384171597659588, + 0.005395535845309496, + 0.0013420782051980495, + 0.0018914584070444107, + 0.0031337072141468525, + 0.0033965574111789465, + 0.003388581797480583, + 0.0023050052113831043, + 0.0026883569080382586, + 0.001642856514081359, + 0.001230324269272387, + 0.0012320164823904634, + 0.002060831757262349, + 0.002627944340929389, + 0.0018311130115762353, + 0.001516318996436894, + 0.003259167307987809, + 0.0013779139844700694, + 0.020944958552718163, + 0.01403750479221344, + 0.026965634897351265, + 0.03191978856921196, + 0.02674764022231102, + 0.01686917617917061, + 0.03139040991663933, + 0.015694405883550644, + 0.027704596519470215, + 0.01568823866546154, + 0.06938108056783676, + 0.03262975439429283, + 0.20303501188755035, + 0.028683993965387344, + 0.03468512371182442, + 0.11246073246002197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.016072239726781845 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_3/index_tensor([2747, 2747])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_3/index_tensor([2747, 2747])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3d54b2dbd1ed3fa62a0dd180668a39435f744c8d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_3/index_tensor([2747, 2747])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a medium shot, the camera moving forward incredibly slowly, creating an intimate feeling of glimpsing a private moment. In a bright, yet empty corner of a late-night convenience store, a young couple cuddles by a large floor-to-ceiling window. The man, wearing a loose, off-white hooded sweatshirt, sits sideways, his arm gently wrapped around the woman's shoulders. The woman, wearing an earth-toned knit cardigan, rests her full weight securely on him, her head resting gently on his shoulder, her hands cradling a steaming white paper cup. The fluorescent lighting inside the convenience store is deliberately rendered in a warm, vintage hue, suffusing the image with a soft golden glow and permeating the air with a nostalgic cinematic quality. Outside, the night is deep, only scattered streetlights and the faint trails of distant car taillights cast a stark contrast to the warm atmosphere inside. As the camera steadily and calmly zooms in, the picture gradually focuses on the faces of the two of them. They both have their eyes closed, with a subtle but very real smile on the corners of their mouths. Their cheeks appear particularly soft under the warm light, and every strand of hair is clearly visible, perfectly capturing the kind of contentment and tranquility found in ordinary moments that requires no words.", + "seed": 23179, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_3/index_2747_seed_23179_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03816173868253827, + "vlm_logprob": -0.03816173868253827, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.03816173868253827, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13217252492904663 + ], + "opposite_logprob_tokens": [ + -2.1321725845336914 + ], + "logprob_reward": -0.13217252492904663, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13217252492904663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07044707983732224, + 0.04081256315112114, + 0.10516530275344849, + 0.20200641453266144, + 0.15158872306346893, + 0.040409885346889496, + 0.10992732644081116, + 0.12421068549156189, + 0.060137197375297546, + 0.0952947586774826 + ], + "credit_map": [ + 0.009774849750101566, + 0.006177866365760565, + 0.008058536797761917, + 0.011214413680136204, + 0.009433133527636528, + 0.011037412099540234, + 0.008096486330032349, + 0.006654381286352873, + 0.005134339444339275, + 0.003594191512092948, + 0.007767635863274336, + 0.006457657087594271, + 0.0041135321371257305, + 0.006050495430827141, + 0.0036426640581339598, + 0.004052046686410904, + 0.0170261450111866, + 0.007062331307679415, + 0.008068406954407692, + 0.033172886818647385, + 0.006966413464397192, + 0.009626464918255806, + 0.00844603031873703, + 0.01479662861675024, + 0.05167509987950325, + 0.01316116563975811, + 0.013542641885578632, + 0.07090768963098526, + 0.012168977409601212, + 0.009981249459087849, + 0.014643899165093899, + 0.015925686806440353, + 0.012338669039309025, + 0.01219856645911932, + 0.008334916085004807, + 0.041165970265865326, + 0.01522817648947239, + 0.009959015995264053, + 0.042303700000047684, + 0.010059702210128307, + 0.003234956646338105, + 0.006319657899439335, + 0.0035625756718218327, + 0.004967983812093735, + 0.007754691876471043, + 0.004751287400722504, + 0.005169412586838007, + 0.004649324342608452, + 0.010686339810490608, + 0.016275886446237564, + 0.011995923705399036, + 0.015213118866086006, + 0.02023637853562832, + 0.011771374382078648, + 0.012463638558983803, + 0.011284665204584599, + 0.014216741546988487, + 0.01804669201374054, + 0.011247445829212666, + 0.01645411178469658, + 0.023929817602038383, + 0.013696735724806786, + 0.016885627061128616, + 0.009733504615724087, + 0.006171469110995531, + 0.008715696632862091, + 0.009441452100872993, + 0.009151091799139977, + 0.008415691554546356, + 0.00471800472587347, + 0.00953820999711752, + 0.003985579591244459, + 0.008267218247056007, + 0.010478129610419273, + 0.018248744308948517, + 0.008228051476180553, + 0.013229160569608212, + 0.005937756504863501, + 0.026354406028985977, + 0.004551293794065714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.13217252492904663 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012567255645990372 + ], + "opposite_logprob_tokens": [ + -4.387567043304443 + ], + "logprob_reward": -0.012567255645990372, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012567255645990372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07902427017688751, + 0.0666528269648552, + 0.17056550085544586, + 0.12402375042438507, + 0.11768664419651031, + 0.03378570079803467, + 0.04925639554858208, + 0.05966645106673241, + 0.1319596767425537, + 0.16737882792949677 + ], + "credit_map": [ + 0.008700608275830746, + 0.00822482630610466, + 0.009289167821407318, + 0.011007341556251049, + 0.01048815343528986, + 0.01560064498335123, + 0.008428169414401054, + 0.007285355124622583, + 0.00658793281763792, + 0.006795079912990332, + 0.007893434725701809, + 0.009510962292551994, + 0.010816281661391258, + 0.007987244985997677, + 0.008263488300144672, + 0.008798402734100819, + 0.021939491853117943, + 0.011890217661857605, + 0.012223494239151478, + 0.06529193371534348, + 0.008513261564075947, + 0.008213181048631668, + 0.01144980825483799, + 0.031044095754623413, + 0.022616920992732048, + 0.012437318451702595, + 0.009680068120360374, + 0.02309957705438137, + 0.015828926116228104, + 0.011020692065358162, + 0.01616436429321766, + 0.01317587774246931, + 0.012273837812244892, + 0.011797904036939144, + 0.011078618466854095, + 0.01915206015110016, + 0.01948079653084278, + 0.012300992384552956, + 0.021595805883407593, + 0.010006630793213844, + 0.002903455402702093, + 0.003054574364796281, + 0.004168496932834387, + 0.005037354771047831, + 0.004712868947535753, + 0.003427551593631506, + 0.006975822150707245, + 0.0035055773332715034, + 0.004229734651744366, + 0.006724510807543993, + 0.0060713994316756725, + 0.007290626410394907, + 0.009159648790955544, + 0.004626377951353788, + 0.007183233741670847, + 0.003970867022871971, + 0.009005250409245491, + 0.006440950557589531, + 0.005693602375686169, + 0.011794443242251873, + 0.007433184422552586, + 0.007528665941208601, + 0.0076068732887506485, + 0.004163477104157209, + 0.009453201666474342, + 0.014040371403098106, + 0.03585672378540039, + 0.03045230731368065, + 0.01122924592345953, + 0.007264076266437769, + 0.012232000939548016, + 0.011431735008955002, + 0.014610457234084606, + 0.019613783806562424, + 0.020992835983633995, + 0.02352573536336422, + 0.026155438274145126, + 0.020261608064174652, + 0.0307257492095232, + 0.01149321161210537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.012567255645990372 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641940541565418 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.008641940541565418, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641940541565418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04789162054657936, + 0.023566560819745064, + 0.28586483001708984, + 0.08223474770784378, + 0.12837880849838257, + 0.12772411108016968, + 0.019801734015345573, + 0.04084159806370735, + 0.20359568297863007, + 0.04010031744837761 + ], + "credit_map": [ + 0.0037062771152704954, + 0.004858695901930332, + 0.0046071805991232395, + 0.005746397655457258, + 0.008828765712678432, + 0.006539104040712118, + 0.006896480917930603, + 0.006708718836307526, + 0.002110752509906888, + 0.0018901759758591652, + 0.0033607250079512596, + 0.003091639606282115, + 0.004181778524070978, + 0.0030094198882579803, + 0.002578025683760643, + 0.0033440440893173218, + 0.04767115041613579, + 0.01992121711373329, + 0.01496930979192257, + 0.1149601936340332, + 0.00946725346148014, + 0.011983174830675125, + 0.019487548619508743, + 0.04740498587489128, + 0.006980789825320244, + 0.01662791520357132, + 0.009848046116530895, + 0.010641603730618954, + 0.012619659304618835, + 0.008804126642644405, + 0.009700941853225231, + 0.007011663168668747, + 0.010046985000371933, + 0.015083671547472477, + 0.006947692483663559, + 0.033008426427841187, + 0.012559118680655956, + 0.010149913839995861, + 0.033147960901260376, + 0.0074350410141050816, + 0.016282664611935616, + 0.01619197241961956, + 0.019828666001558304, + 0.016304077580571175, + 0.02033878304064274, + 0.012252042070031166, + 0.014403201639652252, + 0.012122695334255695, + 0.0021083643659949303, + 0.0025031676050275564, + 0.002905214438214898, + 0.0026286921929568052, + 0.003928179387003183, + 0.0021042742300778627, + 0.001969843404367566, + 0.0016539979260414839, + 0.0046812486834824085, + 0.005081303417682648, + 0.004559995140880346, + 0.008905152790248394, + 0.005328787956386805, + 0.004020188935101032, + 0.0047698053531348705, + 0.0034951178822666407, + 0.016369687393307686, + 0.012716435827314854, + 0.049904074519872665, + 0.07437259703874588, + 0.007077721878886223, + 0.005178005434572697, + 0.028290634974837303, + 0.009686535224318504, + 0.003277047537267208, + 0.004463129676878452, + 0.0074525983072817326, + 0.004050501622259617, + 0.004289825446903706, + 0.00369275058619678, + 0.010852291248738766, + 0.002022172324359417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.008641940541565418 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631197106093168 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631197106093168, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631197106093168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07604916393756866, + 0.055991388857364655, + 0.2732877731323242, + 0.06489107757806778, + 0.04506811499595642, + 0.07847219705581665, + 0.03737153485417366, + 0.05174735188484192, + 0.2868037223815918, + 0.030317697674036026 + ], + "credit_map": [ + 0.0069061266258358955, + 0.008800550363957882, + 0.008305765688419342, + 0.011828295886516571, + 0.01204855926334858, + 0.009911042638123035, + 0.010158547200262547, + 0.008090268820524216, + 0.004712654277682304, + 0.005113744176924229, + 0.004962553270161152, + 0.00796719640493393, + 0.010169987566769123, + 0.0071162148378789425, + 0.007842961698770523, + 0.008106070570647717, + 0.043783340603113174, + 0.01885976828634739, + 0.014629628509283066, + 0.10974034667015076, + 0.00889622326940298, + 0.011594090610742569, + 0.018466446548700333, + 0.047317925840616226, + 0.010914226993918419, + 0.011587601155042648, + 0.00620265444740653, + 0.010839023627340794, + 0.008566455915570259, + 0.006000221241265535, + 0.006424159277230501, + 0.004356732591986656, + 0.0040905955247581005, + 0.005548463203012943, + 0.004419079516083002, + 0.007058521267026663, + 0.007956945337355137, + 0.005194262135773897, + 0.006865874398499727, + 0.003934368025511503, + 0.007797831669449806, + 0.00701522920280695, + 0.009830744937062263, + 0.011130292899906635, + 0.008983522653579712, + 0.007096066605299711, + 0.01798979379236698, + 0.00862869992852211, + 0.003905790625140071, + 0.0033063648734241724, + 0.004957365803420544, + 0.005021892488002777, + 0.005997602827847004, + 0.005265951156616211, + 0.004664372652769089, + 0.004252190236002207, + 0.006912139244377613, + 0.006741876248270273, + 0.0057804519310593605, + 0.008490892127156258, + 0.007379875984042883, + 0.0064820535480976105, + 0.005514796357601881, + 0.004445262253284454, + 0.02208038605749607, + 0.016548754647374153, + 0.06096639111638069, + 0.1041819155216217, + 0.011108391918241978, + 0.00760095939040184, + 0.052278365939855576, + 0.012038524262607098, + 0.0031349852215498686, + 0.0038281050510704517, + 0.004212545230984688, + 0.0047012679278850555, + 0.003899039002135396, + 0.004559369757771492, + 0.003948137164115906, + 0.002034244127571583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.007631197106093168 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795775189995766 + ], + "opposite_logprob_tokens": [ + -3.5297958850860596 + ], + "logprob_reward": -0.029795775189995766, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795775189995766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04548660293221474, + 0.03440643846988678, + 0.24934880435466766, + 0.12658385932445526, + 0.12899024784564972, + 0.07108870148658752, + 0.024628857150673866, + 0.031270164996385574, + 0.2492312490940094, + 0.038965024054050446 + ], + "credit_map": [ + 0.004841384943574667, + 0.005860685836523771, + 0.005002578254789114, + 0.007068172562867403, + 0.006389419548213482, + 0.0060776714235544205, + 0.005144752096384764, + 0.005101940128952265, + 0.003027112688869238, + 0.004528651013970375, + 0.005991140380501747, + 0.004554665647447109, + 0.003913714084774256, + 0.0048383865505456924, + 0.003387845354154706, + 0.004164926242083311, + 0.04373927786946297, + 0.018021048977971077, + 0.013285810127854347, + 0.09794095903635025, + 0.00863208994269371, + 0.011317053809762001, + 0.016890909522771835, + 0.03952166438102722, + 0.029275495558977127, + 0.017361413687467575, + 0.008374877274036407, + 0.037344127893447876, + 0.011387918144464493, + 0.006429385393857956, + 0.008108139038085938, + 0.00830250233411789, + 0.007047086954116821, + 0.008347048424184322, + 0.005388956982642412, + 0.038400065153837204, + 0.010085664689540863, + 0.008258857764303684, + 0.045076001435518265, + 0.006386577617377043, + 0.005218123085796833, + 0.011250997893512249, + 0.007992404513061047, + 0.009095882065594196, + 0.012171879410743713, + 0.007823471911251545, + 0.012288312427699566, + 0.005247634835541248, + 0.0022820662707090378, + 0.004943774547427893, + 0.0032295051496475935, + 0.002969055902212858, + 0.004103574901819229, + 0.0021698635537177324, + 0.0026142369024455547, + 0.0023167806211858988, + 0.004209852311760187, + 0.003418158972635865, + 0.0030285874381661415, + 0.004830754827708006, + 0.005338982213288546, + 0.003621411742642522, + 0.003989579621702433, + 0.002832838101312518, + 0.019592298194766045, + 0.016316043213009834, + 0.05527213215827942, + 0.09082698076963425, + 0.009769359603524208, + 0.006466621067374945, + 0.03966067358851433, + 0.011327114887535572, + 0.0028786053881049156, + 0.00439783651381731, + 0.00815921276807785, + 0.004432162269949913, + 0.00409913482144475, + 0.002821884583681822, + 0.010181906633079052, + 0.001994279446080327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.029795775189995766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a medium shot, the camera moving forward incredibly slowly, creating an intimate feeling of glimpsing a private moment. In a bright, yet empty corner of a late-night convenience store, a young couple cuddles by a large floor-to-ceiling window. The man, wearing a loose, off-white hooded sweatshirt, sits sideways, his arm gently wrapped around the woman's shoulders. The woman, wearing an earth-toned knit cardigan, rests her full weight securely on him, her head resting gently on his shoulder, her hands cradling a steaming white paper cup. The fluorescent lighting inside the convenience store is deliberately rendered in a warm, vintage hue, suffusing the image with a soft golden glow and permeating the air with a nostalgic cinematic quality. Outside, the night is deep, only scattered streetlights and the faint trails of distant car taillights cast a stark contrast to the warm atmosphere inside. As the camera steadily and calmly zooms in, the picture gradually focuses on the faces of the two of them. They both have their eyes closed, with a subtle but very real smile on the corners of their mouths. Their cheeks appear particularly soft under the warm light, and every strand of hair is clearly visible, perfectly capturing the kind of contentment and tranquility found in ordinary moments that requires no words.", + "seed": 23180, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_3/index_2747_seed_23180_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09304708987474442, + "vlm_logprob": -0.09304708987474442, + "vlm_yes_no_margin": 3.8500000119209288, + "vlm_token_credit": -0.09304708987474442, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3709356188774109 + ], + "opposite_logprob_tokens": [ + -1.2459356784820557 + ], + "logprob_reward": -0.3709356188774109, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3709356188774109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10431807488203049, + 0.06478877365589142, + 0.053444840013980865, + 0.0665588304400444, + 0.246725395321846, + 0.178423210978508, + 0.13474512100219727, + 0.03390934690833092, + 0.07271655648946762, + 0.044369813054800034 + ], + "credit_map": [ + 0.014875275082886219, + 0.011483345180749893, + 0.01676628552377224, + 0.008380877785384655, + 0.016528790816664696, + 0.011249849572777748, + 0.012824330478906631, + 0.012209330685436726, + 0.012202030047774315, + 0.01041417010128498, + 0.009644108824431896, + 0.0053705088794231415, + 0.0059997751377522945, + 0.004735448397696018, + 0.005251510534435511, + 0.011171224527060986, + 0.008601951412856579, + 0.010163594968616962, + 0.006015142425894737, + 0.004575888626277447, + 0.008015741594135761, + 0.005272095091640949, + 0.0064091915264725685, + 0.0043912362307310104, + 0.01182989589869976, + 0.015331143513321877, + 0.010606746189296246, + 0.004878650885075331, + 0.007995270192623138, + 0.005244540981948376, + 0.00560690276324749, + 0.005065687000751495, + 0.030405886471271515, + 0.050760626792907715, + 0.045954324305057526, + 0.025542501360177994, + 0.029588546603918076, + 0.022586524486541748, + 0.01815442554652691, + 0.023732567206025124, + 0.026083629578351974, + 0.025933295488357544, + 0.025369126349687576, + 0.011822881177067757, + 0.023728374391794205, + 0.03125132620334625, + 0.01763973757624626, + 0.016594847664237022, + 0.027033977210521698, + 0.01500037219375372, + 0.01826665736734867, + 0.020895162597298622, + 0.01384416501969099, + 0.013643519021570683, + 0.014284525997936726, + 0.011776752769947052, + 0.006165777333080769, + 0.004350254777818918, + 0.0062688314355909824, + 0.0024392714258283377, + 0.0034431235399097204, + 0.0051490385085344315, + 0.002400758909061551, + 0.003692293306812644, + 0.013341836631298065, + 0.008924663998186588, + 0.01219269447028637, + 0.006866720039397478, + 0.00810836348682642, + 0.00849844142794609, + 0.008362283930182457, + 0.0064215580932796, + 0.0032845190726220608, + 0.005188995972275734, + 0.006206497550010681, + 0.005612988956272602, + 0.005057450849562883, + 0.005075214430689812, + 0.010214569047093391, + 0.0037295776419341564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.3709356188774109 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004127318039536476 + ], + "opposite_logprob_tokens": [ + -5.504127502441406 + ], + "logprob_reward": -0.004127318039536476, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004127318039536476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2491641789674759, + 0.1139475405216217, + 0.07552295178174973, + 0.04990033060312271, + 0.08912760019302368, + 0.09634185582399368, + 0.06705844402313232, + 0.10851163417100906, + 0.08223843574523926, + 0.06818701326847076 + ], + "credit_map": [ + 0.03761846572160721, + 0.02666535973548889, + 0.0392509400844574, + 0.019448066130280495, + 0.024436570703983307, + 0.021817795932292938, + 0.019901111721992493, + 0.06002587452530861, + 0.01423610094934702, + 0.020224962383508682, + 0.017353737726807594, + 0.016725260764360428, + 0.01587235927581787, + 0.008095834404230118, + 0.011256445199251175, + 0.01018284447491169, + 0.015868082642555237, + 0.011265953071415424, + 0.010270160622894764, + 0.008206857368350029, + 0.006541197653859854, + 0.008957541547715664, + 0.007380322087556124, + 0.0070328423753380775, + 0.008822433650493622, + 0.0060725086368620396, + 0.011499489657580853, + 0.005439106374979019, + 0.005086819641292095, + 0.004382592625916004, + 0.0032769562676548958, + 0.005320427939295769, + 0.011963801458477974, + 0.013822630047798157, + 0.01822015270590782, + 0.007882453501224518, + 0.012512527406215668, + 0.009186767041683197, + 0.007715865038335323, + 0.007823402993381023, + 0.01729067787528038, + 0.01017997320741415, + 0.01127585582435131, + 0.0067480988800525665, + 0.01636584848165512, + 0.012470368295907974, + 0.012112010270357132, + 0.009899020195007324, + 0.008458491414785385, + 0.008434216491878033, + 0.0069519393146038055, + 0.00783533975481987, + 0.010974608361721039, + 0.01087862253189087, + 0.009046916849911213, + 0.00447830930352211, + 0.018902884796261787, + 0.014562741853296757, + 0.024761440232396126, + 0.008756520226597786, + 0.013130799867212772, + 0.013491572812199593, + 0.007778088096529245, + 0.007127585355192423, + 0.016128405928611755, + 0.010679644532501698, + 0.008714132942259312, + 0.0070855263620615005, + 0.011028395034372807, + 0.010941767133772373, + 0.007375346031039953, + 0.010285211727023125, + 0.007328877225518227, + 0.00838138535618782, + 0.008426730521023273, + 0.006081531755626202, + 0.011748245917260647, + 0.009269065223634243, + 0.009491500444710255, + 0.00745967635884881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.004127318039536476 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525581045076251 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.00525581045076251, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525581045076251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09539730846881866, + 0.0385090634226799, + 0.10705377906560898, + 0.10323690623044968, + 0.17971041798591614, + 0.20740920305252075, + 0.11417374759912491, + 0.06850988417863846, + 0.053104255348443985, + 0.03289540484547615 + ], + "credit_map": [ + 0.018506359308958054, + 0.011996971443295479, + 0.016980014741420746, + 0.009566915221512318, + 0.012625936418771744, + 0.0073471651412546635, + 0.00880421418696642, + 0.009569735266268253, + 0.006508136633783579, + 0.00554933026432991, + 0.005961254704743624, + 0.004626072943210602, + 0.0042109242640435696, + 0.002367333509027958, + 0.004641595762223005, + 0.00464441766962409, + 0.01400329265743494, + 0.01859928108751774, + 0.010463621467351913, + 0.010172036476433277, + 0.012223578058183193, + 0.016457684338092804, + 0.012085876427590847, + 0.013048416003584862, + 0.015753762796521187, + 0.01325321290642023, + 0.022818520665168762, + 0.007674211170524359, + 0.00953841581940651, + 0.009244570508599281, + 0.008602883666753769, + 0.016351325437426567, + 0.021918928250670433, + 0.03569381311535835, + 0.032607242465019226, + 0.016907067969441414, + 0.025319676846265793, + 0.016855809837579727, + 0.013047445565462112, + 0.017360441386699677, + 0.025159886106848717, + 0.03424495458602905, + 0.025101102888584137, + 0.014736136421561241, + 0.029908955097198486, + 0.03545233607292175, + 0.02167896367609501, + 0.021126864477992058, + 0.020905259996652603, + 0.012067455798387527, + 0.01490087155252695, + 0.01848710887134075, + 0.013222865760326385, + 0.011665940284729004, + 0.012769773602485657, + 0.01015446987003088, + 0.004745970945805311, + 0.010621004737913609, + 0.0045706103555858135, + 0.003138526575639844, + 0.005849271081387997, + 0.021415747702121735, + 0.004202544689178467, + 0.013966209255158901, + 0.01004750281572342, + 0.0066275205463171005, + 0.007029679603874683, + 0.004493657499551773, + 0.007251868490129709, + 0.006566735450178385, + 0.005274868570268154, + 0.005812421906739473, + 0.003157560247927904, + 0.003839487209916115, + 0.004340129438787699, + 0.0033477654214948416, + 0.005714583210647106, + 0.004296695813536644, + 0.004378690849989653, + 0.003820493584498763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.00525581045076251 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950829479843378 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950829479843378, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950829479843378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10094382613897324, + 0.03859812766313553, + 0.09133758395910263, + 0.06405086070299149, + 0.1897224634885788, + 0.17931631207466125, + 0.14134548604488373, + 0.05474061146378517, + 0.09148558974266052, + 0.04845914617180824 + ], + "credit_map": [ + 0.017131401225924492, + 0.012802994810044765, + 0.017832918092608452, + 0.010568623431026936, + 0.011068325489759445, + 0.008361144922673702, + 0.012514770962297916, + 0.010663647204637527, + 0.006366129033267498, + 0.004969280678778887, + 0.00610691960901022, + 0.0038477634079754353, + 0.004276853986084461, + 0.002374187810346484, + 0.005133711267262697, + 0.005523279309272766, + 0.011368402279913425, + 0.022290479391813278, + 0.011479858309030533, + 0.010235946625471115, + 0.011786116287112236, + 0.00741860456764698, + 0.007553411647677422, + 0.009204762987792492, + 0.011187375523149967, + 0.01518213003873825, + 0.01009749062359333, + 0.006018329877406359, + 0.006691784132272005, + 0.004973829723894596, + 0.005131820682436228, + 0.004768101032823324, + 0.0226538497954607, + 0.035338740795850754, + 0.03663719445466995, + 0.017887482419610023, + 0.023790845647454262, + 0.015971621498465538, + 0.016226833686232567, + 0.021215895190835, + 0.02258959598839283, + 0.03199281543493271, + 0.023231100291013718, + 0.014688543975353241, + 0.0234611164778471, + 0.029490076005458832, + 0.016223445534706116, + 0.01763962395489216, + 0.02615422196686268, + 0.013899403624236584, + 0.016830863431096077, + 0.020809821784496307, + 0.0188461784273386, + 0.017399083822965622, + 0.016107454895973206, + 0.011298461817204952, + 0.005074115004390478, + 0.007678569760173559, + 0.0054865749552845955, + 0.002647185232490301, + 0.0070943268947303295, + 0.01432938501238823, + 0.00449750479310751, + 0.007932950742542744, + 0.018100840970873833, + 0.012416373006999493, + 0.010554284788668156, + 0.00741332583129406, + 0.011823045089840889, + 0.012670043855905533, + 0.008531734347343445, + 0.009975943714380264, + 0.0056177666410803795, + 0.006381275597959757, + 0.0059859841130673885, + 0.004332156386226416, + 0.00640765018761158, + 0.008173754438757896, + 0.006501135416328907, + 0.005059425253421068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.005950829479843378 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896587252616882 + ], + "opposite_logprob_tokens": [ + -2.578965902328491 + ], + "logprob_reward": -0.07896587252616882, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896587252616882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13655045628547668, + 0.03291456028819084, + 0.04758079722523689, + 0.0667979046702385, + 0.11088839918375015, + 0.2011542171239853, + 0.11651071161031723, + 0.16169953346252441, + 0.04414582997560501, + 0.08175758272409439 + ], + "credit_map": [ + 0.020195534452795982, + 0.01577598601579666, + 0.016089199110865593, + 0.012672023847699165, + 0.01961963064968586, + 0.015499205328524113, + 0.016467807814478874, + 0.020231062546372414, + 0.004758225753903389, + 0.0030131870880723, + 0.0046510412357747555, + 0.004507478792220354, + 0.005034129600971937, + 0.0017564496956765652, + 0.005311618093401194, + 0.0038824293296784163, + 0.0051760138012468815, + 0.009648503735661507, + 0.005659257527440786, + 0.004506360739469528, + 0.006927944719791412, + 0.003988292533904314, + 0.0066944002173841, + 0.004980022087693214, + 0.010977193713188171, + 0.01652541384100914, + 0.009054983034729958, + 0.005468685179948807, + 0.008527758531272411, + 0.005689789075404406, + 0.005668994504958391, + 0.004885086789727211, + 0.015452248044312, + 0.01734674721956253, + 0.018959736451506615, + 0.010779794305562973, + 0.013599113561213017, + 0.00885491631925106, + 0.011775351129472256, + 0.01412049587816, + 0.021068105474114418, + 0.042481161653995514, + 0.025125958025455475, + 0.01486901380121708, + 0.021962881088256836, + 0.03796633705496788, + 0.01917799562215805, + 0.018502764403820038, + 0.02127045951783657, + 0.014145190827548504, + 0.01430925540626049, + 0.016601471230387688, + 0.014246371574699879, + 0.013090255670249462, + 0.012696286663413048, + 0.010151421651244164, + 0.009525769390165806, + 0.025984691455960274, + 0.006595994345843792, + 0.003621735842898488, + 0.007541610859334469, + 0.05990049988031387, + 0.007410516031086445, + 0.041118718683719635, + 0.005923401564359665, + 0.005190331023186445, + 0.005701633170247078, + 0.003382299793884158, + 0.006543670780956745, + 0.005128677003085613, + 0.003952790517359972, + 0.008323021233081818, + 0.006465717684477568, + 0.009504695422947407, + 0.011659961193799973, + 0.010323704220354557, + 0.009127761237323284, + 0.010549072176218033, + 0.01389208436012268, + 0.0102345896884799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.07896587252616882 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a medium shot, the camera moving forward incredibly slowly, creating an intimate feeling of glimpsing a private moment. In a bright, yet empty corner of a late-night convenience store, a young couple cuddles by a large floor-to-ceiling window. The man, wearing a loose, off-white hooded sweatshirt, sits sideways, his arm gently wrapped around the woman's shoulders. The woman, wearing an earth-toned knit cardigan, rests her full weight securely on him, her head resting gently on his shoulder, her hands cradling a steaming white paper cup. The fluorescent lighting inside the convenience store is deliberately rendered in a warm, vintage hue, suffusing the image with a soft golden glow and permeating the air with a nostalgic cinematic quality. Outside, the night is deep, only scattered streetlights and the faint trails of distant car taillights cast a stark contrast to the warm atmosphere inside. As the camera steadily and calmly zooms in, the picture gradually focuses on the faces of the two of them. They both have their eyes closed, with a subtle but very real smile on the corners of their mouths. Their cheeks appear particularly soft under the warm light, and every strand of hair is clearly visible, perfectly capturing the kind of contentment and tranquility found in ordinary moments that requires no words.", + "seed": 23181, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_3/index_2747_seed_23181_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03661999423056841, + "vlm_logprob": -0.03661999423056841, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.03661999423056841, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14060381054878235 + ], + "opposite_logprob_tokens": [ + -2.14060378074646 + ], + "logprob_reward": -0.14060381054878235, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14060381054878235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32261937856674194, + 0.07785886526107788, + 0.056188516318798065, + 0.07174631953239441, + 0.05126847326755524, + 0.08534141629934311, + 0.07391012459993362, + 0.08084361255168915, + 0.04335740953683853, + 0.13686583936214447 + ], + "credit_map": [ + 0.0563080720603466, + 0.03331824019551277, + 0.040264345705509186, + 0.03889736533164978, + 0.02615465596318245, + 0.06305433064699173, + 0.031226519495248795, + 0.03339586779475212, + 0.009870695881545544, + 0.009121288545429707, + 0.009603057987987995, + 0.007451304234564304, + 0.007743432652205229, + 0.010997993871569633, + 0.01094069518148899, + 0.012130393646657467, + 0.006095483433455229, + 0.005385975819081068, + 0.013341298326849937, + 0.00592874875292182, + 0.006778672803193331, + 0.006696013268083334, + 0.004714830778539181, + 0.007247492671012878, + 0.009746767580509186, + 0.008119242265820503, + 0.008682834915816784, + 0.010103668086230755, + 0.008562419563531876, + 0.009747722186148167, + 0.007722148206084967, + 0.009061514399945736, + 0.007558836601674557, + 0.0054136947728693485, + 0.005757421255111694, + 0.005785579793155193, + 0.0070938365533947945, + 0.006897467654198408, + 0.007355601992458105, + 0.0054060365073382854, + 0.012434864416718483, + 0.008211110718548298, + 0.008695661090314388, + 0.011540442705154419, + 0.009533076547086239, + 0.015495647676289082, + 0.01079437043517828, + 0.00863624643534422, + 0.012102080509066582, + 0.006611226592212915, + 0.00883377343416214, + 0.010019884444773197, + 0.010765518061816692, + 0.01027572713792324, + 0.008232091553509235, + 0.0070698196068406105, + 0.013122676871716976, + 0.008470390923321247, + 0.010168532840907574, + 0.01520570833235979, + 0.006070749368518591, + 0.010264476761221886, + 0.007884821854531765, + 0.009656251408159733, + 0.0046498761512339115, + 0.004525660537183285, + 0.004339363891631365, + 0.0076242974027991295, + 0.006676004268229008, + 0.0038362909108400345, + 0.006574498489499092, + 0.005131418816745281, + 0.015645312145352364, + 0.014061914756894112, + 0.013907788321375847, + 0.026988735422492027, + 0.02051457390189171, + 0.012740282341837883, + 0.021809201687574387, + 0.011198014952242374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.14060381054878235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019657351076602936 + ], + "opposite_logprob_tokens": [ + -6.251965522766113 + ], + "logprob_reward": -0.0019657351076602936, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019657351076602936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24449817836284637, + 0.0788092091679573, + 0.1005418598651886, + 0.08940187096595764, + 0.04522731900215149, + 0.03927941992878914, + 0.030499573796987534, + 0.1177864670753479, + 0.09087767452001572, + 0.1630783975124359 + ], + "credit_map": [ + 0.04247557744383812, + 0.025816714391112328, + 0.03199335187673569, + 0.03294726833701134, + 0.028043175116181374, + 0.03771258145570755, + 0.021947231143712997, + 0.023562291637063026, + 0.007931211963295937, + 0.009076320566236973, + 0.011906368657946587, + 0.008126663975417614, + 0.009349687024950981, + 0.013791931793093681, + 0.00844427477568388, + 0.01018274761736393, + 0.012432192452251911, + 0.01445037592202425, + 0.026317013427615166, + 0.011141100898385048, + 0.008605609647929668, + 0.01021352969110012, + 0.007935741916298866, + 0.009446300566196442, + 0.01236832793802023, + 0.005103746894747019, + 0.023125024512410164, + 0.020582735538482666, + 0.0063310000114142895, + 0.008072789758443832, + 0.006482549011707306, + 0.0073356893844902515, + 0.006534737069159746, + 0.006860790774226189, + 0.005110138561576605, + 0.006909494288265705, + 0.00566414138302207, + 0.004559612367302179, + 0.005275893956422806, + 0.004312510602176189, + 0.0054526561871171, + 0.004388989415019751, + 0.004485190846025944, + 0.005374313332140446, + 0.0054367040283977985, + 0.005507003981620073, + 0.004856662359088659, + 0.003777896985411644, + 0.005428206641227007, + 0.002914795186370611, + 0.004034953657537699, + 0.004603968001902103, + 0.003879366908222437, + 0.003439673688262701, + 0.0031403640750795603, + 0.0030582428444176912, + 0.016710378229618073, + 0.010221934877336025, + 0.015916625037789345, + 0.022076910361647606, + 0.008452163077890873, + 0.01823822595179081, + 0.012811376713216305, + 0.013358847238123417, + 0.009101788513362408, + 0.011492302641272545, + 0.010280889458954334, + 0.014298925176262856, + 0.015796922147274017, + 0.007660129573196173, + 0.013438878580927849, + 0.008807839825749397, + 0.023857425898313522, + 0.01966908387839794, + 0.01770009845495224, + 0.023223374038934708, + 0.02788541465997696, + 0.014433794654905796, + 0.01707741804420948, + 0.019231781363487244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0019657351076602936 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002203657291829586 + ], + "opposite_logprob_tokens": [ + -6.127203464508057 + ], + "logprob_reward": -0.002203657291829586, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002203657291829586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09562846273183823, + 0.050798747688531876, + 0.05233457311987877, + 0.05636447295546532, + 0.133576899766922, + 0.3042827546596527, + 0.013681468553841114, + 0.050704486668109894, + 0.0447988323867321, + 0.19782935082912445 + ], + "credit_map": [ + 0.013008004054427147, + 0.011630501598119736, + 0.011056767776608467, + 0.01513879094272852, + 0.010910303331911564, + 0.009686597622931004, + 0.010589318349957466, + 0.0136081762611866, + 0.006184649188071489, + 0.005748120602220297, + 0.006374978926032782, + 0.004580676555633545, + 0.007264244370162487, + 0.006717615760862827, + 0.006513215601444244, + 0.00741524389013648, + 0.005454810801893473, + 0.006912550423294306, + 0.008991586044430733, + 0.005323899444192648, + 0.007256712298840284, + 0.004530994687229395, + 0.00579293305054307, + 0.00807107798755169, + 0.012063149362802505, + 0.0055689215660095215, + 0.005773145705461502, + 0.0077772545628249645, + 0.006066196132451296, + 0.006585434544831514, + 0.006189570762217045, + 0.0063407947309315205, + 0.010204528458416462, + 0.014514503069221973, + 0.016665928065776825, + 0.032271966338157654, + 0.021514898166060448, + 0.01031566597521305, + 0.015895968303084373, + 0.012193429283797741, + 0.028745753690600395, + 0.044607650488615036, + 0.03685566037893295, + 0.049931302666664124, + 0.04256424307823181, + 0.025570783764123917, + 0.03595677763223648, + 0.04005058482289314, + 0.0014407761627808213, + 0.001306435908190906, + 0.0018050571670755744, + 0.002349252812564373, + 0.002548745134845376, + 0.0012467221822589636, + 0.001525814295746386, + 0.0014586630277335644, + 0.004878131672739983, + 0.004707001149654388, + 0.006386070046573877, + 0.009174235165119171, + 0.008047632873058319, + 0.00414947047829628, + 0.00558087695389986, + 0.007781060878187418, + 0.004539239685982466, + 0.004567248281091452, + 0.004912278614938259, + 0.006585416384041309, + 0.006433042231947184, + 0.004806047305464745, + 0.007646290585398674, + 0.0053092665039002895, + 0.02617678791284561, + 0.022621693089604378, + 0.021042240783572197, + 0.030260972678661346, + 0.03107057325541973, + 0.019367052242159843, + 0.023491812869906425, + 0.023798199370503426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.002203657291829586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464111752808094 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.00464111752808094, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464111752808094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12269163131713867, + 0.049476951360702515, + 0.08168759196996689, + 0.10514604300260544, + 0.10384848713874817, + 0.06175851449370384, + 0.06050863489508629, + 0.08881105482578278, + 0.048771508038043976, + 0.2772996127605438 + ], + "credit_map": [ + 0.015784379094839096, + 0.01478871051222086, + 0.01212936732918024, + 0.018576787784695625, + 0.015177648514509201, + 0.013540309853851795, + 0.014338596723973751, + 0.018355829641222954, + 0.0059171998873353004, + 0.0072959549725055695, + 0.006480609532445669, + 0.004836712032556534, + 0.007301232777535915, + 0.004189521074295044, + 0.005942236632108688, + 0.007513486314564943, + 0.008412324823439121, + 0.010233747772872448, + 0.011326398700475693, + 0.0075606987811625, + 0.011736011132597923, + 0.00813036598265171, + 0.00940166786313057, + 0.014886378310620785, + 0.020343201234936714, + 0.0064384606666862965, + 0.01486134808510542, + 0.012618516571819782, + 0.012676059268414974, + 0.010622391477227211, + 0.01574624516069889, + 0.011839822866022587, + 0.009999881498515606, + 0.0136648491024971, + 0.012926933355629444, + 0.02474176697432995, + 0.014173806644976139, + 0.008166144602000713, + 0.012043452821671963, + 0.008131648413836956, + 0.007128268480300903, + 0.008168341591954231, + 0.0070486306212842464, + 0.00934059638530016, + 0.011690614745020866, + 0.00578473974019289, + 0.0069294823333621025, + 0.005667840130627155, + 0.011699315160512924, + 0.006305045448243618, + 0.007904281839728355, + 0.007903371006250381, + 0.005870120134204626, + 0.009368409402668476, + 0.006110482849180698, + 0.005347609985619783, + 0.007666765712201595, + 0.007601735182106495, + 0.013183625414967537, + 0.02176591567695141, + 0.012371339835226536, + 0.006958180572837591, + 0.009247496724128723, + 0.010015987791121006, + 0.005587630905210972, + 0.004363867454230785, + 0.006329639349132776, + 0.008466353639960289, + 0.006296851206570864, + 0.005129830911755562, + 0.007234642747789621, + 0.005362692754715681, + 0.0436384379863739, + 0.03336524963378906, + 0.03178083151578903, + 0.037840794771909714, + 0.041949205100536346, + 0.030601315200328827, + 0.02889743074774742, + 0.029226355254650116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.00464111752808094 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033685650676488876 + ], + "opposite_logprob_tokens": [ + -3.4086856842041016 + ], + "logprob_reward": -0.033685650676488876, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033685650676488876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13407985866069794, + 0.04138780012726784, + 0.07849667221307755, + 0.05389014631509781, + 0.04523152485489845, + 0.03344249352812767, + 0.044120095670223236, + 0.0774078369140625, + 0.06722056865692139, + 0.4247229993343353 + ], + "credit_map": [ + 0.03184821084141731, + 0.014470187947154045, + 0.01913277991116047, + 0.018137123435735703, + 0.010108194313943386, + 0.014476001262664795, + 0.010919363237917423, + 0.014987997710704803, + 0.004728035070002079, + 0.005038542672991753, + 0.0065603177063167095, + 0.004721456207334995, + 0.004867644980549812, + 0.005119152367115021, + 0.003756894962862134, + 0.006595754995942116, + 0.007855480536818504, + 0.006788692437112331, + 0.01227501779794693, + 0.007853054441511631, + 0.011831143870949745, + 0.0077215032652020454, + 0.010948685929179192, + 0.013223100453615189, + 0.008396809920668602, + 0.005088122561573982, + 0.005874163005501032, + 0.00712500000372529, + 0.0067120809108018875, + 0.0072476984933018684, + 0.0062848846428096294, + 0.007161385379731655, + 0.006038867402821779, + 0.005050139967352152, + 0.00416704872623086, + 0.006304038688540459, + 0.006268000695854425, + 0.00583348423242569, + 0.0065500265918672085, + 0.005019919481128454, + 0.0032404784578830004, + 0.005045037250965834, + 0.003776367986574769, + 0.005503891501575708, + 0.005567142739892006, + 0.002604920417070389, + 0.00401897681877017, + 0.003685677656903863, + 0.0051765479147434235, + 0.004316565115004778, + 0.006138301454484463, + 0.009020301513373852, + 0.00562440138310194, + 0.004241722635924816, + 0.005497796460986137, + 0.004104458261281252, + 0.008446712046861649, + 0.007912727072834969, + 0.009227698668837547, + 0.014064418151974678, + 0.012506552040576935, + 0.008157273754477501, + 0.009079095907509327, + 0.008013361133635044, + 0.006698968820273876, + 0.006938242819160223, + 0.006892162840813398, + 0.008706958033144474, + 0.011429421603679657, + 0.007616291753947735, + 0.010709170252084732, + 0.008229351602494717, + 0.05926923826336861, + 0.043748918920755386, + 0.04701056331396103, + 0.07906045019626617, + 0.06220315024256706, + 0.03929917886853218, + 0.047785576432943344, + 0.046345941722393036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.033685650676488876 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a medium shot, the camera moving forward incredibly slowly, creating an intimate feeling of glimpsing a private moment. In a bright, yet empty corner of a late-night convenience store, a young couple cuddles by a large floor-to-ceiling window. The man, wearing a loose, off-white hooded sweatshirt, sits sideways, his arm gently wrapped around the woman's shoulders. The woman, wearing an earth-toned knit cardigan, rests her full weight securely on him, her head resting gently on his shoulder, her hands cradling a steaming white paper cup. The fluorescent lighting inside the convenience store is deliberately rendered in a warm, vintage hue, suffusing the image with a soft golden glow and permeating the air with a nostalgic cinematic quality. Outside, the night is deep, only scattered streetlights and the faint trails of distant car taillights cast a stark contrast to the warm atmosphere inside. As the camera steadily and calmly zooms in, the picture gradually focuses on the faces of the two of them. They both have their eyes closed, with a subtle but very real smile on the corners of their mouths. Their cheeks appear particularly soft under the warm light, and every strand of hair is clearly visible, perfectly capturing the kind of contentment and tranquility found in ordinary moments that requires no words.", + "seed": 23182, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_3/index_2747_seed_23182_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.024432125594466926, + "vlm_logprob": -0.024432125594466926, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.024432125594466926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09484246373176575 + ], + "opposite_logprob_tokens": [ + -2.4698424339294434 + ], + "logprob_reward": -0.09484246373176575, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09484246373176575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4078100919723511, + 0.16841302812099457, + 0.04104986786842346, + 0.04770559072494507, + 0.028002772480249405, + 0.04946080967783928, + 0.15601746737957, + 0.03919695317745209, + 0.03813875466585159, + 0.024204693734645844 + ], + "credit_map": [ + 0.11516576260328293, + 0.031735606491565704, + 0.14374609291553497, + 0.04325404390692711, + 0.018125249072909355, + 0.013387972488999367, + 0.019654011353850365, + 0.022741371765732765, + 0.034026023000478745, + 0.019330112263560295, + 0.023125257343053818, + 0.024996638298034668, + 0.01612783782184124, + 0.019725251942873, + 0.018760614097118378, + 0.012321283109486103, + 0.0044153472408652306, + 0.008610229939222336, + 0.004446614999324083, + 0.00611898535862565, + 0.004786125849932432, + 0.0037690303288400173, + 0.005458668805658817, + 0.003444864647462964, + 0.006189069245010614, + 0.004871762357652187, + 0.003899992909282446, + 0.014575235545635223, + 0.0030192648991942406, + 0.0029950900934636593, + 0.0062386877834796906, + 0.0059164888225495815, + 0.003467248985543847, + 0.003279905067756772, + 0.0036208194214850664, + 0.004904812667518854, + 0.0027283201925456524, + 0.002795650390908122, + 0.003931994084268808, + 0.003274020040407777, + 0.00538878608494997, + 0.004081116523593664, + 0.0036841777618974447, + 0.005754711572080851, + 0.00883918721228838, + 0.0026976270601153374, + 0.013065031729638577, + 0.005950175691395998, + 0.013495584949851036, + 0.017043840140104294, + 0.03494422882795334, + 0.024246815592050552, + 0.02618345618247986, + 0.007402164861559868, + 0.0189250148832798, + 0.013776367530226707, + 0.004663681611418724, + 0.005454267840832472, + 0.002911678981035948, + 0.005683897528797388, + 0.005913752596825361, + 0.0029807600658386946, + 0.00785156525671482, + 0.0037373474333435297, + 0.004771149251610041, + 0.0028605242259800434, + 0.005396044347435236, + 0.007164777256548405, + 0.005391388200223446, + 0.004591670818626881, + 0.004570126999169588, + 0.00339307077229023, + 0.0032529428135603666, + 0.0022027408704161644, + 0.0025333096273243427, + 0.003239572048187256, + 0.0033042486757040024, + 0.0036960605066269636, + 0.0028869870584458113, + 0.003088832600042224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.09484246373176575 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007645274978131056 + ], + "opposite_logprob_tokens": [ + -4.882645130157471 + ], + "logprob_reward": -0.007645274978131056, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007645274978131056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4988917410373688, + 0.0900353416800499, + 0.03854949399828911, + 0.036699313670396805, + 0.02524334192276001, + 0.08634667098522186, + 0.0781143307685852, + 0.06723572313785553, + 0.016919776797294617, + 0.0619642436504364 + ], + "credit_map": [ + 0.1362428069114685, + 0.03834658861160278, + 0.18119949102401733, + 0.0495491661131382, + 0.020413076505064964, + 0.025817250832915306, + 0.02111823298037052, + 0.026205124333500862, + 0.012197135016322136, + 0.010998372919857502, + 0.012366141192615032, + 0.010135641321539879, + 0.011043052189052105, + 0.015943683683872223, + 0.009003016166388988, + 0.008348294533789158, + 0.004784079734236002, + 0.008867374621331692, + 0.0038735365960747004, + 0.006047626491636038, + 0.0032000530045479536, + 0.004818381275981665, + 0.004060958977788687, + 0.0028974823653697968, + 0.0056432862766087055, + 0.0032168698962777853, + 0.004592581186443567, + 0.004139098804444075, + 0.002846844494342804, + 0.0034902351908385754, + 0.004359270911663771, + 0.008411124348640442, + 0.0024259770289063454, + 0.002677104203030467, + 0.004113327711820602, + 0.004892014432698488, + 0.003042228752747178, + 0.001858348143287003, + 0.0033320742659270763, + 0.0029022672679275274, + 0.009047159925103188, + 0.009458035230636597, + 0.00832128245383501, + 0.013669579289853573, + 0.013868040405213833, + 0.004609641619026661, + 0.017805131152272224, + 0.009567802771925926, + 0.006675601005554199, + 0.008086916990578175, + 0.012468401342630386, + 0.012162202969193459, + 0.015229186043143272, + 0.005673969630151987, + 0.010193598456680775, + 0.007624453399330378, + 0.006803177762776613, + 0.005873167887330055, + 0.005642223171889782, + 0.010396524332463741, + 0.011732766404747963, + 0.0064653982408344746, + 0.01370368804782629, + 0.0066187758930027485, + 0.0021522969473153353, + 0.0015576607547700405, + 0.00213266396895051, + 0.002898224163800478, + 0.002535096602514386, + 0.0017267748480662704, + 0.0022634421475231647, + 0.0016536166658625007, + 0.004867487587034702, + 0.00750100240111351, + 0.006724284961819649, + 0.012715616263449192, + 0.00864461064338684, + 0.006626158952713013, + 0.007883773185312748, + 0.007001310586929321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.007645274978131056 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639693535864353 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639693535864353, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639693535864353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21187244355678558, + 0.03771793842315674, + 0.07447459548711777, + 0.04372952878475189, + 0.14589615166187286, + 0.20281653106212616, + 0.14316461980342865, + 0.0385456345975399, + 0.04569341987371445, + 0.056089136749506 + ], + "credit_map": [ + 0.05782787874341011, + 0.013038945384323597, + 0.0768878385424614, + 0.021375184878706932, + 0.010293104685842991, + 0.008277212269604206, + 0.012784495018422604, + 0.011387784965336323, + 0.0057952310889959335, + 0.009049461223185062, + 0.0031815222464501858, + 0.004795859567821026, + 0.0041556707583367825, + 0.004585678223520517, + 0.003831774927675724, + 0.0023227373603731394, + 0.01227667834609747, + 0.013843399472534657, + 0.01565220020711422, + 0.00883736927062273, + 0.004296755883842707, + 0.005747707094997168, + 0.007979071699082851, + 0.005841411184519529, + 0.0043971240520477295, + 0.0052711497992277145, + 0.004759922157973051, + 0.0056172944605350494, + 0.0067692906595766544, + 0.0047931428998708725, + 0.007352471351623535, + 0.004769132938235998, + 0.013497216627001762, + 0.013243981637060642, + 0.02566845528781414, + 0.02506893500685692, + 0.012901282869279385, + 0.009508689865469933, + 0.025510909035801888, + 0.020496679469943047, + 0.02166115865111351, + 0.023757632821798325, + 0.019881462678313255, + 0.02410978637635708, + 0.03277729079127312, + 0.014308388344943523, + 0.04383930563926697, + 0.022481510415673256, + 0.014491825364530087, + 0.01642243005335331, + 0.02376200258731842, + 0.02146131917834282, + 0.02577468380331993, + 0.01022585853934288, + 0.017731567844748497, + 0.013294930569827557, + 0.005978780332952738, + 0.005105782300233841, + 0.004196464084088802, + 0.003926718141883612, + 0.004378743004053831, + 0.005466387141495943, + 0.005570015870034695, + 0.003922746516764164, + 0.0054512033239007, + 0.004527376499027014, + 0.005540214944630861, + 0.006672192364931107, + 0.006510080769658089, + 0.004917484242469072, + 0.007125148084014654, + 0.004949720576405525, + 0.0041508725844323635, + 0.006303955335170031, + 0.00574770150706172, + 0.012558097951114178, + 0.007168366573750973, + 0.005455879494547844, + 0.00732222618535161, + 0.007382035721093416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.004639693535864353 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253794603049755 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.005253794603049755, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253794603049755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27391862869262695, + 0.027661548927426338, + 0.09428665786981583, + 0.06574023514986038, + 0.0661831796169281, + 0.15510299801826477, + 0.1463852822780609, + 0.05042005702853203, + 0.061260100454092026, + 0.059041284024715424 + ], + "credit_map": [ + 0.07251684367656708, + 0.01871742494404316, + 0.0886741504073143, + 0.027682818472385406, + 0.013443365693092346, + 0.018790462985634804, + 0.01737925037741661, + 0.0167143102735281, + 0.003667964367195964, + 0.004060191567987204, + 0.0030672515276819468, + 0.006098627112805843, + 0.002706721890717745, + 0.002961626509204507, + 0.003049611346796155, + 0.0020495548378676176, + 0.009455215185880661, + 0.012105834670364857, + 0.014978481456637383, + 0.015212766826152802, + 0.010498739778995514, + 0.007133961655199528, + 0.016424203291535378, + 0.008477459661662579, + 0.005617512855678797, + 0.007636666763573885, + 0.007511514704674482, + 0.007952755317091942, + 0.010387593880295753, + 0.006917169317603111, + 0.012320353649556637, + 0.007396664470434189, + 0.00639284448698163, + 0.00687342369928956, + 0.011516135185956955, + 0.009321528486907482, + 0.0066442289389669895, + 0.005612637382000685, + 0.011556685902178288, + 0.008265690878033638, + 0.02031082473695278, + 0.021172434091567993, + 0.013643892481923103, + 0.017804212868213654, + 0.023340102285146713, + 0.008129224181175232, + 0.033834200352430344, + 0.016868101432919502, + 0.015719395130872726, + 0.016383111476898193, + 0.026594819501042366, + 0.023682022467255592, + 0.024455131962895393, + 0.010207838378846645, + 0.016650578007102013, + 0.012692374177277088, + 0.006562661379575729, + 0.005899070296436548, + 0.005570191890001297, + 0.006555076688528061, + 0.007148290518671274, + 0.006710564251989126, + 0.007346715312451124, + 0.004627484828233719, + 0.008180206641554832, + 0.0069313510321080685, + 0.008260510861873627, + 0.010304651223123074, + 0.006832617800682783, + 0.006879063323140144, + 0.007500505540519953, + 0.0063711958937346935, + 0.005093988962471485, + 0.006227408070117235, + 0.005841909907758236, + 0.013514295220375061, + 0.007022507954388857, + 0.007243241649121046, + 0.007074942346662283, + 0.0070229885168373585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.005253794603049755 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bright, empty corner of a late-night convenience store?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young couple cuddling by a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a loose, off-white hooded sweatshirt with his arm around the woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing an earth-toned knit cardigan while holding a steaming white paper cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a medium shot and steadily zoom in on the couple's faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779401123523712 + ], + "opposite_logprob_tokens": [ + -4.634779453277588 + ], + "logprob_reward": -0.009779401123523712, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779401123523712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.203435018658638, + 0.044105611741542816, + 0.06290316581726074, + 0.04883566126227379, + 0.055542558431625366, + 0.12797176837921143, + 0.3949930965900421, + 0.021240420639514923, + 0.019725702702999115, + 0.021246975287795067 + ], + "credit_map": [ + 0.0491536408662796, + 0.019187049940228462, + 0.05751962587237358, + 0.02255799062550068, + 0.011891685426235199, + 0.018370499834418297, + 0.011169398203492165, + 0.013585127890110016, + 0.008046447299420834, + 0.005430066026747227, + 0.006576960440725088, + 0.007856437005102634, + 0.005490279756486416, + 0.003221170511096716, + 0.00422004284337163, + 0.0032642041333019733, + 0.013113519176840782, + 0.0055744280107319355, + 0.01596449874341488, + 0.008114936761558056, + 0.005481740925461054, + 0.0040123434737324715, + 0.007237080484628677, + 0.0034046184737235308, + 0.0068063936196267605, + 0.005195583216845989, + 0.006028768606483936, + 0.0055544511415064335, + 0.004953000694513321, + 0.00482361251488328, + 0.007149333134293556, + 0.008324519731104374, + 0.0074864705093204975, + 0.010108628310263157, + 0.007966185919940472, + 0.006550767924636602, + 0.004756794311106205, + 0.003874514950439334, + 0.006832891143858433, + 0.007966306991875172, + 0.0164199098944664, + 0.015896225348114967, + 0.010660897009074688, + 0.01690368354320526, + 0.018494905903935432, + 0.006385242100805044, + 0.028598938137292862, + 0.014611978083848953, + 0.039969366043806076, + 0.04564917832612991, + 0.07622823864221573, + 0.05785985291004181, + 0.06823857873678207, + 0.025095026940107346, + 0.04713333025574684, + 0.03481949865818024, + 0.002596526872366667, + 0.0020992490462958813, + 0.00247656530700624, + 0.0031018781010061502, + 0.0033326309639960527, + 0.001856352435424924, + 0.0038916519843041897, + 0.0018855644157156348, + 0.0018708792049437761, + 0.0017315472941845655, + 0.0026948566082865, + 0.004455952439457178, + 0.0025394344702363014, + 0.0016796920681372285, + 0.002907741582021117, + 0.0018455978715792298, + 0.001798679935745895, + 0.002232662634924054, + 0.0020885805133730173, + 0.004340546205639839, + 0.002883169800043106, + 0.0023083463311195374, + 0.0027427636086940765, + 0.0028522240463644266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.009779401123523712 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_4/index_tensor([7981, 7981])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_4/index_tensor([7981, 7981])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0c7a1ba3435f158807e7136f35497222e1d50a59 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_4/index_tensor([7981, 7981])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static low-angle shot captures a slow-motion territorial battle between two ornate, spiky brown shells, their forms locked in a silent yet powerful struggle. The scene is rendered in a gritty, industrial style, dominated by harsh directional lighting, creating a dramatic black-and-white contrast. Resembling calcareous, gnarled inkstones, these shells are an intricate tapestry of tan and black cynar, their surfaces covered in intricate growth lines and sharp, dangerous spines that are currently interlocked. They push against each other with tremendous, grinding force, a motion that is almost imperceptible yet conveys incredible pressure. They rest on a dark, metallic coral reef, an unnatural surface of pitted, corroded gunmetal gray that gleams and shimmers with oil in the intense light. A powerful light source, directed from high above and to the right, cuts through the frame, casting one side of the combatants in a stark, clinical glow that accentuates every rough texture while plunging the other into deep, impenetrable shadow. The color palette is cool and desaturated, creating a tense, brutal atmosphere that transforms this natural conflict into a raw, mechanical, and unforgiving duel.", + "seed": 18884, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_4/index_7981_seed_18884_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0759208682924509, + "vlm_logprob": -0.0759208682924509, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.0759208682924509, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009243195876479149 + ], + "opposite_logprob_tokens": [ + -5.009243011474609 + ], + "logprob_reward": -0.009243195876479149, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009243195876479149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17504556477069855, + 0.0858219787478447, + 0.09209882467985153, + 0.0616757795214653, + 0.059566546231508255, + 0.06408345699310303, + 0.08383266627788544, + 0.08875308185815811, + 0.24518872797489166, + 0.043933331966400146 + ], + "credit_map": [ + 0.022322693839669228, + 0.013869783841073513, + 0.017034800723195076, + 0.02062872424721718, + 0.019874026998877525, + 0.024599920958280563, + 0.01536790281534195, + 0.04134770855307579, + 0.008786790072917938, + 0.006565547548234463, + 0.01181095652282238, + 0.013147452846169472, + 0.007689368445426226, + 0.011182674206793308, + 0.016832515597343445, + 0.009806671179831028, + 0.007830299437046051, + 0.008124098181724548, + 0.010515510104596615, + 0.015438930131494999, + 0.007955662906169891, + 0.010941527783870697, + 0.022049888968467712, + 0.009242910891771317, + 0.012805464677512646, + 0.004151776432991028, + 0.006938409060239792, + 0.006503662094473839, + 0.007116864435374737, + 0.00803415011614561, + 0.008836563676595688, + 0.007288890890777111, + 0.008386296220123768, + 0.005386214703321457, + 0.009908601641654968, + 0.008356720209121704, + 0.004343152977526188, + 0.00800265185534954, + 0.008006634190678596, + 0.007176271639764309, + 0.006697377655655146, + 0.005353669635951519, + 0.006051904987543821, + 0.007745569571852684, + 0.009673038497567177, + 0.008717783726751804, + 0.009683599695563316, + 0.010160510428249836, + 0.008245829492807388, + 0.009322828613221645, + 0.010002863593399525, + 0.01771644316613674, + 0.006600429769605398, + 0.012926151975989342, + 0.010956975631415844, + 0.008061149157583714, + 0.008523151278495789, + 0.011572908610105515, + 0.009789173491299152, + 0.014726918190717697, + 0.010086709633469582, + 0.0126853808760643, + 0.008678697980940342, + 0.012690143659710884, + 0.02412043698132038, + 0.0267361868172884, + 0.04958563297986984, + 0.03585565835237503, + 0.024455850943922997, + 0.03426740691065788, + 0.027769774198532104, + 0.022397775202989578, + 0.0035047910641878843, + 0.005187374539673328, + 0.005957505665719509, + 0.005253111012279987, + 0.008328082971274853, + 0.0058464366011321545, + 0.0048577082343399525, + 0.004998323041945696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.009243195876479149 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904295414686203 + ], + "opposite_logprob_tokens": [ + -2.464042901992798 + ], + "logprob_reward": -0.08904295414686203, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904295414686203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5047797560691833, + 0.059770748019218445, + 0.03629821911454201, + 0.04367383196949959, + 0.08055642247200012, + 0.02967759035527706, + 0.04721468687057495, + 0.02940484508872032, + 0.14415360987186432, + 0.024470213800668716 + ], + "credit_map": [ + 0.07227380573749542, + 0.048024147748947144, + 0.04271173104643822, + 0.06118737533688545, + 0.06878352165222168, + 0.0639273002743721, + 0.04773189499974251, + 0.10014002025127411, + 0.007873296737670898, + 0.005176975391805172, + 0.006613911595195532, + 0.007691720500588417, + 0.00740292202681303, + 0.008118107914924622, + 0.006494911853224039, + 0.010398903861641884, + 0.0036552969831973314, + 0.0024942404124885798, + 0.007637981791049242, + 0.005187464877963066, + 0.00298118544742465, + 0.005307899788022041, + 0.004600759129971266, + 0.004433393478393555, + 0.005724834278225899, + 0.0041748289950191975, + 0.005560232792049646, + 0.005936257541179657, + 0.005119104404002428, + 0.005488689057528973, + 0.005679177585989237, + 0.0059907096438109875, + 0.005904349032789469, + 0.009017436765134335, + 0.011426941491663456, + 0.013819274492561817, + 0.006661968305706978, + 0.01210226584225893, + 0.013241262175142765, + 0.008382923901081085, + 0.00382321048527956, + 0.003431023797020316, + 0.0031879935413599014, + 0.003156156511977315, + 0.003094577929005027, + 0.0038224169984459877, + 0.004719851538538933, + 0.004442363977432251, + 0.004036803729832172, + 0.0073782699182629585, + 0.005128161050379276, + 0.006061330437660217, + 0.005975381936877966, + 0.007566360756754875, + 0.00551919499412179, + 0.005549190565943718, + 0.003334412816911936, + 0.0034627222921699286, + 0.002870280761271715, + 0.004786361940205097, + 0.0031227916479110718, + 0.004108470864593983, + 0.0030158229637891054, + 0.004703984130173922, + 0.014433305710554123, + 0.024563707411289215, + 0.02299809828400612, + 0.01807042770087719, + 0.011582101695239544, + 0.02029854990541935, + 0.013387241400778294, + 0.018820196390151978, + 0.002549743978306651, + 0.00292754708789289, + 0.003770723706111312, + 0.0025840767193585634, + 0.003907963167876005, + 0.0034062203485518694, + 0.003074419917538762, + 0.002249522367492318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.08904295414686203 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057494968175888 + ], + "opposite_logprob_tokens": [ + -3.8955750465393066 + ], + "logprob_reward": -0.02057494968175888, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057494968175888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6254724860191345, + 0.04032234102487564, + 0.02337203547358513, + 0.030168110504746437, + 0.04607732966542244, + 0.016808578744530678, + 0.04457921162247658, + 0.07079248130321503, + 0.06938905268907547, + 0.03301834687590599 + ], + "credit_map": [ + 0.08957244455814362, + 0.06302224099636078, + 0.052679382264614105, + 0.07709244638681412, + 0.0866495743393898, + 0.07774835079908371, + 0.05836108699440956, + 0.12034698575735092, + 0.004212238360196352, + 0.0032738351728767157, + 0.004001299384981394, + 0.005652771331369877, + 0.003878944320604205, + 0.005537950433790684, + 0.0062358444556593895, + 0.007529453840106726, + 0.002738076727837324, + 0.0017613547388464212, + 0.004379906691610813, + 0.003720678389072418, + 0.0023038943763822317, + 0.0030542309395968914, + 0.0028236962389200926, + 0.0025901959743350744, + 0.003046509576961398, + 0.002438754541799426, + 0.0046158162876963615, + 0.00318545033223927, + 0.003346844343468547, + 0.003938295878469944, + 0.003327127080410719, + 0.006269312463700771, + 0.0049672359600663185, + 0.0035087044816464186, + 0.005699570290744305, + 0.007152169477194548, + 0.004053687676787376, + 0.005859880242496729, + 0.008646133355796337, + 0.006189944222569466, + 0.00225786748342216, + 0.0017699507297948003, + 0.0018629960250109434, + 0.0023418001364916563, + 0.002265641000121832, + 0.0023787457030266523, + 0.0018936280393972993, + 0.002037950325757265, + 0.0038312608376145363, + 0.0051050870679318905, + 0.005357752554118633, + 0.007273081690073013, + 0.0035041216760873795, + 0.006779448129236698, + 0.006567325908690691, + 0.006161136087030172, + 0.007130867801606655, + 0.005608821287751198, + 0.007016660179942846, + 0.007315918803215027, + 0.006649223621934652, + 0.010202714242041111, + 0.008818871341645718, + 0.01804940402507782, + 0.006769032683223486, + 0.00795537605881691, + 0.0069038537330925465, + 0.00771300308406353, + 0.0051344651728868484, + 0.01069426815956831, + 0.008354228921234608, + 0.015864823013544083, + 0.004326213151216507, + 0.0027672883588820696, + 0.003463245229795575, + 0.0034995542373508215, + 0.0039768763817846775, + 0.004111272748559713, + 0.004409631248563528, + 0.006464262492954731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.02057494968175888 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25210803747177124 + ], + "opposite_logprob_tokens": [ + -1.502108097076416 + ], + "logprob_reward": -0.25210803747177124, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25210803747177124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6159462928771973, + 0.02947399392724037, + 0.03273224085569382, + 0.02175368368625641, + 0.03522910177707672, + 0.0719008818268776, + 0.031253308057785034, + 0.050285786390304565, + 0.08558853715658188, + 0.025836210697889328 + ], + "credit_map": [ + 0.09465788304805756, + 0.06101306155323982, + 0.049206554889678955, + 0.07279578596353531, + 0.08856869488954544, + 0.07612704485654831, + 0.0532839298248291, + 0.12029335647821426, + 0.004649023991078138, + 0.002161930548027158, + 0.003995761275291443, + 0.004585200920701027, + 0.00305600231513381, + 0.004322811029851437, + 0.0035461075603961945, + 0.0031571565195918083, + 0.003634735709056258, + 0.0027206353843212128, + 0.006133491639047861, + 0.005447774194180965, + 0.002713908674195409, + 0.003916729707270861, + 0.004222660791128874, + 0.003942304756492376, + 0.0038124118000268936, + 0.001620565541088581, + 0.0025893982965499163, + 0.002651380142197013, + 0.0024358953814953566, + 0.0027605078648775816, + 0.003432304598391056, + 0.002451219130307436, + 0.003520035417750478, + 0.003781374543905258, + 0.005436371546238661, + 0.005030855070799589, + 0.0034031448885798454, + 0.004855053033679724, + 0.005849238485097885, + 0.0033530269283801317, + 0.0074501121416687965, + 0.00816190242767334, + 0.00836328323930502, + 0.010892431251704693, + 0.007953385822474957, + 0.010571255348622799, + 0.008184271864593029, + 0.010324236936867237, + 0.003574584610760212, + 0.0036087764892727137, + 0.0030624126084148884, + 0.004707003943622112, + 0.0034565699752420187, + 0.004673297051340342, + 0.003449639305472374, + 0.004721025004982948, + 0.0049589816480875015, + 0.005067151505500078, + 0.006731053814291954, + 0.006629949901252985, + 0.005712123587727547, + 0.0068938247859478, + 0.005177072249352932, + 0.00911562517285347, + 0.008821635507047176, + 0.011519784107804298, + 0.012352674268186092, + 0.009388013742864132, + 0.006815133150666952, + 0.012618325650691986, + 0.007418695371598005, + 0.016654275357723236, + 0.002158472081646323, + 0.0025980763603001833, + 0.0031848561484366655, + 0.0028067557141184807, + 0.004032074473798275, + 0.0035782880149781704, + 0.0043501779437065125, + 0.00312750949524343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.25210803747177124 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635204285383224 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008635204285383224, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635204285383224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36562633514404297, + 0.09356153011322021, + 0.09041334688663483, + 0.03097415342926979, + 0.0408366434276104, + 0.04123682156205177, + 0.09082106500864029, + 0.06740179657936096, + 0.1496560126543045, + 0.02947230450809002 + ], + "credit_map": [ + 0.05401216447353363, + 0.028567681089043617, + 0.03234090283513069, + 0.041966889053583145, + 0.04520048201084137, + 0.04800313711166382, + 0.03359832987189293, + 0.08193676173686981, + 0.016386114060878754, + 0.006545827258378267, + 0.01205446571111679, + 0.010198652744293213, + 0.012202519923448563, + 0.012210196815431118, + 0.009064409881830215, + 0.014899344183504581, + 0.007290596608072519, + 0.007063003722578287, + 0.021110791712999344, + 0.012589880265295506, + 0.00805190671235323, + 0.01287524402141571, + 0.012389716692268848, + 0.009042205289006233, + 0.0021801770199090242, + 0.002682521240785718, + 0.006411273498088121, + 0.0039917887188494205, + 0.0022908987011760473, + 0.00393255427479744, + 0.005385035183280706, + 0.0040999045595526695, + 0.0042489939369261265, + 0.00338238594122231, + 0.0066939485259354115, + 0.004755539819598198, + 0.00465377327054739, + 0.00503872474655509, + 0.006793380714952946, + 0.005269896239042282, + 0.0037390331272035837, + 0.005482847802340984, + 0.004121947567909956, + 0.007229223381727934, + 0.0035041216760873795, + 0.00624047452583909, + 0.006357226520776749, + 0.004561945796012878, + 0.007603407371789217, + 0.011474929749965668, + 0.011657824739813805, + 0.01634543389081955, + 0.00861288234591484, + 0.013564285822212696, + 0.011548335663974285, + 0.010013959370553493, + 0.005673215724527836, + 0.010893710888922215, + 0.007311766501516104, + 0.010822548530995846, + 0.006281421519815922, + 0.009317595511674881, + 0.008957560174167156, + 0.008143974468111992, + 0.016268715262413025, + 0.01980533078312874, + 0.026606788858771324, + 0.017160870134830475, + 0.01315750740468502, + 0.021926941350102425, + 0.013386049307882786, + 0.021343817934393883, + 0.0030253289733082056, + 0.002808752004057169, + 0.0036103492602705956, + 0.0033238546457141638, + 0.0033616304863244295, + 0.0041365670040249825, + 0.004967146553099155, + 0.004238675814121962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.008635204285383224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static low-angle shot captures a slow-motion territorial battle between two ornate, spiky brown shells, their forms locked in a silent yet powerful struggle. The scene is rendered in a gritty, industrial style, dominated by harsh directional lighting, creating a dramatic black-and-white contrast. Resembling calcareous, gnarled inkstones, these shells are an intricate tapestry of tan and black cynar, their surfaces covered in intricate growth lines and sharp, dangerous spines that are currently interlocked. They push against each other with tremendous, grinding force, a motion that is almost imperceptible yet conveys incredible pressure. They rest on a dark, metallic coral reef, an unnatural surface of pitted, corroded gunmetal gray that gleams and shimmers with oil in the intense light. A powerful light source, directed from high above and to the right, cuts through the frame, casting one side of the combatants in a stark, clinical glow that accentuates every rough texture while plunging the other into deep, impenetrable shadow. The color palette is cool and desaturated, creating a tense, brutal atmosphere that transforms this natural conflict into a raw, mechanical, and unforgiving duel.", + "seed": 18885, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_4/index_7981_seed_18885_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08631627969443798, + "vlm_logprob": -0.08631627969443798, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.08631627969443798, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01992216147482395 + ], + "opposite_logprob_tokens": [ + -4.019922256469727 + ], + "logprob_reward": -0.01992216147482395, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01992216147482395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09257502853870392, + 0.05673089995980263, + 0.03912368044257164, + 0.09326988458633423, + 0.11204195022583008, + 0.2634583115577698, + 0.2096131294965744, + 0.04465677589178085, + 0.061492204666137695, + 0.02703809179365635 + ], + "credit_map": [ + 0.01178724505007267, + 0.015471790917217731, + 0.01439211331307888, + 0.01023123413324356, + 0.008671395480632782, + 0.00907975621521473, + 0.013661066070199013, + 0.009280426427721977, + 0.003762793494388461, + 0.004153479356318712, + 0.006969595793634653, + 0.016849160194396973, + 0.004110151436179876, + 0.0049171303398907185, + 0.009763829410076141, + 0.00620475597679615, + 0.0031256377696990967, + 0.0022646980360150337, + 0.0035243681631982327, + 0.003235071897506714, + 0.002195753389969468, + 0.0026248202193528414, + 0.012498513795435429, + 0.0096548181027174, + 0.011130031198263168, + 0.008550108410418034, + 0.012629061937332153, + 0.013809142634272575, + 0.01322114747017622, + 0.010848472826182842, + 0.013268359936773777, + 0.009813555516302586, + 0.013653709553182125, + 0.018095819279551506, + 0.014252102002501488, + 0.015218512155115604, + 0.013775746338069439, + 0.010606292635202408, + 0.013060943223536015, + 0.013378828763961792, + 0.021228190511465073, + 0.02216479927301407, + 0.043263278901576996, + 0.02025749534368515, + 0.02008119225502014, + 0.018898366019129753, + 0.06281594932079315, + 0.05474907159805298, + 0.020818917080760002, + 0.030135206878185272, + 0.03554878011345863, + 0.019911879673600197, + 0.016373299062252045, + 0.0166926309466362, + 0.038166992366313934, + 0.03196541965007782, + 0.004085620865225792, + 0.005253602284938097, + 0.007831060327589512, + 0.0036570457741618156, + 0.003829941852018237, + 0.008486940525472164, + 0.006081258412450552, + 0.005431304685771465, + 0.005803181789815426, + 0.007833406329154968, + 0.010393495671451092, + 0.0066443439573049545, + 0.006226804573088884, + 0.008349472656846046, + 0.008015234023332596, + 0.008226265199482441, + 0.0034762979485094547, + 0.0036840110551565886, + 0.005815811455249786, + 0.002462004544213414, + 0.0024914760142564774, + 0.0031887113582342863, + 0.00226205145008862, + 0.003657728200778365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.01992216147482395 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902866393327713 + ], + "opposite_logprob_tokens": [ + -2.4640285968780518 + ], + "logprob_reward": -0.08902866393327713, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902866393327713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1056666448712349, + 0.09431759268045425, + 0.021074984222650528, + 0.05366694927215576, + 0.21720696985721588, + 0.19321247935295105, + 0.12150007486343384, + 0.02680135704576969, + 0.13184219598770142, + 0.03471076115965843 + ], + "credit_map": [ + 0.009883134625852108, + 0.018526587635278702, + 0.01174171082675457, + 0.0196057315915823, + 0.011453053914010525, + 0.007335675414651632, + 0.01636599190533161, + 0.010754767805337906, + 0.005656416527926922, + 0.006467681378126144, + 0.009782570414245129, + 0.03205332159996033, + 0.008086896501481533, + 0.00600469671189785, + 0.016909169033169746, + 0.009356854483485222, + 0.0023764672223478556, + 0.0021434405352920294, + 0.003153074299916625, + 0.00213389634154737, + 0.002301812870427966, + 0.001778214005753398, + 0.0034422387834638357, + 0.00374584156088531, + 0.007233039475977421, + 0.006165663246065378, + 0.00760759087279439, + 0.007569336332380772, + 0.007517574355006218, + 0.005275959614664316, + 0.006769316270947456, + 0.005528470501303673, + 0.023413699120283127, + 0.033794160932302475, + 0.023556962609291077, + 0.03415467217564583, + 0.022938871756196022, + 0.021425429731607437, + 0.031076058745384216, + 0.0268471110612154, + 0.016480563208460808, + 0.014785968698561192, + 0.03108966164290905, + 0.015155507251620293, + 0.013948239386081696, + 0.012331577017903328, + 0.04857030138373375, + 0.040850672870874405, + 0.015161082148551941, + 0.01777057535946369, + 0.019753092899918556, + 0.011327976360917091, + 0.010087340138852596, + 0.010383430868387222, + 0.020336603745818138, + 0.016679983586072922, + 0.0033184958156198263, + 0.00331750581972301, + 0.004815701860934496, + 0.002149922074750066, + 0.0022405877243727446, + 0.0035424053203314543, + 0.003667950863018632, + 0.003748789196833968, + 0.010834313929080963, + 0.012127634137868881, + 0.02413008362054825, + 0.015251269564032555, + 0.01563970372080803, + 0.017278797924518585, + 0.018145224079489708, + 0.018435170873999596, + 0.003723726375028491, + 0.004857023246586323, + 0.0071475026197731495, + 0.003520219586789608, + 0.004106958396732807, + 0.0038815161678940058, + 0.0030279646161943674, + 0.00444585457444191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.08902866393327713 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033687151968479156 + ], + "opposite_logprob_tokens": [ + -3.408687114715576 + ], + "logprob_reward": -0.033687151968479156, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033687151968479156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1162455752491951, + 0.034268662333488464, + 0.04996570572257042, + 0.08990126103162766, + 0.03132884204387665, + 0.4494602680206299, + 0.06764103472232819, + 0.06069786846637726, + 0.05412082001566887, + 0.04636995494365692 + ], + "credit_map": [ + 0.015649380162358284, + 0.013139506801962852, + 0.0185499656945467, + 0.011806524358689785, + 0.012160034850239754, + 0.012189455330371857, + 0.017739268019795418, + 0.01501144003123045, + 0.0025969457346946, + 0.0030951283406466246, + 0.003884237026795745, + 0.004722439218312502, + 0.003685383126139641, + 0.004843610804527998, + 0.005978662520647049, + 0.005462251603603363, + 0.006408539600670338, + 0.004652920179069042, + 0.006436030380427837, + 0.003799357684329152, + 0.005706394091248512, + 0.0035798533353954554, + 0.008724573999643326, + 0.010658039711415768, + 0.011760860681533813, + 0.007693309336900711, + 0.012857758440077305, + 0.016915425658226013, + 0.007075154222548008, + 0.008423530496656895, + 0.012170834466814995, + 0.013004389591515064, + 0.0031281504780054092, + 0.0027119657024741173, + 0.004796815104782581, + 0.0042143953032791615, + 0.0027804833371192217, + 0.0040277885273098946, + 0.004629009403288364, + 0.005040236283093691, + 0.04484120011329651, + 0.037736471742391586, + 0.08324478566646576, + 0.033888090401887894, + 0.03321794420480728, + 0.02594190463423729, + 0.10211703181266785, + 0.08847282826900482, + 0.006702156271785498, + 0.009739062748849392, + 0.013943102210760117, + 0.006311428267508745, + 0.004843994043767452, + 0.004904116503894329, + 0.010524989105761051, + 0.010672187432646751, + 0.005841651000082493, + 0.00751052750274539, + 0.009815382771193981, + 0.004655955359339714, + 0.007957366295158863, + 0.009592296555638313, + 0.008871392346918583, + 0.006453298032283783, + 0.004569449461996555, + 0.0070642814971506596, + 0.010304565541446209, + 0.004502379335463047, + 0.008064267225563526, + 0.007838387042284012, + 0.004389085341244936, + 0.0073884017765522, + 0.004742990713566542, + 0.01095079816877842, + 0.003974333871155977, + 0.004003694746643305, + 0.003897804534062743, + 0.008087139576673508, + 0.0046872892417013645, + 0.00602590199559927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.033687151968479156 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813195586204529 + ], + "opposite_logprob_tokens": [ + -1.4063196182250977 + ], + "logprob_reward": -0.2813195586204529, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813195586204529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1360047310590744, + 0.06031009182333946, + 0.05627076327800751, + 0.05768965557217598, + 0.2469422072172165, + 0.16645103693008423, + 0.1585298329591751, + 0.04220472276210785, + 0.035895492881536484, + 0.03970151022076607 + ], + "credit_map": [ + 0.01286481972783804, + 0.02431347593665123, + 0.013363024219870567, + 0.02031867764890194, + 0.021760301664471626, + 0.011895691975951195, + 0.016386771574616432, + 0.015101978555321693, + 0.004255456849932671, + 0.005080733448266983, + 0.004504386335611343, + 0.013171791099011898, + 0.008898260071873665, + 0.006643544416874647, + 0.010119213722646236, + 0.007636702619493008, + 0.005504526197910309, + 0.0036782873794436455, + 0.00611121067777276, + 0.006990876980125904, + 0.007547841873019934, + 0.004874840844422579, + 0.010705392807722092, + 0.010857784189283848, + 0.006770188454538584, + 0.006488954182714224, + 0.008774268440902233, + 0.009284518659114838, + 0.004384296014904976, + 0.00509211840108037, + 0.00807496253401041, + 0.008820346556603909, + 0.02952057309448719, + 0.0316636860370636, + 0.031412649899721146, + 0.03541387617588043, + 0.020626598969101906, + 0.024678967893123627, + 0.03904397040605545, + 0.03458187356591225, + 0.01752730831503868, + 0.01629532501101494, + 0.03761383891105652, + 0.014694576151669025, + 0.01370849460363388, + 0.012216944247484207, + 0.029836488887667656, + 0.024558046832680702, + 0.01839040219783783, + 0.02340734750032425, + 0.027200890704989433, + 0.014978785999119282, + 0.012159052304923534, + 0.012750137597322464, + 0.026285644620656967, + 0.023357585072517395, + 0.003911562263965607, + 0.006427655927836895, + 0.00713280588388443, + 0.005002847407013178, + 0.00509617617353797, + 0.00426438357681036, + 0.00493208970874548, + 0.005437205545604229, + 0.003812015289440751, + 0.004818924702703953, + 0.008216016925871372, + 0.0030752934981137514, + 0.0037773470394313335, + 0.003740081563591957, + 0.0037025229539722204, + 0.00475328927859664, + 0.003253617323935032, + 0.006001485511660576, + 0.005835163407027721, + 0.003687909571453929, + 0.004472770728170872, + 0.00780750485137105, + 0.00454585300758481, + 0.004097206052392721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.2813195586204529 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623862475156784 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.007623862475156784, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623862475156784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09344670921564102, + 0.14126043021678925, + 0.06527786701917648, + 0.08989335596561432, + 0.09408444911241531, + 0.21650265157222748, + 0.06647547334432602, + 0.1340363472700119, + 0.06208200007677078, + 0.03694074973464012 + ], + "credit_map": [ + 0.009552333503961563, + 0.01676776073873043, + 0.012917626649141312, + 0.012741670943796635, + 0.009570124559104443, + 0.008891049772500992, + 0.010291233658790588, + 0.012714904733002186, + 0.008070273324847221, + 0.008988210931420326, + 0.012126541696488857, + 0.05083806440234184, + 0.013434674590826035, + 0.008045154623687267, + 0.026309475302696228, + 0.013448028825223446, + 0.015670256689190865, + 0.0046875146217644215, + 0.007720861118286848, + 0.005426395684480667, + 0.0035735764540731907, + 0.0045488737523555756, + 0.013308149762451649, + 0.01034223660826683, + 0.012096105143427849, + 0.009747268632054329, + 0.011670642532408237, + 0.015554316341876984, + 0.008437172509729862, + 0.008729071356356144, + 0.012141900137066841, + 0.011516881175339222, + 0.009759224951267242, + 0.015054630115628242, + 0.011770982295274734, + 0.014942863024771214, + 0.009005208499729633, + 0.008326873183250427, + 0.01338919997215271, + 0.011835465207695961, + 0.017847076058387756, + 0.019247759133577347, + 0.03507796302437782, + 0.01703941635787487, + 0.014686251059174538, + 0.012334157712757587, + 0.0534384548664093, + 0.04683157056570053, + 0.006129712797701359, + 0.006981622893363237, + 0.01190630067139864, + 0.007496376056224108, + 0.004238943103700876, + 0.00790757592767477, + 0.009568372741341591, + 0.012246564961969852, + 0.013097759336233139, + 0.01547706313431263, + 0.021378885954618454, + 0.014204745180904865, + 0.012760120443999767, + 0.024779273197054863, + 0.012283424846827984, + 0.020055077970027924, + 0.005281493067741394, + 0.007206988055258989, + 0.011972473002970219, + 0.0052908556535840034, + 0.009807879105210304, + 0.0075619337148964405, + 0.006822999566793442, + 0.00813737977296114, + 0.00386106432415545, + 0.005068939179182053, + 0.007635477464646101, + 0.002268671290948987, + 0.003727062838152051, + 0.006436723284423351, + 0.0025289487093687057, + 0.005413861013948917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.007623862475156784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static low-angle shot captures a slow-motion territorial battle between two ornate, spiky brown shells, their forms locked in a silent yet powerful struggle. The scene is rendered in a gritty, industrial style, dominated by harsh directional lighting, creating a dramatic black-and-white contrast. Resembling calcareous, gnarled inkstones, these shells are an intricate tapestry of tan and black cynar, their surfaces covered in intricate growth lines and sharp, dangerous spines that are currently interlocked. They push against each other with tremendous, grinding force, a motion that is almost imperceptible yet conveys incredible pressure. They rest on a dark, metallic coral reef, an unnatural surface of pitted, corroded gunmetal gray that gleams and shimmers with oil in the intense light. A powerful light source, directed from high above and to the right, cuts through the frame, casting one side of the combatants in a stark, clinical glow that accentuates every rough texture while plunging the other into deep, impenetrable shadow. The color palette is cool and desaturated, creating a tense, brutal atmosphere that transforms this natural conflict into a raw, mechanical, and unforgiving duel.", + "seed": 18886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_4/index_7981_seed_18886_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16541657894849776, + "vlm_logprob": -0.16541657894849776, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.16541657894849776, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009844915941357613 + ], + "opposite_logprob_tokens": [ + -4.884844779968262 + ], + "logprob_reward": -0.009844915941357613, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009844915941357613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20261342823505402, + 0.08909296244382858, + 0.11867702007293701, + 0.07407383620738983, + 0.12425925582647324, + 0.1428845226764679, + 0.10279671847820282, + 0.06349977105855942, + 0.04975253716111183, + 0.03234993293881416 + ], + "credit_map": [ + 0.019671155139803886, + 0.023744473233819008, + 0.023327572271227837, + 0.03005163185298443, + 0.025799306109547615, + 0.016876546666026115, + 0.04006677120923996, + 0.023075977340340614, + 0.008852925151586533, + 0.012873131781816483, + 0.00982805248349905, + 0.013643224723637104, + 0.008661842904984951, + 0.011929476633667946, + 0.010232801549136639, + 0.013071510940790176, + 0.0072072953917086124, + 0.0189841128885746, + 0.008177561685442924, + 0.018336467444896698, + 0.010792179964482784, + 0.009685750119388103, + 0.02342863194644451, + 0.022065015509724617, + 0.010068637318909168, + 0.008801206946372986, + 0.006616870407015085, + 0.010495087131857872, + 0.006948770023882389, + 0.007810206618160009, + 0.011777795851230621, + 0.011555260978639126, + 0.010295971296727657, + 0.018229130655527115, + 0.013064558617770672, + 0.02224075421690941, + 0.013236867263913155, + 0.0158261526376009, + 0.015947163105010986, + 0.015418666414916515, + 0.014429702423512936, + 0.01371538732200861, + 0.013514326885342598, + 0.022850092500448227, + 0.013208906166255474, + 0.020816504955291748, + 0.01734100468456745, + 0.027008607983589172, + 0.006980734411627054, + 0.009662169963121414, + 0.008799897506833076, + 0.012187831103801727, + 0.0060885073617100716, + 0.01012764684855938, + 0.030573735013604164, + 0.018376190215349197, + 0.004234748426824808, + 0.005127394571900368, + 0.0037245163694024086, + 0.014066798612475395, + 0.003759912448003888, + 0.0059272293001413345, + 0.012281605042517185, + 0.014377566054463387, + 0.0054094321094453335, + 0.004323675762861967, + 0.0042001232504844666, + 0.007861417718231678, + 0.0043690879829227924, + 0.007006749976426363, + 0.007527241483330727, + 0.009054809808731079, + 0.0026306365616619587, + 0.0034787131007760763, + 0.003297896357253194, + 0.0032126586884260178, + 0.004233988933265209, + 0.006151990965008736, + 0.0044443788938224316, + 0.004899671301245689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.009844915941357613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761107802391052 + ], + "opposite_logprob_tokens": [ + -0.8261107802391052 + ], + "logprob_reward": -0.5761107802391052, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761107802391052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1773167997598648, + 0.11478756368160248, + 0.06876146048307419, + 0.09249813109636307, + 0.06055533513426781, + 0.1538131833076477, + 0.19299906492233276, + 0.06592201441526413, + 0.025241417810320854, + 0.048104964196681976 + ], + "credit_map": [ + 0.014171050861477852, + 0.024676671251654625, + 0.021760588511824608, + 0.02752046473324299, + 0.018225600942969322, + 0.023709816858172417, + 0.023054061457514763, + 0.02419855259358883, + 0.023502638563513756, + 0.016465885564684868, + 0.013096285052597523, + 0.013408869504928589, + 0.008486793376505375, + 0.01288294792175293, + 0.013125755824148655, + 0.013818391598761082, + 0.005956156179308891, + 0.014614807441830635, + 0.004261836875230074, + 0.01119090523570776, + 0.004955661483108997, + 0.0054143560118973255, + 0.011270781047642231, + 0.011096954345703125, + 0.008817421272397041, + 0.012706339359283447, + 0.007985771633684635, + 0.012734887190163136, + 0.007964557968080044, + 0.011157352477312088, + 0.017540309578180313, + 0.013591484166681767, + 0.009122907184064388, + 0.008347881957888603, + 0.006461611017584801, + 0.008567772805690765, + 0.005708558950573206, + 0.00653902068734169, + 0.008629132993519306, + 0.007178452331572771, + 0.017558645457029343, + 0.014138746075332165, + 0.014874754473567009, + 0.030129147693514824, + 0.014645259827375412, + 0.018954714760184288, + 0.021653831005096436, + 0.021858077496290207, + 0.015169710852205753, + 0.018196986988186836, + 0.016951339319348335, + 0.023216580972075462, + 0.013223916292190552, + 0.022256648167967796, + 0.05141575261950493, + 0.03256813809275627, + 0.003671643789857626, + 0.0034643446560949087, + 0.003369927639141679, + 0.01596846617758274, + 0.0032763320486992598, + 0.005950111895799637, + 0.014482489787042141, + 0.015738699585199356, + 0.0015379157848656178, + 0.002263996982946992, + 0.0022276812233030796, + 0.005270875059068203, + 0.002235345309600234, + 0.00267956405878067, + 0.004251886159181595, + 0.004774153232574463, + 0.003855503164231777, + 0.004310654941946268, + 0.003557814285159111, + 0.004564502276480198, + 0.004398610442876816, + 0.008738608099520206, + 0.010803288780152798, + 0.00787598080933094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.5761107802391052 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794851318001747 + ], + "opposite_logprob_tokens": [ + -3.529794931411743 + ], + "logprob_reward": -0.029794851318001747, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794851318001747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18313033878803253, + 0.028531784191727638, + 0.08729486167430878, + 0.05746757239103317, + 0.05369248986244202, + 0.2285776436328888, + 0.13881705701351166, + 0.08079475909471512, + 0.049236539751291275, + 0.09245697408914566 + ], + "credit_map": [ + 0.022852139547467232, + 0.018508315086364746, + 0.021366339176893234, + 0.027253901585936546, + 0.0228522177785635, + 0.013060650788247585, + 0.03570910170674324, + 0.021527662873268127, + 0.0032104188576340675, + 0.003305096412077546, + 0.0027348443400114775, + 0.004110721405595541, + 0.0030729712452739477, + 0.0040034344419837, + 0.004034045618027449, + 0.004060249775648117, + 0.004895877558737993, + 0.013286435976624489, + 0.005619343835860491, + 0.013539350591599941, + 0.0070160734467208385, + 0.006039183586835861, + 0.021202875301241875, + 0.015695707872509956, + 0.005120052024722099, + 0.006370801478624344, + 0.004817251581698656, + 0.010715139098465443, + 0.005743447691202164, + 0.006478514987975359, + 0.010062525980174541, + 0.008159836754202843, + 0.0045827035792171955, + 0.005751185119152069, + 0.00438444409519434, + 0.009044279344379902, + 0.005205246619880199, + 0.008443931117653847, + 0.009293465875089169, + 0.0069872308522462845, + 0.02573401853442192, + 0.021617619320750237, + 0.01750948093831539, + 0.03359755128622055, + 0.017780331894755363, + 0.03413653373718262, + 0.038755182176828384, + 0.039446912705898285, + 0.00815640389919281, + 0.012286324054002762, + 0.011066547594964504, + 0.015017129480838776, + 0.007669810671359301, + 0.011233785189688206, + 0.04522371292114258, + 0.02816334366798401, + 0.0036629545502364635, + 0.004656380973756313, + 0.004815997555851936, + 0.019123151898384094, + 0.004866013303399086, + 0.006122394930571318, + 0.01741415448486805, + 0.020133700221776962, + 0.003872992703691125, + 0.004641842562705278, + 0.004661859944462776, + 0.00544700538739562, + 0.004886490758508444, + 0.009293528273701668, + 0.00839901715517044, + 0.008033798076212406, + 0.0070815314538776875, + 0.010127137415111065, + 0.009024814702570438, + 0.012211401015520096, + 0.009147373028099537, + 0.015072278678417206, + 0.01573885977268219, + 0.01405357476323843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.029794851318001747 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155483484268188 + ], + "opposite_logprob_tokens": [ + -1.701554775238037 + ], + "logprob_reward": -0.20155483484268188, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155483484268188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06822030991315842, + 0.05359528586268425, + 0.08679747581481934, + 0.21276332437992096, + 0.03634831681847572, + 0.07457685470581055, + 0.13859932124614716, + 0.09771434217691422, + 0.11198943853378296, + 0.11939532309770584 + ], + "credit_map": [ + 0.0058959112502634525, + 0.0117301344871521, + 0.008554134517908096, + 0.010713519528508186, + 0.007208303082734346, + 0.0069835251197218895, + 0.007679561618715525, + 0.009455221705138683, + 0.010354583151638508, + 0.008597051724791527, + 0.004731560591608286, + 0.006339293904602528, + 0.004689305555075407, + 0.007002123631536961, + 0.0060252295807003975, + 0.005856139585375786, + 0.005838320124894381, + 0.017874063923954964, + 0.0053668394684791565, + 0.013308916240930557, + 0.008833741769194603, + 0.007164953742176294, + 0.013376203365623951, + 0.015034439042210579, + 0.0161780696362257, + 0.02478679269552231, + 0.015385043807327747, + 0.030294761061668396, + 0.011145619675517082, + 0.022517431527376175, + 0.058612722903490067, + 0.03384288400411606, + 0.004088396672159433, + 0.004170714411884546, + 0.0036288034170866013, + 0.005861872341483831, + 0.0030453072395175695, + 0.0038792421109974384, + 0.00738507742062211, + 0.004288904834538698, + 0.007408822420984507, + 0.008012441918253899, + 0.006310064345598221, + 0.008018567226827145, + 0.005322435405105352, + 0.010761276818811893, + 0.01666620746254921, + 0.012077043764293194, + 0.010098475962877274, + 0.012775718234479427, + 0.007538996636867523, + 0.01966436207294464, + 0.007848914712667465, + 0.01150432974100113, + 0.04234829172492027, + 0.026820242404937744, + 0.008037561550736427, + 0.012589549645781517, + 0.00656015332788229, + 0.014721734449267387, + 0.006134580355137587, + 0.009238149970769882, + 0.0225791335105896, + 0.017853477969765663, + 0.008762545883655548, + 0.009097924456000328, + 0.008959760889410973, + 0.020398469641804695, + 0.007928840816020966, + 0.015181286260485649, + 0.023419952020049095, + 0.018240656703710556, + 0.010579454712569714, + 0.012968851253390312, + 0.0075315008871257305, + 0.008176006376743317, + 0.015676358714699745, + 0.025990571826696396, + 0.017751330509781837, + 0.02072124555706978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.20155483484268188 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777512401342392 + ], + "opposite_logprob_tokens": [ + -4.634777545928955 + ], + "logprob_reward": -0.009777512401342392, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777512401342392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11508704721927643, + 0.1204252615571022, + 0.19426453113555908, + 0.05276139825582504, + 0.10790332406759262, + 0.11252898722887039, + 0.1375655233860016, + 0.02985016256570816, + 0.08374136686325073, + 0.045872416347265244 + ], + "credit_map": [ + 0.01481000054627657, + 0.01086324080824852, + 0.016410013660788536, + 0.016697857528924942, + 0.013199943117797375, + 0.012175587937235832, + 0.017138034105300903, + 0.013792365789413452, + 0.02724456414580345, + 0.014389154501259327, + 0.0179237462580204, + 0.013461639173328876, + 0.011082404293119907, + 0.010371309705078602, + 0.012174885720014572, + 0.013777564279735088, + 0.011550681665539742, + 0.034122467041015625, + 0.011786491610109806, + 0.030309785157442093, + 0.016532307490706444, + 0.012696233578026295, + 0.03982115164399147, + 0.03744540736079216, + 0.005839275661855936, + 0.008157515898346901, + 0.005878439638763666, + 0.007285196799784899, + 0.0041592298075556755, + 0.00691804988309741, + 0.007574514485895634, + 0.006949173752218485, + 0.007614584639668465, + 0.011979104951024055, + 0.008280307054519653, + 0.01777433417737484, + 0.009950115345418453, + 0.017800763249397278, + 0.019773976877331734, + 0.01473014336079359, + 0.010394638404250145, + 0.010382541455328465, + 0.010067256167531013, + 0.014588024467229843, + 0.008797395043075085, + 0.014556900598108768, + 0.023105930536985397, + 0.02063629776239395, + 0.013282159343361855, + 0.013808873482048512, + 0.012786349281668663, + 0.018017364665865898, + 0.010747229680418968, + 0.02100188285112381, + 0.02397076040506363, + 0.023950902745127678, + 0.002142571611329913, + 0.0027876212261617184, + 0.0019031903939321637, + 0.0038744069170206785, + 0.0021590497344732285, + 0.004338663071393967, + 0.006225509103387594, + 0.0064191496931016445, + 0.007233882322907448, + 0.007829923182725906, + 0.006987971719354391, + 0.012602819129824638, + 0.007547327317297459, + 0.015303821302950382, + 0.011863501742482185, + 0.014372125267982483, + 0.0043542650528252125, + 0.005036442074924707, + 0.003986047115176916, + 0.006835742387920618, + 0.0036274725571274757, + 0.00535354670137167, + 0.008510790765285492, + 0.008168109692633152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.009777512401342392 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static low-angle shot captures a slow-motion territorial battle between two ornate, spiky brown shells, their forms locked in a silent yet powerful struggle. The scene is rendered in a gritty, industrial style, dominated by harsh directional lighting, creating a dramatic black-and-white contrast. Resembling calcareous, gnarled inkstones, these shells are an intricate tapestry of tan and black cynar, their surfaces covered in intricate growth lines and sharp, dangerous spines that are currently interlocked. They push against each other with tremendous, grinding force, a motion that is almost imperceptible yet conveys incredible pressure. They rest on a dark, metallic coral reef, an unnatural surface of pitted, corroded gunmetal gray that gleams and shimmers with oil in the intense light. A powerful light source, directed from high above and to the right, cuts through the frame, casting one side of the combatants in a stark, clinical glow that accentuates every rough texture while plunging the other into deep, impenetrable shadow. The color palette is cool and desaturated, creating a tense, brutal atmosphere that transforms this natural conflict into a raw, mechanical, and unforgiving duel.", + "seed": 18887, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_4/index_7981_seed_18887_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08546374058350921, + "vlm_logprob": -0.08546374058350921, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.08546374058350921, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005341543350368738 + ], + "opposite_logprob_tokens": [ + -5.630341529846191 + ], + "logprob_reward": -0.005341543350368738, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.005341543350368738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15401992201805115, + 0.09656167775392532, + 0.1256442666053772, + 0.1401762068271637, + 0.03559902682900429, + 0.07476416975259781, + 0.036183133721351624, + 0.15814350545406342, + 0.09037361294031143, + 0.08853448182344437 + ], + "credit_map": [ + 0.01706123538315296, + 0.013763475231826305, + 0.018461542204022408, + 0.020922210067510605, + 0.017650706693530083, + 0.025221552699804306, + 0.019616052508354187, + 0.021323151886463165, + 0.010695643723011017, + 0.00613976176828146, + 0.013970569707453251, + 0.0125516252592206, + 0.012192762456834316, + 0.013605298474431038, + 0.015094944275915623, + 0.012311076745390892, + 0.012948242016136646, + 0.016407975926995277, + 0.012750014662742615, + 0.016010141000151634, + 0.015589922666549683, + 0.015276165679097176, + 0.020427491515874863, + 0.01623431406915188, + 0.018171751871705055, + 0.010437688790261745, + 0.02355790138244629, + 0.015526054427027702, + 0.012587564997375011, + 0.020120447501540184, + 0.01387086696922779, + 0.02590392343699932, + 0.005847190972417593, + 0.0028181830421090126, + 0.007399562280625105, + 0.0029089106246829033, + 0.0022490755654871464, + 0.003342426847666502, + 0.00446685217320919, + 0.006566822528839111, + 0.012163941748440266, + 0.007981597445905209, + 0.009113956242799759, + 0.008997492492198944, + 0.004600603599101305, + 0.009829254820942879, + 0.01018450129777193, + 0.011892829090356827, + 0.005088355392217636, + 0.002779834670946002, + 0.005743668880313635, + 0.004204506520181894, + 0.002929499838501215, + 0.0032362521160393953, + 0.006175020709633827, + 0.006025999318808317, + 0.025748489424586296, + 0.013440476730465889, + 0.02404559776186943, + 0.01995321549475193, + 0.007912032306194305, + 0.01523322332650423, + 0.02482779696583748, + 0.02698267064988613, + 0.01012816745787859, + 0.011581557802855968, + 0.009791726246476173, + 0.014452428556978703, + 0.008898279629647732, + 0.010905678384006023, + 0.011031894013285637, + 0.013583878055214882, + 0.010662727989256382, + 0.01007375493645668, + 0.011585760861635208, + 0.01493134256452322, + 0.00744264991953969, + 0.010908869095146656, + 0.011310246773064137, + 0.011619132943451405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.005341543350368738 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988397181034088 + ], + "opposite_logprob_tokens": [ + -1.8048839569091797 + ], + "logprob_reward": -0.17988397181034088, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988397181034088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14201034605503082, + 0.05654091015458107, + 0.42199233174324036, + 0.07625266164541245, + 0.046377602964639664, + 0.040101099759340286, + 0.06389658898115158, + 0.03105524741113186, + 0.0836763009428978, + 0.03809690102934837 + ], + "credit_map": [ + 0.02231941744685173, + 0.012712203897535801, + 0.02045639604330063, + 0.023297911509871483, + 0.009957192465662956, + 0.015500899404287338, + 0.018794918432831764, + 0.018971409648656845, + 0.007570809684693813, + 0.005143481306731701, + 0.006907477974891663, + 0.00697162514552474, + 0.0067777871154248714, + 0.010198395699262619, + 0.006207174155861139, + 0.006764158606529236, + 0.04783459007740021, + 0.04939219728112221, + 0.04525396227836609, + 0.05621165782213211, + 0.04056629166007042, + 0.048409026116132736, + 0.07073304057121277, + 0.06359154731035233, + 0.011593851260840893, + 0.00735766626894474, + 0.011236660182476044, + 0.009910046122968197, + 0.005899297539144754, + 0.008059695363044739, + 0.009453780949115753, + 0.01274165976792574, + 0.004888839554041624, + 0.00424521928653121, + 0.004580789245665073, + 0.0054372455924749374, + 0.004850634373724461, + 0.010537956841289997, + 0.004498079884797335, + 0.007338839583098888, + 0.005211349111050367, + 0.004311771597713232, + 0.004925788380205631, + 0.005322449840605259, + 0.002736796857789159, + 0.006224202457815409, + 0.005302595440298319, + 0.006066146772354841, + 0.009437271393835545, + 0.006987631320953369, + 0.009183661080896854, + 0.007892434485256672, + 0.005063368007540703, + 0.008462047204375267, + 0.008282219991087914, + 0.008587954565882683, + 0.004399242345243692, + 0.0030868565663695335, + 0.004589871969074011, + 0.004228493664413691, + 0.0018319889204576612, + 0.0035677028354257345, + 0.004417175892740488, + 0.004933915566653013, + 0.010365852154791355, + 0.008949738927185535, + 0.009687218815088272, + 0.011988857761025429, + 0.007734295912086964, + 0.00839966256171465, + 0.01245163194835186, + 0.014099044725298882, + 0.0055529954843223095, + 0.004834785591810942, + 0.004912257194519043, + 0.004850358236581087, + 0.0030453060753643513, + 0.005111693870276213, + 0.0041009667329490185, + 0.005688536446541548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.17988397181034088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026339344680309296 + ], + "opposite_logprob_tokens": [ + -3.651339292526245 + ], + "logprob_reward": -0.026339344680309296, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026339344680309296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2316845953464508, + 0.09327585995197296, + 0.07749279588460922, + 0.09607213735580444, + 0.024148518219590187, + 0.08778682351112366, + 0.04866359755396843, + 0.1915256828069687, + 0.08836998045444489, + 0.060980070382356644 + ], + "credit_map": [ + 0.03517676889896393, + 0.01815105974674225, + 0.04984608665108681, + 0.027265703305602074, + 0.018073633313179016, + 0.02300374209880829, + 0.02643560990691185, + 0.03373194858431816, + 0.011919661425054073, + 0.006568187847733498, + 0.007371572777628899, + 0.016486434265971184, + 0.009662124328315258, + 0.01286242250353098, + 0.014029520563781261, + 0.014375918544828892, + 0.007512812968343496, + 0.009095854125916958, + 0.008255040273070335, + 0.008982876315712929, + 0.008618168532848358, + 0.009034207090735435, + 0.015224114991724491, + 0.010769703425467014, + 0.017670055851340294, + 0.005480969324707985, + 0.018245266750454903, + 0.01071174070239067, + 0.005019508767873049, + 0.006143867038190365, + 0.013629317283630371, + 0.019171396270394325, + 0.0033693923614919186, + 0.0017262010369449854, + 0.00539417564868927, + 0.0020348636899143457, + 0.0016637272201478481, + 0.002025884808972478, + 0.0032618253026157618, + 0.004672447219491005, + 0.016116639599204063, + 0.00806818064302206, + 0.015688598155975342, + 0.008339067921042442, + 0.004567047581076622, + 0.006828603334724903, + 0.013947678729891777, + 0.014230995438992977, + 0.009159602224826813, + 0.00395365571603179, + 0.007633003871887922, + 0.004623180255293846, + 0.0030363595578819513, + 0.004797113593667746, + 0.007329341024160385, + 0.00813133642077446, + 0.03269059211015701, + 0.015364949591457844, + 0.03182262182235718, + 0.024571385234594345, + 0.009188567288219929, + 0.012669678777456284, + 0.03248519450426102, + 0.032732680439949036, + 0.012537137605249882, + 0.010010967962443829, + 0.012229803949594498, + 0.01210069376975298, + 0.007146279327571392, + 0.006960031110793352, + 0.012628326192498207, + 0.014756724238395691, + 0.0071934242732822895, + 0.004357265308499336, + 0.00869558285921812, + 0.009828810580074787, + 0.005667844321578741, + 0.004805413074791431, + 0.012912461534142494, + 0.007519262842833996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.026339344680309296 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015625387430191 + ], + "opposite_logprob_tokens": [ + -1.701562523841858 + ], + "logprob_reward": -0.2015625387430191, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015625387430191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1588972508907318, + 0.21302677690982819, + 0.1452402025461197, + 0.08696532994508743, + 0.09475139528512955, + 0.033095769584178925, + 0.05095935985445976, + 0.05139463394880295, + 0.06764914840459824, + 0.09802012145519257 + ], + "credit_map": [ + 0.029638191685080528, + 0.011493456549942493, + 0.025974033400416374, + 0.015529943630099297, + 0.011932545341551304, + 0.019788319244980812, + 0.02395194210112095, + 0.020588790997862816, + 0.03994070738554001, + 0.01048996765166521, + 0.015774475410580635, + 0.04212690144777298, + 0.01594473421573639, + 0.02983912266790867, + 0.026422005146741867, + 0.03248884156346321, + 0.014451954513788223, + 0.017170364037156105, + 0.02501368522644043, + 0.016229594126343727, + 0.011623705737292767, + 0.01824781484901905, + 0.02649012953042984, + 0.01601294055581093, + 0.013542893342673779, + 0.007412244100123644, + 0.013688667677342892, + 0.008591803722083569, + 0.005913287401199341, + 0.008801614865660667, + 0.014900857582688332, + 0.014113953337073326, + 0.012280058115720749, + 0.007970491424202919, + 0.017532343044877052, + 0.009888171218335629, + 0.006874424871057272, + 0.009000500664114952, + 0.011014359071850777, + 0.020191039890050888, + 0.005726254545152187, + 0.003262929618358612, + 0.004435693845152855, + 0.003794493153691292, + 0.0023611050564795732, + 0.003554109949618578, + 0.004975365940481424, + 0.004985814448446035, + 0.006538938265293837, + 0.005043754354119301, + 0.005233049392700195, + 0.007202237378805876, + 0.004607820417732, + 0.009299318306148052, + 0.00587848573923111, + 0.0071557555347681046, + 0.00832313671708107, + 0.004530913196504116, + 0.00475799385458231, + 0.005951117258518934, + 0.003403137205168605, + 0.00842235516756773, + 0.007289445959031582, + 0.008716531097888947, + 0.008295399136841297, + 0.007337809540331364, + 0.009234410710632801, + 0.008167627267539501, + 0.005221857223659754, + 0.005041361320763826, + 0.014495220966637135, + 0.009855460375547409, + 0.012367574498057365, + 0.007078648544847965, + 0.02030792646110058, + 0.013157643377780914, + 0.008763371966779232, + 0.006503640208393335, + 0.015033242292702198, + 0.014808062463998795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.2015625387430191 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two ornate, spiky brown shells visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the shells engaged in a slow-motion struggle where their forms are locked together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a dark, metallic coral reef with a pitted, corroded gunmetal gray surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting harsh and directional, coming from high above and to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an extreme close-up with a static camera angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191304333508015 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014191304333508015, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191304333508015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1804250180721283, + 0.08818581700325012, + 0.38940468430519104, + 0.07274319976568222, + 0.06044992804527283, + 0.024898672476410866, + 0.0378708690404892, + 0.05416061356663704, + 0.04232625663280487, + 0.04953501373529434 + ], + "credit_map": [ + 0.024742696434259415, + 0.010032479651272297, + 0.026091990992426872, + 0.027917008846998215, + 0.017184356227517128, + 0.023768264800310135, + 0.0214984193444252, + 0.029189780354499817, + 0.01127801276743412, + 0.007026415318250656, + 0.007298152428120375, + 0.021292023360729218, + 0.008512023836374283, + 0.010271457955241203, + 0.013023986481130123, + 0.009483728557825089, + 0.04620346426963806, + 0.03906573727726936, + 0.04473048076033592, + 0.05072600021958351, + 0.03589671477675438, + 0.041219670325517654, + 0.07752278447151184, + 0.054039787501096725, + 0.012481407262384892, + 0.007027208339422941, + 0.008827702142298222, + 0.008798737078905106, + 0.006321740336716175, + 0.011292233131825924, + 0.008278664201498032, + 0.009715504013001919, + 0.008272596634924412, + 0.004720015451312065, + 0.008665471337735653, + 0.006111884489655495, + 0.004080197308212519, + 0.005811137612909079, + 0.00755047332495451, + 0.015238143503665924, + 0.004016765858978033, + 0.0020252554677426815, + 0.004844324197620153, + 0.0029468319844454527, + 0.001680118846707046, + 0.0018841851269826293, + 0.0037275489885360003, + 0.0037736394442617893, + 0.005345330573618412, + 0.003705034963786602, + 0.005545551422983408, + 0.004219869617372751, + 0.0020751317497342825, + 0.004233774729073048, + 0.006526662968099117, + 0.006219513714313507, + 0.008789228275418282, + 0.004166743252426386, + 0.00874275527894497, + 0.0066090356558561325, + 0.0025243491400033236, + 0.0033506546169519424, + 0.009826451539993286, + 0.010151393711566925, + 0.004654653836041689, + 0.004488054662942886, + 0.005582611076533794, + 0.005103966686874628, + 0.003642421681433916, + 0.005376772955060005, + 0.007423927541822195, + 0.0060538481920957565, + 0.005859022028744221, + 0.003641125513240695, + 0.009115760214626789, + 0.007544301450252533, + 0.0037814672105014324, + 0.005380905698984861, + 0.008240965195000172, + 0.005971465725451708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.014191304333508015 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_5/index_tensor([8535, 8535])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_5/index_tensor([8535, 8535])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a765bef74162afaa1c9b1e8c9d3d6953c302ee15 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_5/index_tensor([8535, 8535])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning close-up shot in the style of a nature documentary, the camera, in a slow, smooth orbiting motion, rotates 360 degrees around a helmet jellyfish. Its nearly transparent bell, sculpted from crystal, emanates a faint, ghostly, pale-blue bioluminescence against the inky blackness of the deep sea, revealing the intricate structure within. The ghostly creature gracefully and rhythmically contracts and expands its bell, each pulse leisurely propelling it upward through the endless darkness. Descending from its base, delicate, thread-like tentacles float and dance softly behind it like a dreamy veil. The background is a bottomless inky black, shimmering only in the soft beam of the photographic light, tiny particles of marine snow, like dust motes, further accentuating the jellyfish's solitary and otherworldly presence. The entire rotating lens captures every dreamlike moment in astonishing detail, vividly displaying the fragile beauty and mystery of this deep-sea life, creating a tranquil, ethereal and awe-inspiring atmosphere.", + "seed": 14669, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_5/index_8535_seed_14669_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06001828275620937, + "vlm_logprob": -0.06001828275620937, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.06001828275620937, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018276968970894814 + ], + "opposite_logprob_tokens": [ + -4.143277168273926 + ], + "logprob_reward": -0.018276968970894814, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.018276968970894814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2980438768863678, + 0.08573928475379944, + 0.06798778474330902, + 0.11056090891361237, + 0.11207140237092972, + 0.035624366253614426, + 0.07727791368961334, + 0.058605387806892395, + 0.06837876886129379, + 0.08571039885282516 + ], + "credit_map": [ + 0.038307543843984604, + 0.026170959696173668, + 0.07314138114452362, + 0.025517838075757027, + 0.020646050572395325, + 0.05237703025341034, + 0.030355263501405716, + 0.031527791172266006, + 0.009606261737644672, + 0.01078871265053749, + 0.014361213892698288, + 0.009771679528057575, + 0.01050527859479189, + 0.009730692021548748, + 0.011947134509682655, + 0.009028314612805843, + 0.006753095425665379, + 0.006967512890696526, + 0.013269265182316303, + 0.00641104020178318, + 0.008028990589082241, + 0.012726943008601665, + 0.006006181240081787, + 0.007824748754501343, + 0.01047428883612156, + 0.010751754976809025, + 0.014956586062908173, + 0.010846325196325779, + 0.015659959986805916, + 0.01435050182044506, + 0.012156516313552856, + 0.021364979445934296, + 0.014640340581536293, + 0.010523735545575619, + 0.01435568742454052, + 0.01052902638912201, + 0.015789750963449478, + 0.01343666110187769, + 0.021624930202960968, + 0.011171271093189716, + 0.002990253735333681, + 0.0036328036803752184, + 0.004793470725417137, + 0.004096230957657099, + 0.0040087150409817696, + 0.005123601760715246, + 0.004885670728981495, + 0.006093619391322136, + 0.012788902036845684, + 0.006082391832023859, + 0.012820851989090443, + 0.00797282438725233, + 0.010667690075933933, + 0.004044311586767435, + 0.013142569921910763, + 0.009758374653756618, + 0.01348817441612482, + 0.004467740189284086, + 0.010151111520826817, + 0.006176294758915901, + 0.004857812542468309, + 0.003955465275794268, + 0.0062155043706297874, + 0.009293286129832268, + 0.010572477243840694, + 0.006821522489190102, + 0.009900463744997978, + 0.00830961111932993, + 0.010885009542107582, + 0.00859337393194437, + 0.007100818678736687, + 0.0061954897828400135, + 0.008428888395428658, + 0.011877905577421188, + 0.010046851821243763, + 0.009530648589134216, + 0.014359889551997185, + 0.0071150134317576885, + 0.01592261530458927, + 0.008428583852946758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.018276968970894814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864796806126833 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.00864796806126833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864796806126833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2806360721588135, + 0.2553260922431946, + 0.1005914956331253, + 0.11191588640213013, + 0.07504667341709137, + 0.020605826750397682, + 0.029476433992385864, + 0.03935689106583595, + 0.05692470446228981, + 0.03011987917125225 + ], + "credit_map": [ + 0.03470353037118912, + 0.02686258591711521, + 0.04801395162940025, + 0.023660048842430115, + 0.02128751575946808, + 0.03824950382113457, + 0.04267728701233864, + 0.04518168047070503, + 0.029804443940520287, + 0.02810572274029255, + 0.04994891211390495, + 0.025195837020874023, + 0.026350295171141624, + 0.041134025901556015, + 0.0367853157222271, + 0.018001539632678032, + 0.00980081781744957, + 0.005886184051632881, + 0.02710299752652645, + 0.00714058056473732, + 0.007800542749464512, + 0.024368707090616226, + 0.006099745165556669, + 0.012391927652060986, + 0.013236449100077152, + 0.011233881115913391, + 0.016427360475063324, + 0.010972263291478157, + 0.015166870318353176, + 0.016933834180235863, + 0.013440228998661041, + 0.014504997059702873, + 0.007787887006998062, + 0.008749382570385933, + 0.009992335923016071, + 0.008870454505085945, + 0.009456184692680836, + 0.008093724027276039, + 0.012582856230437756, + 0.009513850323855877, + 0.0028057119343429804, + 0.0014188748318701982, + 0.003033694578334689, + 0.0015462794108316302, + 0.0022559999488294125, + 0.0025662584230303764, + 0.0020962469279766083, + 0.004882760811597109, + 0.004994888324290514, + 0.0021784633863717318, + 0.004489042330533266, + 0.003363931318745017, + 0.0037797042168676853, + 0.002206145552918315, + 0.0048283543437719345, + 0.003635906847193837, + 0.0077313389629125595, + 0.00378547515720129, + 0.0058563463389873505, + 0.004903455730527639, + 0.0038123310077935457, + 0.0029385550878942013, + 0.005374823231250048, + 0.004954565782099962, + 0.00932594295591116, + 0.007878151722252369, + 0.007853229530155659, + 0.008239846676588058, + 0.0051030851900577545, + 0.0070863631553947926, + 0.005648498423397541, + 0.005789586808532476, + 0.0039155674166977406, + 0.0032448009587824345, + 0.004503187257796526, + 0.0035524419508874416, + 0.005612859036773443, + 0.0021604085341095924, + 0.003693314967676997, + 0.003437299281358719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.00864796806126833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603444516658783 + ], + "opposite_logprob_tokens": [ + -1.9103444814682007 + ], + "logprob_reward": -0.1603444516658783, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603444516658783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25325116515159607, + 0.13018330931663513, + 0.10495533049106598, + 0.029673099517822266, + 0.11568307131528854, + 0.05712809041142464, + 0.02967582829296589, + 0.1653462052345276, + 0.036294467747211456, + 0.07780950516462326 + ], + "credit_map": [ + 0.033902037888765335, + 0.02041138894855976, + 0.05434199422597885, + 0.02015095017850399, + 0.015227876603603363, + 0.04131358116865158, + 0.026188895106315613, + 0.04171445965766907, + 0.010481372475624084, + 0.009036551229655743, + 0.03506297245621681, + 0.014164394699037075, + 0.007069137413054705, + 0.036555517464876175, + 0.009689733386039734, + 0.008123625069856644, + 0.008227147161960602, + 0.010258467867970467, + 0.020323097705841064, + 0.0081999022513628, + 0.01327984593808651, + 0.019197285175323486, + 0.009088678285479546, + 0.01638091169297695, + 0.0035764172207564116, + 0.0020775008015334606, + 0.0042623551562428474, + 0.00430107768625021, + 0.003096778644248843, + 0.004079297184944153, + 0.0035764905624091625, + 0.00470318179577589, + 0.015615534037351608, + 0.011745710857212543, + 0.0180430319160223, + 0.01161597017198801, + 0.013945543207228184, + 0.014415059238672256, + 0.018669361248612404, + 0.011632866226136684, + 0.005991295445710421, + 0.004857982974499464, + 0.008173620328307152, + 0.00601714663207531, + 0.005789855029433966, + 0.00755271315574646, + 0.00678570382297039, + 0.011959774419665337, + 0.003420302178710699, + 0.002679306548088789, + 0.0037831622175872326, + 0.003071683691814542, + 0.004490299616008997, + 0.002621124731376767, + 0.004534448031336069, + 0.005075500346720219, + 0.028472820296883583, + 0.013382394798099995, + 0.028460538014769554, + 0.02298515848815441, + 0.016883648931980133, + 0.011689893901348114, + 0.019839219748973846, + 0.023632533848285675, + 0.004913415294140577, + 0.0036530368961393833, + 0.00495098065584898, + 0.005840074270963669, + 0.00548132648691535, + 0.004344585817307234, + 0.003976229578256607, + 0.003134818049147725, + 0.008164791390299797, + 0.00964075606316328, + 0.010213117115199566, + 0.009702932089567184, + 0.011942500248551369, + 0.006200497969985008, + 0.013865825720131397, + 0.008079083636403084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.1603444516658783 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029090940952301 + ], + "opposite_logprob_tokens": [ + -2.3502910137176514 + ], + "logprob_reward": -0.10029090940952301, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029090940952301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4879398047924042, + 0.0797821432352066, + 0.03375969082117081, + 0.12884588539600372, + 0.04899607598781586, + 0.09115373343229294, + 0.015619320794939995, + 0.07432115823030472, + 0.01678810827434063, + 0.02279413491487503 + ], + "credit_map": [ + 0.05571754276752472, + 0.07304145395755768, + 0.0572521947324276, + 0.054746855050325394, + 0.03884259611368179, + 0.06468819826841354, + 0.049923140555620193, + 0.09372780472040176, + 0.004665841814130545, + 0.004592114128172398, + 0.026688849553465843, + 0.004924269393086433, + 0.004585758317261934, + 0.02531271055340767, + 0.005136049352586269, + 0.0038765391800552607, + 0.0031561716459691525, + 0.002465439960360527, + 0.006793355569243431, + 0.003200179198756814, + 0.0039011496119201183, + 0.005238120444118977, + 0.0030910931527614594, + 0.005914175882935524, + 0.014134815894067287, + 0.010517594404518604, + 0.016878139227628708, + 0.012783470563590527, + 0.016145693138241768, + 0.016496187075972557, + 0.013442181050777435, + 0.028447801247239113, + 0.006192505359649658, + 0.00509825861081481, + 0.007723581977188587, + 0.00542102288454771, + 0.007264998741447926, + 0.006265627220273018, + 0.005566578824073076, + 0.005463501438498497, + 0.00916949100792408, + 0.007805224973708391, + 0.01321882288902998, + 0.009766853414475918, + 0.01006338745355606, + 0.010595262981951237, + 0.01074887439608574, + 0.019785813987255096, + 0.0015986509388312697, + 0.0013421517796814442, + 0.0020505357533693314, + 0.0019411032553762197, + 0.0028385398909449577, + 0.0011232240358367562, + 0.0026210620999336243, + 0.0021040516439825296, + 0.013167397119104862, + 0.007057443261146545, + 0.011248383671045303, + 0.011124948039650917, + 0.007617673836648464, + 0.005508983973413706, + 0.008961870335042477, + 0.009634454734623432, + 0.002550526987761259, + 0.0018190492410212755, + 0.0023109910544008017, + 0.0021438964176923037, + 0.0032974863424897194, + 0.0018316612113267183, + 0.0016111809527501464, + 0.0012233145534992218, + 0.002111922251060605, + 0.003586843376979232, + 0.0027268549893051386, + 0.0022824518382549286, + 0.0038650522474199533, + 0.001853014575317502, + 0.004281194880604744, + 0.0020867998246103525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.10029090940952301 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531115673482418 + ], + "opposite_logprob_tokens": [ + -4.387531280517578 + ], + "logprob_reward": -0.012531115673482418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531115673482418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13795772194862366, + 0.1911625862121582, + 0.06548135727643967, + 0.1716017872095108, + 0.22574487328529358, + 0.07091600447893143, + 0.029034802690148354, + 0.025293564423918724, + 0.05321070924401283, + 0.02959655039012432 + ], + "credit_map": [ + 0.01621631719172001, + 0.02000441774725914, + 0.020613666623830795, + 0.020863380283117294, + 0.010905742645263672, + 0.01716635562479496, + 0.013349981978535652, + 0.018837863579392433, + 0.020826470106840134, + 0.02327549271285534, + 0.03703998774290085, + 0.02052151784300804, + 0.01797185093164444, + 0.02968650870025158, + 0.028482340276241302, + 0.013358407653868198, + 0.009144939482212067, + 0.005225324537605047, + 0.011060316115617752, + 0.0072955782525241375, + 0.005973855033516884, + 0.010343580506742, + 0.005639048293232918, + 0.010798710398375988, + 0.028983058407902718, + 0.017840338870882988, + 0.026426801458001137, + 0.01628030091524124, + 0.018109355121850967, + 0.025843607261776924, + 0.015778955072164536, + 0.022339362651109695, + 0.03236790746450424, + 0.021531937643885612, + 0.034157730638980865, + 0.03270206227898598, + 0.021873697638511658, + 0.03060821443796158, + 0.02205791138112545, + 0.03044542856514454, + 0.01058533787727356, + 0.005650151986628771, + 0.010090146213769913, + 0.005277401767671108, + 0.006108027417212725, + 0.011630586348474026, + 0.006097031757235527, + 0.015477323904633522, + 0.004337994381785393, + 0.0020255090203136206, + 0.004122893791645765, + 0.0025643089320510626, + 0.004108636174350977, + 0.0031682997941970825, + 0.003713873215019703, + 0.004993289243429899, + 0.0029472701717168093, + 0.003097256412729621, + 0.003517657984048128, + 0.0036247565876692533, + 0.002864756155759096, + 0.0022420398890972137, + 0.003858765587210655, + 0.00314106117002666, + 0.008737100288271904, + 0.005956077482551336, + 0.00767897954210639, + 0.008540237322449684, + 0.005330262705683708, + 0.006982678081840277, + 0.005134000908583403, + 0.004851373843848705, + 0.0026931853499263525, + 0.004967400338500738, + 0.003637987654656172, + 0.0028479776810854673, + 0.005175544880330563, + 0.0023461857344955206, + 0.005522892810404301, + 0.002405374078080058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.012531115673482418 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up shot in the style of a nature documentary, the camera, in a slow, smooth orbiting motion, rotates 360 degrees around a helmet jellyfish. Its nearly transparent bell, sculpted from crystal, emanates a faint, ghostly, pale-blue bioluminescence against the inky blackness of the deep sea, revealing the intricate structure within. The ghostly creature gracefully and rhythmically contracts and expands its bell, each pulse leisurely propelling it upward through the endless darkness. Descending from its base, delicate, thread-like tentacles float and dance softly behind it like a dreamy veil. The background is a bottomless inky black, shimmering only in the soft beam of the photographic light, tiny particles of marine snow, like dust motes, further accentuating the jellyfish's solitary and otherworldly presence. The entire rotating lens captures every dreamlike moment in astonishing detail, vividly displaying the fragile beauty and mystery of this deep-sea life, creating a tranquil, ethereal and awe-inspiring atmosphere.", + "seed": 14670, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_5/index_8535_seed_14670_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18324541011825204, + "vlm_logprob": -0.18324541011825204, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.18324541011825204, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035421695560216904 + ], + "opposite_logprob_tokens": [ + -3.41042160987854 + ], + "logprob_reward": -0.035421695560216904, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.035421695560216904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26113301515579224, + 0.1465228796005249, + 0.08152765780687332, + 0.059316959232091904, + 0.043798260390758514, + 0.08511439710855484, + 0.06078386679291725, + 0.1222265437245369, + 0.0704585388302803, + 0.06911788135766983 + ], + "credit_map": [ + 0.023429447785019875, + 0.020314296707510948, + 0.025992274284362793, + 0.05254741013050079, + 0.0276317335665226, + 0.028064917773008347, + 0.04596198722720146, + 0.037190958857536316, + 0.013589722104370594, + 0.01620607078075409, + 0.019409891217947006, + 0.023304741829633713, + 0.015857987105846405, + 0.01679575815796852, + 0.02510024793446064, + 0.01625846140086651, + 0.01619788631796837, + 0.007927396334707737, + 0.014195915311574936, + 0.007882338017225266, + 0.007598092779517174, + 0.009073799476027489, + 0.012509294785559177, + 0.006142931059002876, + 0.006502418778836727, + 0.0032456412445753813, + 0.008078287355601788, + 0.00950650218874216, + 0.008256924338638783, + 0.005798733793199062, + 0.009324217215180397, + 0.00860423594713211, + 0.005731025245040655, + 0.003978501074016094, + 0.006231496576219797, + 0.004017877858132124, + 0.005924935452640057, + 0.002778927329927683, + 0.011276036500930786, + 0.0038594617508351803, + 0.008513116277754307, + 0.008997178636491299, + 0.011402481235563755, + 0.007180743385106325, + 0.012273778207600117, + 0.008866438642144203, + 0.01728600636124611, + 0.010594651103019714, + 0.00821371003985405, + 0.007406875491142273, + 0.00901506096124649, + 0.006384450010955334, + 0.007892124354839325, + 0.004819604102522135, + 0.011151188984513283, + 0.005900850053876638, + 0.012879184447228909, + 0.01438848115503788, + 0.01708293706178665, + 0.02188427932560444, + 0.01541359443217516, + 0.009560921229422092, + 0.018534665927290916, + 0.01248248666524887, + 0.00685508456081152, + 0.004326730966567993, + 0.008713903836905956, + 0.01054415199905634, + 0.01213928498327732, + 0.004648443777114153, + 0.015151738189160824, + 0.00807920377701521, + 0.009861642494797707, + 0.005045756231993437, + 0.01046262588351965, + 0.008654585108160973, + 0.010472389869391918, + 0.0072552477940917015, + 0.01046633068472147, + 0.006899300497025251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.035421695560216904 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007640779484063387 + ], + "opposite_logprob_tokens": [ + -4.882640838623047 + ], + "logprob_reward": -0.007640779484063387, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007640779484063387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3882802724838257, + 0.0790289118885994, + 0.046708278357982635, + 0.031128307804465294, + 0.04337816685438156, + 0.08078039437532425, + 0.025028642266988754, + 0.10953563451766968, + 0.04057871922850609, + 0.1555526852607727 + ], + "credit_map": [ + 0.06484576314687729, + 0.023463264107704163, + 0.050866663455963135, + 0.05452755466103554, + 0.027206484228372574, + 0.07634615898132324, + 0.062493737787008286, + 0.028530649840831757, + 0.011233366094529629, + 0.008576680906116962, + 0.011119651608169079, + 0.009099164977669716, + 0.007854389026761055, + 0.012820041738450527, + 0.011021025478839874, + 0.007304590195417404, + 0.0076464093290269375, + 0.004822924733161926, + 0.007891100831329823, + 0.005934807006269693, + 0.0042570787481963634, + 0.005473426077514887, + 0.0067602540366351604, + 0.003922278061509132, + 0.003744719550013542, + 0.0017567521426826715, + 0.004712140653282404, + 0.005479726474732161, + 0.003985539544373751, + 0.002137113129720092, + 0.005336547736078501, + 0.003975765313953161, + 0.007610723376274109, + 0.0034495817963033915, + 0.007247081026434898, + 0.003352403175085783, + 0.004588541574776173, + 0.003184571163728833, + 0.008334576152265072, + 0.0056106895208358765, + 0.01018674485385418, + 0.01067020371556282, + 0.012269169092178345, + 0.006384364329278469, + 0.008353857323527336, + 0.00882001779973507, + 0.015360570512712002, + 0.00873546116054058, + 0.004388417117297649, + 0.0020947111770510674, + 0.003993656020611525, + 0.0024833218194544315, + 0.003715082071721554, + 0.0024658418260514736, + 0.003816422540694475, + 0.002071189694106579, + 0.013668020255863667, + 0.010065976530313492, + 0.01740352250635624, + 0.021185992285609245, + 0.016728244721889496, + 0.008362346328794956, + 0.012661044485867023, + 0.00946048367768526, + 0.007937311194837093, + 0.0025047746021300554, + 0.007247463334351778, + 0.004773471038788557, + 0.004630073439329863, + 0.0035162114072591066, + 0.005238000303506851, + 0.004731414839625359, + 0.032075248658657074, + 0.011398366652429104, + 0.027625346556305885, + 0.015208553522825241, + 0.020544853061437607, + 0.013400422409176826, + 0.02106194570660591, + 0.014237947762012482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.007640779484063387 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577953934669495 + ], + "opposite_logprob_tokens": [ + -0.6327953934669495 + ], + "logprob_reward": -0.7577953934669495, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577953934669495, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.30483487248420715, + 0.05428323149681091, + 0.05875347554683685, + 0.048737529665231705, + 0.07482577860355377, + 0.08058267086744308, + 0.04852811619639397, + 0.16348160803318024, + 0.062198247760534286, + 0.10377450287342072 + ], + "credit_map": [ + 0.04756447672843933, + 0.021024653688073158, + 0.0406673364341259, + 0.04771055281162262, + 0.02545330300927162, + 0.0411725714802742, + 0.057200297713279724, + 0.024041660130023956, + 0.008418826386332512, + 0.00583734642714262, + 0.009167931042611599, + 0.005946075078099966, + 0.0069884369149804115, + 0.004410371650010347, + 0.007565709762275219, + 0.005948531907051802, + 0.0076741380617022514, + 0.005928306374698877, + 0.007072658743709326, + 0.008426091633737087, + 0.0066443197429180145, + 0.004724998492747545, + 0.014267360791563988, + 0.0040156040340662, + 0.005564470309764147, + 0.0031851865351200104, + 0.005931274499744177, + 0.007806321140378714, + 0.006286012940108776, + 0.0038562784902751446, + 0.010159571655094624, + 0.005948411300778389, + 0.008134858682751656, + 0.006888676900416613, + 0.010957174003124237, + 0.005942836403846741, + 0.010573876090347767, + 0.00567666906863451, + 0.015240246430039406, + 0.011411439627408981, + 0.009524904191493988, + 0.00835332553833723, + 0.011254959739744663, + 0.0075493305921554565, + 0.008562606759369373, + 0.008022649213671684, + 0.012781363911926746, + 0.014533530920743942, + 0.009383347816765308, + 0.004751639906316996, + 0.008810884319245815, + 0.004726926330476999, + 0.006137568969279528, + 0.004291783086955547, + 0.006454717833548784, + 0.003971248399466276, + 0.02766217291355133, + 0.016208577901124954, + 0.028385113924741745, + 0.03606079891324043, + 0.013185339979827404, + 0.010945209302008152, + 0.019386716187000275, + 0.011647669598460197, + 0.005693064536899328, + 0.0036118857096880674, + 0.00794150959700346, + 0.008022643625736237, + 0.010335074737668037, + 0.004809030331671238, + 0.015003903768956661, + 0.006781133357435465, + 0.008982306346297264, + 0.008381354622542858, + 0.012883906252682209, + 0.012434063479304314, + 0.019369222223758698, + 0.01245138794183731, + 0.017385588958859444, + 0.011886672116816044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.7577953934669495 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903390169143677 + ], + "opposite_logprob_tokens": [ + -2.464033842086792 + ], + "logprob_reward": -0.08903390169143677, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903390169143677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11559171974658966, + 0.2125566303730011, + 0.13929250836372375, + 0.08630097657442093, + 0.07346406579017639, + 0.1281171590089798, + 0.026809783652424812, + 0.08436069637537003, + 0.03443119302392006, + 0.09907526522874832 + ], + "credit_map": [ + 0.01292019709944725, + 0.009345341473817825, + 0.014762052334845066, + 0.02448556013405323, + 0.00927859265357256, + 0.012692607007920742, + 0.018317226320505142, + 0.013790150173008442, + 0.029298201203346252, + 0.03870758041739464, + 0.03031885251402855, + 0.022174669429659843, + 0.020947333425283432, + 0.02710796520113945, + 0.026628287509083748, + 0.017373740673065186, + 0.014955611899495125, + 0.019190480932593346, + 0.01611826755106449, + 0.01808907464146614, + 0.01810307800769806, + 0.0098861800506711, + 0.03362693637609482, + 0.009322873316705227, + 0.006921546533703804, + 0.00484191719442606, + 0.011168847791850567, + 0.016236230731010437, + 0.015314202755689621, + 0.006386764347553253, + 0.01613156870007515, + 0.009299897588789463, + 0.00950195174664259, + 0.00462036207318306, + 0.009763556532561779, + 0.006101993378251791, + 0.010165475308895111, + 0.005081325303763151, + 0.019687622785568237, + 0.00854177214205265, + 0.017431456595659256, + 0.009110716171562672, + 0.01971500553190708, + 0.01114996150135994, + 0.013221264816820621, + 0.012275289744138718, + 0.01883026584982872, + 0.026383209973573685, + 0.00298012001439929, + 0.002174986293539405, + 0.004118151031434536, + 0.0031080199405550957, + 0.0045451088808476925, + 0.003567727515473962, + 0.0042494828812778, + 0.00206618569791317, + 0.009682577103376389, + 0.006167005747556686, + 0.011446469463407993, + 0.011627092026174068, + 0.01320875808596611, + 0.007247444242238998, + 0.01425319816917181, + 0.010728158988058567, + 0.003432267112657428, + 0.0021355950739234686, + 0.004872065037488937, + 0.0038364704232662916, + 0.005729454103857279, + 0.003363350871950388, + 0.00748479412868619, + 0.0035771969705820084, + 0.010204820893704891, + 0.009372326545417309, + 0.011774232611060143, + 0.011352693662047386, + 0.01622636429965496, + 0.010525664314627647, + 0.017390843480825424, + 0.012228313833475113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.08903390169143677 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026335280388593674 + ], + "opposite_logprob_tokens": [ + -3.6513352394104004 + ], + "logprob_reward": -0.026335280388593674, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026335280388593674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18358725309371948, + 0.10653325170278549, + 0.08288338780403137, + 0.06768239289522171, + 0.04404313862323761, + 0.08338095247745514, + 0.044433776289224625, + 0.07715914398431778, + 0.05045498162508011, + 0.2598416805267334 + ], + "credit_map": [ + 0.02483946830034256, + 0.011186460964381695, + 0.023372044786810875, + 0.029850246384739876, + 0.016492241993546486, + 0.034391071647405624, + 0.020628662779927254, + 0.022827042266726494, + 0.017467940226197243, + 0.012920268811285496, + 0.01583215594291687, + 0.012388668023049831, + 0.008743340149521828, + 0.017234379425644875, + 0.011597031727433205, + 0.010349463671445847, + 0.00960250198841095, + 0.013153575360774994, + 0.010860716924071312, + 0.010568537749350071, + 0.01026174146682024, + 0.007094653323292732, + 0.016017096117138863, + 0.005324568599462509, + 0.014371808618307114, + 0.004041601438075304, + 0.012360721826553345, + 0.008868878707289696, + 0.0061590662226080894, + 0.00559580372646451, + 0.009161584079265594, + 0.007122925017029047, + 0.006559961475431919, + 0.004536280874162912, + 0.006989335175603628, + 0.004590173251926899, + 0.005205615423619747, + 0.0023233576212078333, + 0.009989709593355656, + 0.0038487061392515898, + 0.009157818742096424, + 0.007275382988154888, + 0.011795812286436558, + 0.007418157998472452, + 0.011429063975811005, + 0.007027794606983662, + 0.015208775177598, + 0.014068148098886013, + 0.007151390891522169, + 0.004880180116742849, + 0.007266553118824959, + 0.0050036548636853695, + 0.005992827471345663, + 0.003756504040211439, + 0.006002469919621944, + 0.0043801963329315186, + 0.01080765575170517, + 0.004948874469846487, + 0.011989693157374859, + 0.010433584451675415, + 0.008837457746267319, + 0.004959517158567905, + 0.014558185823261738, + 0.010624169372022152, + 0.010479960590600967, + 0.004214059561491013, + 0.009041444398462772, + 0.00562864588573575, + 0.004653165582567453, + 0.004732027649879456, + 0.006261799484491348, + 0.005443880334496498, + 0.051792554557323456, + 0.0214729905128479, + 0.04500824213027954, + 0.02487611025571823, + 0.03605617955327034, + 0.023683447390794754, + 0.03409092128276825, + 0.022861240431666374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.026335280388593674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up shot in the style of a nature documentary, the camera, in a slow, smooth orbiting motion, rotates 360 degrees around a helmet jellyfish. Its nearly transparent bell, sculpted from crystal, emanates a faint, ghostly, pale-blue bioluminescence against the inky blackness of the deep sea, revealing the intricate structure within. The ghostly creature gracefully and rhythmically contracts and expands its bell, each pulse leisurely propelling it upward through the endless darkness. Descending from its base, delicate, thread-like tentacles float and dance softly behind it like a dreamy veil. The background is a bottomless inky black, shimmering only in the soft beam of the photographic light, tiny particles of marine snow, like dust motes, further accentuating the jellyfish's solitary and otherworldly presence. The entire rotating lens captures every dreamlike moment in astonishing detail, vividly displaying the fragile beauty and mystery of this deep-sea life, creating a tranquil, ethereal and awe-inspiring atmosphere.", + "seed": 14671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_5/index_8535_seed_14671_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.048269732668995854, + "vlm_logprob": -0.048269732668995854, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.048269732668995854, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02253449335694313 + ], + "opposite_logprob_tokens": [ + -3.8975346088409424 + ], + "logprob_reward": -0.02253449335694313, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02253449335694313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046530287712812424, + 0.10736899077892303, + 0.051745373755693436, + 0.12239336222410202, + 0.03314880654215813, + 0.04081157222390175, + 0.183092400431633, + 0.33245742321014404, + 0.0338827408850193, + 0.04856901243329048 + ], + "credit_map": [ + 0.004548670258373022, + 0.004518131725490093, + 0.006030059885233641, + 0.006696817930787802, + 0.005477569531649351, + 0.007162314839661121, + 0.005371750332415104, + 0.006724980194121599, + 0.012665074318647385, + 0.008947309106588364, + 0.008194082416594028, + 0.01266148965805769, + 0.009394384920597076, + 0.017381059005856514, + 0.020375600084662437, + 0.01774999126791954, + 0.005770326592028141, + 0.005688983015716076, + 0.005825027823448181, + 0.006846598349511623, + 0.0047544254921376705, + 0.007473435718566179, + 0.007256209850311279, + 0.008130371570587158, + 0.015294166281819344, + 0.009157760068774223, + 0.017014963552355766, + 0.01312998216599226, + 0.008478879928588867, + 0.024082304909825325, + 0.01278009545058012, + 0.02245521917939186, + 0.0035050222650170326, + 0.004211859777569771, + 0.004119005519896746, + 0.0034515263978391886, + 0.004609268624335527, + 0.004252771846950054, + 0.004031007178127766, + 0.004968347027897835, + 0.0041837007738649845, + 0.003924502991139889, + 0.0044924914836883545, + 0.005509738810360432, + 0.003787430003285408, + 0.007577815093100071, + 0.005158640909940004, + 0.006177259609103203, + 0.02647031471133232, + 0.01799868419766426, + 0.029164215549826622, + 0.020874613896012306, + 0.011578738689422607, + 0.02809414453804493, + 0.016336575150489807, + 0.032575130462646484, + 0.05259091407060623, + 0.022102374583482742, + 0.04856870323419571, + 0.03302006423473358, + 0.020432760939002037, + 0.05289550870656967, + 0.029157547280192375, + 0.07368957996368408, + 0.005748151335865259, + 0.003100639209151268, + 0.0048890020698308945, + 0.0051788450218737125, + 0.00270823878236115, + 0.0041731903329491615, + 0.0038279136642813683, + 0.004256763961166143, + 0.010121242143213749, + 0.0032353715505450964, + 0.008288373239338398, + 0.003805518615990877, + 0.00291768042370677, + 0.006308249197900295, + 0.0037026579957455397, + 0.010189922526478767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.02253449335694313 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787309914827347 + ], + "opposite_logprob_tokens": [ + -4.634787082672119 + ], + "logprob_reward": -0.009787309914827347, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787309914827347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05456450954079628, + 0.283185750246048, + 0.1788068264722824, + 0.06930065900087357, + 0.13961274921894073, + 0.047448549419641495, + 0.07681788504123688, + 0.04544469341635704, + 0.06234137341380119, + 0.04247693717479706 + ], + "credit_map": [ + 0.005552432965487242, + 0.005473581608384848, + 0.0098938699811697, + 0.00659734383225441, + 0.006904483772814274, + 0.006752220448106527, + 0.006386032793670893, + 0.007004547864198685, + 0.04663611575961113, + 0.021788399666547775, + 0.02403879351913929, + 0.01907017081975937, + 0.02943158894777298, + 0.04987331107258797, + 0.033967532217502594, + 0.05837985500693321, + 0.01750202476978302, + 0.015972020104527473, + 0.018595218658447266, + 0.020298609510064125, + 0.018015911802649498, + 0.027421943843364716, + 0.03547443076968193, + 0.025526685640215874, + 0.0076934234239161015, + 0.004855591803789139, + 0.010722991079092026, + 0.008201805874705315, + 0.0044958689250051975, + 0.013737683184444904, + 0.006615811958909035, + 0.012977493926882744, + 0.01897197589278221, + 0.013015147298574448, + 0.019405165687203407, + 0.017375104129314423, + 0.010530063882470131, + 0.02157026156783104, + 0.015192815102636814, + 0.023552240803837776, + 0.005446387454867363, + 0.00565257016569376, + 0.006484184879809618, + 0.008743447251617908, + 0.0028946606907993555, + 0.007916604168713093, + 0.004605431575328112, + 0.005705263931304216, + 0.014623159542679787, + 0.004570093937218189, + 0.01578979380428791, + 0.0045556314289569855, + 0.005319736432284117, + 0.010075347498059273, + 0.00523413484916091, + 0.0166500024497509, + 0.006490368861705065, + 0.004384894855320454, + 0.005933647509664297, + 0.005680564325302839, + 0.0023970426991581917, + 0.007032260298728943, + 0.004867183044552803, + 0.008658736012876034, + 0.01105118915438652, + 0.004773313645273447, + 0.00889509730041027, + 0.008539129048585892, + 0.004320476669818163, + 0.008433327078819275, + 0.0071212430484592915, + 0.009207598865032196, + 0.009207557886838913, + 0.003583867335692048, + 0.006170433480292559, + 0.004457428585737944, + 0.0029209237545728683, + 0.004359664395451546, + 0.0052179028280079365, + 0.00655916053801775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.009787309914827347 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987580597400665 + ], + "opposite_logprob_tokens": [ + -1.8048758506774902 + ], + "logprob_reward": -0.17987580597400665, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987580597400665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13630591332912445, + 0.04497763141989708, + 0.06432356685400009, + 0.11447370797395706, + 0.26589158177375793, + 0.12344050407409668, + 0.0935901328921318, + 0.10413314402103424, + 0.03180266544222832, + 0.021061187610030174 + ], + "credit_map": [ + 0.015828337520360947, + 0.012856248766183853, + 0.023017562925815582, + 0.018375305458903313, + 0.012533365748822689, + 0.020112695172429085, + 0.014813641086220741, + 0.018768753856420517, + 0.006393361371010542, + 0.005383004900068045, + 0.0036787583958357573, + 0.0061086625792086124, + 0.0041620912961661816, + 0.006730571389198303, + 0.0051835933700203896, + 0.007337586954236031, + 0.007419355679303408, + 0.007228586822748184, + 0.006297512445598841, + 0.01017675269395113, + 0.007263611536473036, + 0.00832707341760397, + 0.008307364769279957, + 0.009303313679993153, + 0.018300198018550873, + 0.007508696988224983, + 0.01641150936484337, + 0.011506499722599983, + 0.006394950207322836, + 0.02232796512544155, + 0.009337890893220901, + 0.02268599532544613, + 0.03927187621593475, + 0.020730633288621902, + 0.035911086946725845, + 0.02519557625055313, + 0.024388914927840233, + 0.04538699612021446, + 0.02545846439898014, + 0.04954802617430687, + 0.012919383123517036, + 0.013604198582470417, + 0.016224896535277367, + 0.019439686089754105, + 0.008252532221376896, + 0.023727450519800186, + 0.012616162188351154, + 0.016656192019581795, + 0.019159920513629913, + 0.004693188238888979, + 0.018731674179434776, + 0.0063953339122235775, + 0.005016565788537264, + 0.014230897650122643, + 0.0059502446092665195, + 0.019412308931350708, + 0.015808401629328728, + 0.007761884015053511, + 0.017184995114803314, + 0.0062665739096701145, + 0.006319040898233652, + 0.015785424038767815, + 0.007124839350581169, + 0.027881983667612076, + 0.007389927748590708, + 0.0020783625077456236, + 0.004480901639908552, + 0.004798525013029575, + 0.0024797613732516766, + 0.0031650359742343426, + 0.003559670876711607, + 0.003850480541586876, + 0.0042204479686915874, + 0.0022522981744259596, + 0.0029528227169066668, + 0.0026718126609921455, + 0.0021240473724901676, + 0.0017305550863966346, + 0.002163215074688196, + 0.0029459877405315638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.17987580597400665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263367909938097 + ], + "opposite_logprob_tokens": [ + -3.651336908340454 + ], + "logprob_reward": -0.0263367909938097, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263367909938097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021589679643511772, + 0.028294391930103302, + 0.03903861716389656, + 0.03810323029756546, + 0.15878601372241974, + 0.06599200516939163, + 0.05623168498277664, + 0.182337686419487, + 0.09685221314430237, + 0.31277450919151306 + ], + "credit_map": [ + 0.0024682332295924425, + 0.0021658577024936676, + 0.0023874654434621334, + 0.0036147451028227806, + 0.002231053775176406, + 0.0030112676322460175, + 0.0025503644719719887, + 0.0031606925185769796, + 0.005512773524969816, + 0.0031428239308297634, + 0.002437272109091282, + 0.0026170879136770964, + 0.0031125457026064396, + 0.0035805983934551477, + 0.002791689708828926, + 0.0050995987839996815, + 0.004459783900529146, + 0.003261801553890109, + 0.0038339977618306875, + 0.004145466256886721, + 0.0035980104003101587, + 0.007903017103672028, + 0.006193533539772034, + 0.005643008276820183, + 0.005749059841036797, + 0.0022999870125204325, + 0.006654292810708284, + 0.0036392505280673504, + 0.001873251167126, + 0.008857330307364464, + 0.002648010617122054, + 0.006382045801728964, + 0.014997852966189384, + 0.0182785764336586, + 0.023688340559601784, + 0.021575966849923134, + 0.0148748978972435, + 0.025002067908644676, + 0.020281262695789337, + 0.02008702978491783, + 0.007094208151102066, + 0.006148563232272863, + 0.009807481430470943, + 0.006137664429843426, + 0.004913173615932465, + 0.018146509304642677, + 0.005095137748867273, + 0.008649264462292194, + 0.007738202344626188, + 0.004333599470555782, + 0.008049304597079754, + 0.006061719264835119, + 0.0038123249541968107, + 0.010751924477517605, + 0.005933139473199844, + 0.009551470167934895, + 0.028101656585931778, + 0.013828512281179428, + 0.025740766897797585, + 0.01543382741510868, + 0.012429822236299515, + 0.03163798525929451, + 0.013084079138934612, + 0.042081043124198914, + 0.017056282609701157, + 0.007971095852553844, + 0.014247133396565914, + 0.01602017879486084, + 0.007075671572238207, + 0.012085504829883575, + 0.010973037220537663, + 0.011423307470977306, + 0.06154080480337143, + 0.03345701843500137, + 0.04930160194635391, + 0.024916667491197586, + 0.025621412321925163, + 0.03561596944928169, + 0.029800400137901306, + 0.05252063274383545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.0263367909938097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002814263105392456 + ], + "opposite_logprob_tokens": [ + -5.877814292907715 + ], + "logprob_reward": -0.002814263105392456, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002814263105392456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08501484990119934, + 0.05970189720392227, + 0.09677338600158691, + 0.12681812047958374, + 0.07079091668128967, + 0.3212430477142334, + 0.059980541467666626, + 0.06300461292266846, + 0.05354814976453781, + 0.06312445551156998 + ], + "credit_map": [ + 0.007847829721868038, + 0.0062898448668420315, + 0.01380021870136261, + 0.011651784181594849, + 0.008006582036614418, + 0.014726325869560242, + 0.007368598598986864, + 0.015323670580983162, + 0.010310455225408077, + 0.005631104577332735, + 0.005768618080765009, + 0.00573714217171073, + 0.005394979380071163, + 0.008114119060337543, + 0.008763072080910206, + 0.009982406161725521, + 0.012756112031638622, + 0.007841016165912151, + 0.011316773481667042, + 0.00962008349597454, + 0.009153495542705059, + 0.017562875524163246, + 0.014355449937283993, + 0.014167570509016514, + 0.015195436775684357, + 0.006981146056205034, + 0.02069387026131153, + 0.01282739918678999, + 0.0075010815635323524, + 0.028242960572242737, + 0.014848574995994568, + 0.020527660846710205, + 0.00667440053075552, + 0.007814238779246807, + 0.011666991747915745, + 0.008310997858643532, + 0.0072873057797551155, + 0.011539812199771404, + 0.008524475619196892, + 0.008972691372036934, + 0.05059920996427536, + 0.021417710930109024, + 0.05844956636428833, + 0.03066643700003624, + 0.01544590387493372, + 0.08260606974363327, + 0.021110782399773598, + 0.040947336703538895, + 0.008025960996747017, + 0.00527441781014204, + 0.007876361720263958, + 0.005952068138867617, + 0.0036174168344587088, + 0.015702400356531143, + 0.004975253250449896, + 0.008556668646633625, + 0.010185880586504936, + 0.004278908018022776, + 0.010307950899004936, + 0.006055172067135572, + 0.003973054233938456, + 0.007107819430530071, + 0.0059052384458482265, + 0.01519058644771576, + 0.009562002494931221, + 0.00461158249527216, + 0.0078849196434021, + 0.008300461806356907, + 0.004063401371240616, + 0.005857732146978378, + 0.005614428780972958, + 0.0076536210253834724, + 0.01380270253866911, + 0.00516977533698082, + 0.00846121646463871, + 0.005970020312815905, + 0.003649067832157016, + 0.0077667986042797565, + 0.009956181049346924, + 0.008348693139851093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.002814263105392456 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up shot in the style of a nature documentary, the camera, in a slow, smooth orbiting motion, rotates 360 degrees around a helmet jellyfish. Its nearly transparent bell, sculpted from crystal, emanates a faint, ghostly, pale-blue bioluminescence against the inky blackness of the deep sea, revealing the intricate structure within. The ghostly creature gracefully and rhythmically contracts and expands its bell, each pulse leisurely propelling it upward through the endless darkness. Descending from its base, delicate, thread-like tentacles float and dance softly behind it like a dreamy veil. The background is a bottomless inky black, shimmering only in the soft beam of the photographic light, tiny particles of marine snow, like dust motes, further accentuating the jellyfish's solitary and otherworldly presence. The entire rotating lens captures every dreamlike moment in astonishing detail, vividly displaying the fragile beauty and mystery of this deep-sea life, creating a tranquil, ethereal and awe-inspiring atmosphere.", + "seed": 14672, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_5/index_8535_seed_14672_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0822245879098773, + "vlm_logprob": -0.0822245879098773, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.0822245879098773, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028559232130646706 + ], + "opposite_logprob_tokens": [ + -3.6535592079162598 + ], + "logprob_reward": -0.028559232130646706, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028559232130646706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14502984285354614, + 0.3005298674106598, + 0.03541450574994087, + 0.11585773527622223, + 0.1067570298910141, + 0.12938950955867767, + 0.0381203256547451, + 0.04654337465763092, + 0.028708208352327347, + 0.05364959314465523 + ], + "credit_map": [ + 0.013007091358304024, + 0.01484206784516573, + 0.022723328322172165, + 0.013978465460240841, + 0.014240632764995098, + 0.02416127547621727, + 0.022903859615325928, + 0.019173122942447662, + 0.021247562021017075, + 0.01665617525577545, + 0.029630480334162712, + 0.013876689597964287, + 0.015128303319215775, + 0.09149530529975891, + 0.018644316121935844, + 0.09385105967521667, + 0.004153780173510313, + 0.002925627864897251, + 0.005064505152404308, + 0.00341202225536108, + 0.00304105831310153, + 0.006097271107137203, + 0.004603356122970581, + 0.006116883829236031, + 0.008017687126994133, + 0.00935500767081976, + 0.020218990743160248, + 0.00924656167626381, + 0.01254724059253931, + 0.02406376786530018, + 0.012081295251846313, + 0.020327186211943626, + 0.010632379911839962, + 0.009421896189451218, + 0.012202822603285313, + 0.012599611654877663, + 0.013280619867146015, + 0.012122501619160175, + 0.016619054600596428, + 0.019878143444657326, + 0.015716491267085075, + 0.013841262087225914, + 0.023652922362089157, + 0.011353274807333946, + 0.015446965582668781, + 0.017898190766572952, + 0.010096855461597443, + 0.02138354443013668, + 0.004353880882263184, + 0.004397394135594368, + 0.004795270971953869, + 0.003843714250251651, + 0.006332009565085173, + 0.003610788844525814, + 0.0053416104055941105, + 0.005445657763630152, + 0.004462016746401787, + 0.0054053012281656265, + 0.005833997391164303, + 0.007702427916228771, + 0.007399571593850851, + 0.005009419750422239, + 0.006374753545969725, + 0.004355886485427618, + 0.00599287124350667, + 0.002143976977095008, + 0.006090342998504639, + 0.0027227806858718395, + 0.0019470116822049022, + 0.0024378749076277018, + 0.002099092584103346, + 0.005274258088320494, + 0.005656779743731022, + 0.006704621948301792, + 0.007177285384386778, + 0.0041023255325853825, + 0.006234494037926197, + 0.006166309118270874, + 0.01100669614970684, + 0.006601082626730204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.028559232130646706 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067454250529408455 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.0067454250529408455, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067454250529408455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12727399170398712, + 0.1613326221704483, + 0.1330137699842453, + 0.08994530886411667, + 0.11467038840055466, + 0.08746074140071869, + 0.04785725846886635, + 0.11110670119524002, + 0.048996537923812866, + 0.07834260165691376 + ], + "credit_map": [ + 0.01674402318894863, + 0.013557408936321735, + 0.02081160433590412, + 0.010241972282528877, + 0.010244912467896938, + 0.020278630778193474, + 0.011819855310022831, + 0.023575596511363983, + 0.009201512672007084, + 0.00827670469880104, + 0.01838635466992855, + 0.015731224790215492, + 0.010690775699913502, + 0.04599475488066673, + 0.017176998779177666, + 0.03587431088089943, + 0.01109919510781765, + 0.010289279744029045, + 0.013174244202673435, + 0.008695893920958042, + 0.014056636020541191, + 0.028452524915337563, + 0.02115965075790882, + 0.026086363941431046, + 0.010727521032094955, + 0.005461344961076975, + 0.01448556873947382, + 0.004381400533020496, + 0.01168256439268589, + 0.018081609159708023, + 0.007872016169130802, + 0.017253287136554718, + 0.014850500971078873, + 0.010398549027740955, + 0.020751018077135086, + 0.012483000755310059, + 0.01228679995983839, + 0.014034124091267586, + 0.016308071091771126, + 0.013558336533606052, + 0.012115688994526863, + 0.009273399598896503, + 0.01216930989176035, + 0.009059269912540913, + 0.007522511761635542, + 0.013286293484270573, + 0.008298387750983238, + 0.015735888853669167, + 0.006841456051915884, + 0.0033906016033142805, + 0.008017541840672493, + 0.0063817501068115234, + 0.004018682986497879, + 0.005114918574690819, + 0.00397524144500494, + 0.010117070749402046, + 0.013486574403941631, + 0.00782737135887146, + 0.01659419946372509, + 0.02752341330051422, + 0.008428128436207771, + 0.011227935552597046, + 0.014684068039059639, + 0.011335021816194057, + 0.01027342677116394, + 0.004495448432862759, + 0.009357841685414314, + 0.004214624408632517, + 0.004415060859173536, + 0.004857710096985102, + 0.0037668920122087, + 0.007615537848323584, + 0.010162397287786007, + 0.007950111292302608, + 0.010958703234791756, + 0.009086755104362965, + 0.0072887009009718895, + 0.007884843274950981, + 0.013552631251513958, + 0.011458472348749638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0067454250529408455 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987950146198273 + ], + "opposite_logprob_tokens": [ + -1.8048795461654663 + ], + "logprob_reward": -0.17987950146198273, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987950146198273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06661184132099152, + 0.16057930886745453, + 0.047954875975847244, + 0.10834857076406479, + 0.09302522987127304, + 0.06259380280971527, + 0.03443827107548714, + 0.04371950775384903, + 0.039695486426353455, + 0.34303316473960876 + ], + "credit_map": [ + 0.006132218986749649, + 0.006080829072743654, + 0.009296138770878315, + 0.00797339715063572, + 0.0061341761611402035, + 0.011621481738984585, + 0.008360804989933968, + 0.011012784205377102, + 0.01057675015181303, + 0.008135426789522171, + 0.018671968951821327, + 0.017777668312191963, + 0.014031878672540188, + 0.03948279842734337, + 0.021329443901777267, + 0.03057335689663887, + 0.004468918312340975, + 0.0061867376789450645, + 0.006747061852365732, + 0.0047887819819152355, + 0.004727794323116541, + 0.007939599454402924, + 0.007281286641955376, + 0.005814694333821535, + 0.009917364455759525, + 0.007062543649226427, + 0.022004326805472374, + 0.00844945851713419, + 0.009766657836735249, + 0.01759893074631691, + 0.01347433589398861, + 0.020074943080544472, + 0.008190940134227276, + 0.006893820129334927, + 0.01575237326323986, + 0.009741517715156078, + 0.009982069954276085, + 0.009965863078832626, + 0.011970123276114464, + 0.02052852138876915, + 0.008699046447873116, + 0.004737901035696268, + 0.010875746607780457, + 0.0035482028033584356, + 0.006945094559341669, + 0.008973854593932629, + 0.004787537269294262, + 0.014026413671672344, + 0.003870159387588501, + 0.00391368567943573, + 0.004333364777266979, + 0.003368505509570241, + 0.005173846613615751, + 0.0040019783191382885, + 0.004010465461760759, + 0.005766259040683508, + 0.004670127760618925, + 0.0034861513413488865, + 0.0055988747626543045, + 0.007341776974499226, + 0.006070724222809076, + 0.004261794965714216, + 0.007030216045677662, + 0.005259837955236435, + 0.007007387932389975, + 0.0037092622369527817, + 0.007241074927151203, + 0.006621044594794512, + 0.0029187428299337626, + 0.0033065893221646547, + 0.0037947185337543488, + 0.005096661392599344, + 0.07000908255577087, + 0.01887463591992855, + 0.07786823809146881, + 0.01810901053249836, + 0.02196560986340046, + 0.0301115233451128, + 0.031511832028627396, + 0.07458321750164032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.17987950146198273 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987780272960663 + ], + "opposite_logprob_tokens": [ + -1.804877758026123 + ], + "logprob_reward": -0.17987780272960663, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987780272960663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10501499474048615, + 0.1622774600982666, + 0.06035006791353226, + 0.2069222331047058, + 0.1251319944858551, + 0.041469231247901917, + 0.04819208011031151, + 0.06678327172994614, + 0.04436526447534561, + 0.1394934058189392 + ], + "credit_map": [ + 0.01072277408093214, + 0.01074158400297165, + 0.018041037023067474, + 0.008048293180763721, + 0.011390874162316322, + 0.018696030601859093, + 0.014687403105199337, + 0.012686999514698982, + 0.012720604427158833, + 0.012478257529437542, + 0.020144423469901085, + 0.02069045417010784, + 0.022093800827860832, + 0.0254293754696846, + 0.026128845289349556, + 0.022591697052121162, + 0.004210180137306452, + 0.007028022315353155, + 0.009842843748629093, + 0.005409810692071915, + 0.006416034419089556, + 0.012228710576891899, + 0.006924733053892851, + 0.008289732970297337, + 0.01929599978029728, + 0.01162418257445097, + 0.03125395625829697, + 0.00923957023769617, + 0.02493695169687271, + 0.05047918111085892, + 0.01770716719329357, + 0.04238521307706833, + 0.0118997972458601, + 0.011694534681737423, + 0.011680237017571926, + 0.016196835786104202, + 0.017756089568138123, + 0.01617559790611267, + 0.02094944380223751, + 0.01877945102751255, + 0.0038442625664174557, + 0.0036672044079750776, + 0.00651082769036293, + 0.0023986087180674076, + 0.010676752775907516, + 0.005277208983898163, + 0.005265319719910622, + 0.0038290475495159626, + 0.00743827922269702, + 0.004004737827926874, + 0.008001431822776794, + 0.0032371568959206343, + 0.005269899033010006, + 0.0037972002755850554, + 0.004565896932035685, + 0.011877474375069141, + 0.009784143418073654, + 0.006093951873481274, + 0.008535295724868774, + 0.007474833633750677, + 0.01054952759295702, + 0.006395281292498112, + 0.008456177078187466, + 0.009494057856500149, + 0.008411282673478127, + 0.003861803561449051, + 0.008602140471339226, + 0.005085803102701902, + 0.0038170989137142897, + 0.0037307054735720158, + 0.0037086119409650564, + 0.007147819269448519, + 0.017957523465156555, + 0.01420485321432352, + 0.023240596055984497, + 0.01025957241654396, + 0.013549608178436756, + 0.01252426765859127, + 0.021661795675754547, + 0.026095183566212654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.17987780272960663 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a helmet jellyfish with a nearly transparent, crystal-like bell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the jellyfish emit a faint, ghostly, pale-blue bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow, smooth orbiting motion rotating 360 degrees around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are delicate, thread-like tentacles visible descending from the base of the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background an inky black deep sea containing tiny particles of marine snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016060978174209595 + ], + "opposite_logprob_tokens": [ + -4.141060829162598 + ], + "logprob_reward": -0.016060978174209595, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016060978174209595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22946105897426605, + 0.21841688454151154, + 0.03178815543651581, + 0.11346463859081268, + 0.07604524493217468, + 0.07446792721748352, + 0.08074500411748886, + 0.028857950121164322, + 0.0981447845697403, + 0.048608288168907166 + ], + "credit_map": [ + 0.025860469788312912, + 0.017541436478495598, + 0.05077291280031204, + 0.01688770018517971, + 0.01373396348208189, + 0.04323799908161163, + 0.025896061211824417, + 0.03553052991628647, + 0.014502295292913914, + 0.011601420119404793, + 0.020901378244161606, + 0.011053487658500671, + 0.011969051323831081, + 0.06994741410017014, + 0.01248565036803484, + 0.06595618277788162, + 0.003537840209901333, + 0.0035210340283811092, + 0.0071602072566747665, + 0.003122571622952819, + 0.0031690967734903097, + 0.005537068471312523, + 0.00309878122061491, + 0.002641559112817049, + 0.016987783834338188, + 0.006193909794092178, + 0.023220334202051163, + 0.0069768899120390415, + 0.009238176979124546, + 0.023211343213915825, + 0.009304321371018887, + 0.01833188906311989, + 0.006790520157665014, + 0.004223369527608156, + 0.022370794788002968, + 0.00546016963198781, + 0.006427299696952105, + 0.005493822041898966, + 0.006491570267826319, + 0.018787700682878494, + 0.008795985020697117, + 0.00678757531568408, + 0.01203832309693098, + 0.005378352478146553, + 0.008678458631038666, + 0.009579984471201897, + 0.006571820937097073, + 0.016637422144412994, + 0.013516062870621681, + 0.006887487601488829, + 0.01339386310428381, + 0.006663379725068808, + 0.006898635067045689, + 0.008762743324041367, + 0.005698046647012234, + 0.018924787640571594, + 0.003540657926350832, + 0.0024514503311365843, + 0.0037310244515538216, + 0.003598424606025219, + 0.00424330634996295, + 0.003644840093329549, + 0.004355525132268667, + 0.003292722860351205, + 0.021552899852395058, + 0.008571960031986237, + 0.01857798919081688, + 0.008450505323708057, + 0.009148646146059036, + 0.009837799705564976, + 0.007599704898893833, + 0.014405274763703346, + 0.006693494971841574, + 0.0034458443988114595, + 0.008825610391795635, + 0.00494723953306675, + 0.003930495120584965, + 0.004831474274396896, + 0.006905355490744114, + 0.009028772823512554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.016060978174209595 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_6/index_tensor([6544, 6544])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_6/index_tensor([6544, 6544])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5221a189f7908059bf59425ccd024cfed3b8099e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_6/index_tensor([6544, 6544])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a perfectly locked-off, medium close-up shot, a vibrant mantis shrimp hovers quietly in crystal-clear tropical waters. Its segmented carapace, clad in iridescent emerald, sapphire blue, and vibrant orange, shimmers like porcelain under the bright sunlight filtering through the water's surface, creating swaying caustics. Its body is slightly curled, its prominent, independently movable compound eyes warily observing the horizon. Its hammer-like raptors (predatory appendages) are tucked tightly beneath its body, like a fully wound spring, gathering terrifying power ready to explode. Suddenly, the entire scene explodes with Bones Studio's signature high-energy, fluid animation. With a speed barely perceptible to the naked eye, the raptor's limbs, rendered in a blur of colorful afterimages at the extreme frame rate, rip through the tranquil waters. Then, at the front of its impact path, a spherical shockwave, instantly formed from countless tiny, seething cavitation bubbles, violently spread outward—a silent \"howl\"—creating a visual feast in absolute silence. This piercing white ring of energy contrasted sharply with the vibrant neon colors of the mantis shrimp itself. Throughout this process, the completely still camera remained motionless, its extreme stability creating a thrilling dramatic tension with the creature's instantaneous, all-shattering burst of violent power, vividly conveying this momentary physical wonder through animated form.", + "seed": 25752, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_6/index_6544_seed_25752_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2628507365472615, + "vlm_logprob": -0.2628507365472615, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.2628507365472615, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01575932838022709 + ], + "opposite_logprob_tokens": [ + -4.390759468078613 + ], + "logprob_reward": -0.01575932838022709, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01575932838022709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047276370227336884, + 0.21022844314575195, + 0.02270437590777874, + 0.141664057970047, + 0.039258722215890884, + 0.4269266128540039, + 0.014562444761395454, + 0.025153936818242073, + 0.020921476185321808, + 0.05130360275506973 + ], + "credit_map": [ + 0.004640396684408188, + 0.00959040317684412, + 0.003966204822063446, + 0.005433250684291124, + 0.0052559468895196915, + 0.00736765842884779, + 0.004241493996232748, + 0.006781014148145914, + 0.012110995128750801, + 0.006836587563157082, + 0.00358411087654531, + 0.0045764329843223095, + 0.004090643022209406, + 0.08830893039703369, + 0.004240855108946562, + 0.08647987991571426, + 0.003736768616363406, + 0.0020267677027732134, + 0.0014933871570974588, + 0.0019525885581970215, + 0.0020304678473621607, + 0.005025793332606554, + 0.001709104049950838, + 0.0047294981777668, + 0.006446728948503733, + 0.003683223621919751, + 0.003573958296328783, + 0.0031726022716611624, + 0.002797724213451147, + 0.05972561985254288, + 0.0027189168613404036, + 0.05954527109861374, + 0.0033977245911955833, + 0.003516287775710225, + 0.002665376989170909, + 0.0030106964986771345, + 0.0019278539111837745, + 0.011517618782818317, + 0.0020890317391604185, + 0.011134130880236626, + 0.06956429779529572, + 0.03043793886899948, + 0.02840404212474823, + 0.019600074738264084, + 0.03832150623202324, + 0.11853310465812683, + 0.034951742738485336, + 0.08711392432451248, + 0.0013102933298796415, + 0.0011933509958907962, + 0.0019312725635245442, + 0.0016072633443400264, + 0.0010613102931529284, + 0.0030791580211371183, + 0.0014828371349722147, + 0.0028969591949135065, + 0.0027017230167984962, + 0.002216097665950656, + 0.004291298799216747, + 0.0021611296106129885, + 0.004324241541326046, + 0.0037629285361617804, + 0.002633488504216075, + 0.003063029842451215, + 0.0014511346817016602, + 0.0021502377931028605, + 0.0018992711557075381, + 0.0033386547584086657, + 0.00401492603123188, + 0.0016979863867163658, + 0.00481986440718174, + 0.0015494013205170631, + 0.001600516028702259, + 0.0029279214795678854, + 0.0017801994690671563, + 0.0033516506664454937, + 0.01799929328262806, + 0.003309565829113126, + 0.017228683456778526, + 0.003105767536908388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.01575932838022709 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019596675410866737 + ], + "opposite_logprob_tokens": [ + -6.251959800720215 + ], + "logprob_reward": -0.0019596675410866737, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019596675410866737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14793603122234344, + 0.09837646782398224, + 0.025243323296308517, + 0.05697710067033768, + 0.24486465752124786, + 0.1554405391216278, + 0.03172323480248451, + 0.03185177221894264, + 0.0535561703145504, + 0.15403063595294952 + ], + "credit_map": [ + 0.019753240048885345, + 0.02993673086166382, + 0.015717145055532455, + 0.015704356133937836, + 0.017602212727069855, + 0.021045047789812088, + 0.015049141831696033, + 0.013128162361681461, + 0.015422171913087368, + 0.008044242858886719, + 0.006344354245811701, + 0.006276693660765886, + 0.004940863698720932, + 0.026643235236406326, + 0.004492036532610655, + 0.026212869212031364, + 0.0035438493359833956, + 0.0024777196813374758, + 0.0017667284701019526, + 0.0018719499930739403, + 0.001943737966939807, + 0.005343473516404629, + 0.002756027039140463, + 0.005539839155972004, + 0.004292221739888191, + 0.002304665045812726, + 0.003874109825119376, + 0.003019919153302908, + 0.0037645448464900255, + 0.01869174651801586, + 0.003841121681034565, + 0.017188772559165955, + 0.016286877915263176, + 0.01702665165066719, + 0.010974040254950523, + 0.01555234007537365, + 0.010107209905982018, + 0.0834885686635971, + 0.010497561655938625, + 0.08093142509460449, + 0.024833688512444496, + 0.011100759729743004, + 0.00981872621923685, + 0.006659160368144512, + 0.011959780007600784, + 0.04515688493847847, + 0.010769319720566273, + 0.03514222428202629, + 0.003188340924680233, + 0.0026385970413684845, + 0.003280786331743002, + 0.003463226370513439, + 0.005488547496497631, + 0.00541600352153182, + 0.0038475326728075743, + 0.004400202073156834, + 0.0044918302446603775, + 0.003084886120632291, + 0.004908194299787283, + 0.003738217521458864, + 0.0036303966771811247, + 0.0035525215789675713, + 0.005392838269472122, + 0.003052887972444296, + 0.004892564844340086, + 0.006227575242519379, + 0.005710785277187824, + 0.00897863507270813, + 0.007739092223346233, + 0.00488946121186018, + 0.008919649757444859, + 0.006198413670063019, + 0.004113071132451296, + 0.009095026180148125, + 0.004662416409701109, + 0.009182608686387539, + 0.05804906040430069, + 0.008931208401918411, + 0.052755579352378845, + 0.007241679821163416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.0019596675410866737 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005627453327179 + ], + "opposite_logprob_tokens": [ + -2.695056200027466 + ], + "logprob_reward": -0.07005627453327179, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005627453327179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04508361965417862, + 0.48741650581359863, + 0.04643412306904793, + 0.1182485818862915, + 0.11055217683315277, + 0.062233418226242065, + 0.026163622736930847, + 0.023005208000540733, + 0.016478775069117546, + 0.06438394635915756 + ], + "credit_map": [ + 0.004332251846790314, + 0.004742676857858896, + 0.0032182831782847643, + 0.003611489664763212, + 0.0033348232973366976, + 0.01170377992093563, + 0.0031594326719641685, + 0.010980883613228798, + 0.00919225625693798, + 0.007480912376195192, + 0.004169689957052469, + 0.005687529221177101, + 0.007471810095012188, + 0.22496867179870605, + 0.006962527986615896, + 0.22148314118385315, + 0.006742265075445175, + 0.003554884111508727, + 0.0028904106002300978, + 0.0031194870825856924, + 0.0061297183856368065, + 0.011757886968553066, + 0.0034502875059843063, + 0.008789187297224998, + 0.0041166041046381, + 0.00272791157476604, + 0.0020640003494918346, + 0.0029377012979239225, + 0.00395578658208251, + 0.049378279596567154, + 0.0037550735287368298, + 0.04931323230266571, + 0.01097212452441454, + 0.014085298404097557, + 0.012518096715211868, + 0.010991344228386879, + 0.015500584617257118, + 0.018844889476895332, + 0.010585356503725052, + 0.017054492607712746, + 0.008044502697885036, + 0.00278823752887547, + 0.003300837939605117, + 0.002231298480182886, + 0.003128485754132271, + 0.021045224741101265, + 0.0033441130071878433, + 0.018350724130868912, + 0.002867568051442504, + 0.002207154640927911, + 0.0036967983469367027, + 0.003321757772937417, + 0.0027802863623946905, + 0.003483725478872657, + 0.0034823017194867134, + 0.004324033390730619, + 0.002018885686993599, + 0.002750404179096222, + 0.0018216146854683757, + 0.003388305427506566, + 0.0038411992136389017, + 0.0038188935723155737, + 0.002784256124868989, + 0.002581650158390403, + 0.0016747021581977606, + 0.0025167418643832207, + 0.002209620550274849, + 0.00209549511782825, + 0.002862925874069333, + 0.0012970591196790338, + 0.0020279698073863983, + 0.0017942605772987008, + 0.0016388647491112351, + 0.003810502588748932, + 0.0018056448316201568, + 0.0033977979328483343, + 0.024665653705596924, + 0.0041068256832659245, + 0.021601568907499313, + 0.0033570886589586735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.07005627453327179 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819731365889311 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819731365889311, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819731365889311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03406193479895592, + 0.20110996067523956, + 0.0093027139082551, + 0.058832284063100815, + 0.043415676802396774, + 0.15542320907115936, + 0.028158338740468025, + 0.022750483825802803, + 0.04239928722381592, + 0.40454617142677307 + ], + "credit_map": [ + 0.0031558906193822622, + 0.00621592253446579, + 0.0023784632794559, + 0.0027247644029557705, + 0.002208374673500657, + 0.007601508405059576, + 0.0025873230770230293, + 0.007189688738435507, + 0.004438949748873711, + 0.0034430043306201696, + 0.0024681405629962683, + 0.0022543463855981827, + 0.003139226231724024, + 0.09184275567531586, + 0.0030104827601462603, + 0.09051303565502167, + 0.0010464498773217201, + 0.0008889689925126731, + 0.0017351889982819557, + 0.0009146055090241134, + 0.0008312544669024646, + 0.0013417568989098072, + 0.0011476236395537853, + 0.0013968658167868853, + 0.0026534695643931627, + 0.0017065276624634862, + 0.0014824745012447238, + 0.001677875523455441, + 0.0019356702687218785, + 0.023484379053115845, + 0.002357863588258624, + 0.023534024134278297, + 0.004924848210066557, + 0.006413940340280533, + 0.004282752051949501, + 0.004722031764686108, + 0.004030441399663687, + 0.008363517932593822, + 0.002811481012031436, + 0.007866663858294487, + 0.028271639719605446, + 0.010567446239292622, + 0.011715688742697239, + 0.006468173116445541, + 0.0110597750172019, + 0.04257844388484955, + 0.01178264431655407, + 0.03297939524054527, + 0.003544526407495141, + 0.002197186928242445, + 0.006147780455648899, + 0.0028825760819017887, + 0.002419453114271164, + 0.004139802418649197, + 0.0028188233263790607, + 0.004008189775049686, + 0.002865527756512165, + 0.0024728255812078714, + 0.003825276391580701, + 0.0027724129613488913, + 0.003963393159210682, + 0.0021919780410826206, + 0.002776649547740817, + 0.0018824213184416294, + 0.0028470605611801147, + 0.004472186788916588, + 0.003752024844288826, + 0.006140321027487516, + 0.010249815881252289, + 0.003962178714573383, + 0.006440498400479555, + 0.0045351991429924965, + 0.009438588283956051, + 0.022397557273507118, + 0.005369165912270546, + 0.018081212416291237, + 0.16255609691143036, + 0.021046586334705353, + 0.15078328549861908, + 0.014873680658638477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.002819731365889311 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236586809158325 + ], + "opposite_logprob_tokens": [ + -0.3486586809158325 + ], + "logprob_reward": -1.2236586809158325, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236586809158325, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07775189727544785, + 0.12016132473945618, + 0.010399287566542625, + 0.019361181184649467, + 0.06573323905467987, + 0.12030336260795593, + 0.03331159055233002, + 0.033474668860435486, + 0.042335156351327896, + 0.47716835141181946 + ], + "credit_map": [ + 0.005976862274110317, + 0.00725302379578352, + 0.0035740926396101713, + 0.004431535955518484, + 0.004765349440276623, + 0.023944202810525894, + 0.003686476731672883, + 0.024120349436998367, + 0.006976587697863579, + 0.003083490766584873, + 0.0034400231670588255, + 0.0019981632940471172, + 0.002089581685140729, + 0.050588853657245636, + 0.00215827114880085, + 0.04982633888721466, + 0.0012093993136659265, + 0.001082119531929493, + 0.0008933206554502249, + 0.0010968190617859364, + 0.0011540282284840941, + 0.002166495192795992, + 0.0012054442195221782, + 0.0015916606644168496, + 0.0021250962745398283, + 0.001836599432863295, + 0.0012535123387351632, + 0.0019807738717645407, + 0.002551293931901455, + 0.003847668180242181, + 0.0021044500172138214, + 0.003661787137389183, + 0.008036279119551182, + 0.006344500929117203, + 0.005207494832575321, + 0.004806619603186846, + 0.003830504138022661, + 0.017239617183804512, + 0.003437984036281705, + 0.016830233857035637, + 0.00931535568088293, + 0.004837047774344683, + 0.004568829201161861, + 0.00448843277990818, + 0.008296349085867405, + 0.0445404089987278, + 0.006202781107276678, + 0.038054145872592926, + 0.0034304356668144464, + 0.0030856127850711346, + 0.003883136436343193, + 0.0037199801299721003, + 0.004175631795078516, + 0.005971536040306091, + 0.0034473666455596685, + 0.005597887095063925, + 0.003823588602244854, + 0.0042931195348501205, + 0.003588438965380192, + 0.004577252548187971, + 0.005042314529418945, + 0.0035788994282484055, + 0.0053313723765313625, + 0.0032396793831139803, + 0.0034189040306955576, + 0.003787883324548602, + 0.005374513100832701, + 0.007621204014867544, + 0.008077782578766346, + 0.00287402025423944, + 0.007785481866449118, + 0.003395362291485071, + 0.010844280943274498, + 0.025567110627889633, + 0.0063767689280211926, + 0.021482471376657486, + 0.19255772233009338, + 0.023675037547945976, + 0.1786801517009735, + 0.01798475719988346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -1.2236586809158325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In a perfectly locked-off, medium close-up shot, a vibrant mantis shrimp hovers quietly in crystal-clear tropical waters. Its segmented carapace, clad in iridescent emerald, sapphire blue, and vibrant orange, shimmers like porcelain under the bright sunlight filtering through the water's surface, creating swaying caustics. Its body is slightly curled, its prominent, independently movable compound eyes warily observing the horizon. Its hammer-like raptors (predatory appendages) are tucked tightly beneath its body, like a fully wound spring, gathering terrifying power ready to explode. Suddenly, the entire scene explodes with Bones Studio's signature high-energy, fluid animation. With a speed barely perceptible to the naked eye, the raptor's limbs, rendered in a blur of colorful afterimages at the extreme frame rate, rip through the tranquil waters. Then, at the front of its impact path, a spherical shockwave, instantly formed from countless tiny, seething cavitation bubbles, violently spread outward—a silent \"howl\"—creating a visual feast in absolute silence. This piercing white ring of energy contrasted sharply with the vibrant neon colors of the mantis shrimp itself. Throughout this process, the completely still camera remained motionless, its extreme stability creating a thrilling dramatic tension with the creature's instantaneous, all-shattering burst of violent power, vividly conveying this momentary physical wonder through animated form.", + "seed": 25753, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_6/index_6544_seed_25753_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.30999837862327695, + "vlm_logprob": -0.30999837862327695, + "vlm_yes_no_margin": 3.649999976158142, + "vlm_token_credit": -0.30999837862327695, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013012022711336613 + ], + "opposite_logprob_tokens": [ + -4.638011932373047 + ], + "logprob_reward": -0.013012022711336613, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013012022711336613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40381407737731934, + 0.05987919867038727, + 0.04276830703020096, + 0.05854996293783188, + 0.07283328473567963, + 0.07766430079936981, + 0.13060660660266876, + 0.07742512971162796, + 0.038377098739147186, + 0.038082052022218704 + ], + "credit_map": [ + 0.050869766622781754, + 0.03913145139813423, + 0.04722109064459801, + 0.04679033160209656, + 0.03306673467159271, + 0.05721669644117355, + 0.038677047938108444, + 0.09084093570709229, + 0.004259053617715836, + 0.008532273583114147, + 0.005914722569286823, + 0.00625198520720005, + 0.006896077189594507, + 0.010043147951364517, + 0.008267276920378208, + 0.009714656509459019, + 0.005088496953248978, + 0.007024641614407301, + 0.004618917126208544, + 0.004562593996524811, + 0.0060254349373281, + 0.005214596167206764, + 0.006475325673818588, + 0.0037582956720143557, + 0.010707173496484756, + 0.013562891632318497, + 0.004032684490084648, + 0.006502693053334951, + 0.004006429109722376, + 0.00755822379142046, + 0.007586812134832144, + 0.004593051970005035, + 0.014581053517758846, + 0.01525250356644392, + 0.007477154955267906, + 0.009234954603016376, + 0.006115224678069353, + 0.005175045691430569, + 0.01058290246874094, + 0.004414432682096958, + 0.013232727535068989, + 0.011069465428590775, + 0.0036843090783804655, + 0.011721418239176273, + 0.003241159487515688, + 0.002600722946226597, + 0.028503015637397766, + 0.0036114687100052834, + 0.008359349332749844, + 0.010390380397439003, + 0.005239978898316622, + 0.042209066450595856, + 0.006460145115852356, + 0.008114233613014221, + 0.04100188985466957, + 0.008831556886434555, + 0.011198909021914005, + 0.0156540609896183, + 0.005501076579093933, + 0.008402268402278423, + 0.007363974116742611, + 0.007155011408030987, + 0.012450541369616985, + 0.009699285961687565, + 0.005271046422421932, + 0.005241774953901768, + 0.003635478438809514, + 0.005816732533276081, + 0.004731205757707357, + 0.003967282362282276, + 0.0036570981610566378, + 0.006056476850062609, + 0.01023331843316555, + 0.004772781860083342, + 0.0017648044740781188, + 0.0029227964114397764, + 0.004392104689031839, + 0.002208123682066798, + 0.007795310579240322, + 0.0039928085170686245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.013012022711336613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022159088402986526 + ], + "opposite_logprob_tokens": [ + -6.12721586227417 + ], + "logprob_reward": -0.0022159088402986526, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022159088402986526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17086002230644226, + 0.06909539550542831, + 0.08439026027917862, + 0.13179896771907806, + 0.1991007775068283, + 0.08060353994369507, + 0.10810820013284683, + 0.07498235255479813, + 0.04354003444314003, + 0.037520404905080795 + ], + "credit_map": [ + 0.023933500051498413, + 0.01922924816608429, + 0.017612481489777565, + 0.021176006644964218, + 0.016814976930618286, + 0.022884568199515343, + 0.016956666484475136, + 0.03225257620215416, + 0.007469020783901215, + 0.009110248647630215, + 0.006838762201368809, + 0.011948692612349987, + 0.007570590823888779, + 0.010711261071264744, + 0.008168457075953484, + 0.007278363220393658, + 0.0047867330722510815, + 0.0077059827744960785, + 0.007189074531197548, + 0.008663654327392578, + 0.01017759833484888, + 0.013035964220762253, + 0.014139766804873943, + 0.01869148202240467, + 0.01376339327543974, + 0.0446450375020504, + 0.00435436749830842, + 0.006130060646682978, + 0.006609858013689518, + 0.04174203798174858, + 0.007410718593746424, + 0.00714349839836359, + 0.02234135940670967, + 0.03463110700249672, + 0.025194475427269936, + 0.024627983570098877, + 0.023019248619675636, + 0.021719181910157204, + 0.03107873722910881, + 0.016488682478666306, + 0.021242717280983925, + 0.02536826953291893, + 0.004807749763131142, + 0.007870827801525593, + 0.004684416111558676, + 0.0053549292497336864, + 0.00622144341468811, + 0.005053186323493719, + 0.009070981293916702, + 0.012355130165815353, + 0.007197254337370396, + 0.026618946343660355, + 0.008183285593986511, + 0.012023610062897205, + 0.021792735904455185, + 0.010866256430745125, + 0.008940308354794979, + 0.012309395708143711, + 0.005450030788779259, + 0.010084457695484161, + 0.008208041079342365, + 0.008120113983750343, + 0.01028791069984436, + 0.01158209703862667, + 0.00679417047649622, + 0.00582887465134263, + 0.004700939636677504, + 0.006070379167795181, + 0.004184645600616932, + 0.003907694015651941, + 0.005071096122264862, + 0.0069822389632463455, + 0.011024023406207561, + 0.004995603580027819, + 0.002660489873960614, + 0.0028389012441039085, + 0.003993648570030928, + 0.002404189435765147, + 0.0063399081118404865, + 0.0032636416144669056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0022159088402986526 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708815932273865 + ], + "opposite_logprob_tokens": [ + -2.1270880699157715 + ], + "logprob_reward": -0.12708815932273865, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12708815932273865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0907401293516159, + 0.11956721544265747, + 0.07337279617786407, + 0.373821884393692, + 0.0260009802877903, + 0.0508996844291687, + 0.1622241735458374, + 0.03620418161153793, + 0.039155807346105576, + 0.028013145551085472 + ], + "credit_map": [ + 0.008864232338964939, + 0.007771146949380636, + 0.010450482368469238, + 0.012747466564178467, + 0.009500720538198948, + 0.013042572885751724, + 0.014225823804736137, + 0.01413767784833908, + 0.008544733747839928, + 0.015095071867108345, + 0.015231767669320107, + 0.013684363104403019, + 0.012908653356134892, + 0.02061123587191105, + 0.016231514513492584, + 0.017259882763028145, + 0.0063161225989460945, + 0.006276048719882965, + 0.006690490059554577, + 0.00733503233641386, + 0.011067423038184643, + 0.010402947664260864, + 0.011054761707782745, + 0.014229964464902878, + 0.05159580707550049, + 0.141518697142601, + 0.015356653369963169, + 0.018004870042204857, + 0.018863976001739502, + 0.09482606500387192, + 0.017779316753149033, + 0.01587647944688797, + 0.004239805042743683, + 0.003571022767573595, + 0.002722516655921936, + 0.004103983286768198, + 0.004260776098817587, + 0.002548441756516695, + 0.0028702043928205967, + 0.0016842311015352607, + 0.009781801141798496, + 0.01201652456074953, + 0.002122897654771805, + 0.006942497566342354, + 0.0038814640138298273, + 0.0037211186718195677, + 0.008504713885486126, + 0.003928667865693569, + 0.02134157344698906, + 0.023314496502280235, + 0.012252837419509888, + 0.025300737470388412, + 0.013495050370693207, + 0.01650898903608322, + 0.027175888419151306, + 0.02283460460603237, + 0.0045287287794053555, + 0.00730863818898797, + 0.0025741183198988438, + 0.00491977296769619, + 0.003677189815789461, + 0.0024028222542256117, + 0.006620215717703104, + 0.004172694869339466, + 0.007980979979038239, + 0.004751267377287149, + 0.0030523850582540035, + 0.00525255361571908, + 0.004649248905479908, + 0.003918211441487074, + 0.004032912198454142, + 0.005518247373402119, + 0.005339937284588814, + 0.0036002842243760824, + 0.0027863001450896263, + 0.0026858854107558727, + 0.003015226684510708, + 0.0016721005085855722, + 0.005867075640708208, + 0.003046334721148014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.12708815932273865 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013372297398746014 + ], + "opposite_logprob_tokens": [ + -6.626337051391602 + ], + "logprob_reward": -0.0013372297398746014, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013372297398746014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12299700826406479, + 0.056814249604940414, + 0.040208812803030014, + 0.13748662173748016, + 0.06974519789218903, + 0.09856506437063217, + 0.22185346484184265, + 0.053639791905879974, + 0.07829365134239197, + 0.1203961968421936 + ], + "credit_map": [ + 0.01623004674911499, + 0.012444640509784222, + 0.010484359227120876, + 0.016204554587602615, + 0.012940947897732258, + 0.016533516347408295, + 0.014451357536017895, + 0.02370757982134819, + 0.007850185967981815, + 0.007988879457116127, + 0.0060644131153821945, + 0.00563872279599309, + 0.004816301632672548, + 0.009191872552037239, + 0.00675754901021719, + 0.008506323210895061, + 0.003983048722147942, + 0.004160034470260143, + 0.003798726247623563, + 0.0037870812229812145, + 0.0038993440102785826, + 0.006908842828124762, + 0.004735433496534824, + 0.008936302736401558, + 0.019865745678544044, + 0.051838651299476624, + 0.003772701369598508, + 0.006237283814698458, + 0.006193060893565416, + 0.03824404999613762, + 0.005190518219023943, + 0.006144619081169367, + 0.00791207142174244, + 0.012015711516141891, + 0.009254039265215397, + 0.008166814222931862, + 0.009103970602154732, + 0.007250927854329348, + 0.009856805205345154, + 0.006184852682054043, + 0.02069181390106678, + 0.02034418284893036, + 0.00766223669052124, + 0.013971904292702675, + 0.005206894129514694, + 0.004330075811594725, + 0.019426055252552032, + 0.006931906100362539, + 0.01360700000077486, + 0.02357850782573223, + 0.009265133179724216, + 0.06536426395177841, + 0.013528233394026756, + 0.01753109320998192, + 0.06376423686742783, + 0.015214984305202961, + 0.00892476737499237, + 0.007499403320252895, + 0.004341354593634605, + 0.0076751708984375, + 0.0058665806427598, + 0.003784748027101159, + 0.009094507433474064, + 0.006453256588429213, + 0.00815537665039301, + 0.011272285133600235, + 0.007796532474458218, + 0.011717822402715683, + 0.00993296131491661, + 0.00839136354625225, + 0.007366272620856762, + 0.013661039061844349, + 0.018457984551787376, + 0.013503956608474255, + 0.014094168320298195, + 0.015351482667028904, + 0.016863500699400902, + 0.010148130357265472, + 0.01653245836496353, + 0.015444514341652393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.0013372297398746014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063385725021362 + ], + "opposite_logprob_tokens": [ + -0.28133854269981384 + ], + "logprob_reward": -1.4063385725021362, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063385725021362, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12591388821601868, + 0.04584193602204323, + 0.038761354982852936, + 0.058931779116392136, + 0.07770263403654099, + 0.2948397696018219, + 0.1719174087047577, + 0.0820041075348854, + 0.05178171396255493, + 0.05230538174510002 + ], + "credit_map": [ + 0.017990196123719215, + 0.013436486944556236, + 0.010645456612110138, + 0.015400523319840431, + 0.013028405606746674, + 0.016942724585533142, + 0.013583785854279995, + 0.02488631010055542, + 0.004007523413747549, + 0.00471559027209878, + 0.00448006484657526, + 0.0037737698294222355, + 0.004706161562353373, + 0.007873247377574444, + 0.008576828055083752, + 0.007708752062171698, + 0.0034178488422185183, + 0.003996013663709164, + 0.0034605939872562885, + 0.0037284432910382748, + 0.0040746149607002735, + 0.006365211680531502, + 0.005134683568030596, + 0.008583943359553814, + 0.010538174770772457, + 0.019565721973776817, + 0.0033389716409146786, + 0.004694403614848852, + 0.004136309027671814, + 0.008258859626948833, + 0.005156341940164566, + 0.003243000479415059, + 0.010906629264354706, + 0.011687039397656918, + 0.009411772713065147, + 0.011995186097919941, + 0.008971380069851875, + 0.006393617019057274, + 0.010057180188596249, + 0.0082798320800066, + 0.06991613656282425, + 0.10768630355596542, + 0.014768626540899277, + 0.01779542863368988, + 0.016425855457782745, + 0.012778614647686481, + 0.0354502908885479, + 0.020018497481942177, + 0.016399262472987175, + 0.02074446715414524, + 0.010232661850750446, + 0.04095211252570152, + 0.011400185525417328, + 0.011533957906067371, + 0.0428788959980011, + 0.017775867134332657, + 0.012502780184149742, + 0.012758834287524223, + 0.005338074639439583, + 0.008922215551137924, + 0.008211659267544746, + 0.009052730165421963, + 0.01564534567296505, + 0.009572459384799004, + 0.006366643588989973, + 0.006870617158710957, + 0.00599024910479784, + 0.00815932359546423, + 0.006171791814267635, + 0.005548163782805204, + 0.0063234115950763226, + 0.006351511459797621, + 0.0069069573655724525, + 0.008110872469842434, + 0.006972622592002153, + 0.006056597921997309, + 0.006319690030068159, + 0.004241629503667355, + 0.008353719487786293, + 0.005343292374163866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -1.4063385725021362 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In a perfectly locked-off, medium close-up shot, a vibrant mantis shrimp hovers quietly in crystal-clear tropical waters. Its segmented carapace, clad in iridescent emerald, sapphire blue, and vibrant orange, shimmers like porcelain under the bright sunlight filtering through the water's surface, creating swaying caustics. Its body is slightly curled, its prominent, independently movable compound eyes warily observing the horizon. Its hammer-like raptors (predatory appendages) are tucked tightly beneath its body, like a fully wound spring, gathering terrifying power ready to explode. Suddenly, the entire scene explodes with Bones Studio's signature high-energy, fluid animation. With a speed barely perceptible to the naked eye, the raptor's limbs, rendered in a blur of colorful afterimages at the extreme frame rate, rip through the tranquil waters. Then, at the front of its impact path, a spherical shockwave, instantly formed from countless tiny, seething cavitation bubbles, violently spread outward—a silent \"howl\"—creating a visual feast in absolute silence. This piercing white ring of energy contrasted sharply with the vibrant neon colors of the mantis shrimp itself. Throughout this process, the completely still camera remained motionless, its extreme stability creating a thrilling dramatic tension with the creature's instantaneous, all-shattering burst of violent power, vividly conveying this momentary physical wonder through animated form.", + "seed": 25754, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_6/index_6544_seed_25754_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3461244505131617, + "vlm_logprob": -0.3461244505131617, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.3461244505131617, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011921468190848827 + ], + "opposite_logprob_tokens": [ + -5.136921405792236 + ], + "logprob_reward": -0.011921468190848827, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.011921468190848827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12861065566539764, + 0.09695055335760117, + 0.10812633484601974, + 0.07052850723266602, + 0.07399417459964752, + 0.0668395385146141, + 0.24545679986476898, + 0.08058173209428787, + 0.09901171177625656, + 0.029899999499320984 + ], + "credit_map": [ + 0.02885187417268753, + 0.015485865995287895, + 0.007872470654547215, + 0.011842884123325348, + 0.017816763371229172, + 0.01584857888519764, + 0.014730840921401978, + 0.016161376610398293, + 0.027525771409273148, + 0.009717706590890884, + 0.007905693724751472, + 0.007545403204858303, + 0.008541969582438469, + 0.013891615904867649, + 0.007743694819509983, + 0.014078698121011257, + 0.028573334217071533, + 0.007929971441626549, + 0.01208898052573204, + 0.011366857215762138, + 0.008266127668321133, + 0.015440610237419605, + 0.009662780910730362, + 0.014797660522162914, + 0.006756413262337446, + 0.00608966825529933, + 0.007983592338860035, + 0.009327837266027927, + 0.007607671432197094, + 0.016878247261047363, + 0.007319367956370115, + 0.008565707132220268, + 0.007866214029490948, + 0.010877437889575958, + 0.006441282108426094, + 0.00862770527601242, + 0.011797045357525349, + 0.009799041785299778, + 0.01039059553295374, + 0.008194849826395512, + 0.006738381460309029, + 0.006022173445671797, + 0.009800918400287628, + 0.007337779272347689, + 0.011151641607284546, + 0.01333834882825613, + 0.004000674933195114, + 0.0084496159106493, + 0.016821805387735367, + 0.08275255560874939, + 0.0071420930325984955, + 0.048842113465070724, + 0.02418558858335018, + 0.02055840939283371, + 0.008228675462305546, + 0.03692556545138359, + 0.009919985197484493, + 0.007979189045727253, + 0.007868834771215916, + 0.012289646081626415, + 0.014936799183487892, + 0.012784678488969803, + 0.006068732589483261, + 0.008733870461583138, + 0.01627749390900135, + 0.008444128558039665, + 0.03073522262275219, + 0.009184835478663445, + 0.005288336426019669, + 0.017902597784996033, + 0.005940734408795834, + 0.005238363053649664, + 0.0037211233284324408, + 0.0036350463051348925, + 0.00425391411408782, + 0.003914640750735998, + 0.0028182785026729107, + 0.003737442661076784, + 0.0032432074658572674, + 0.004576348699629307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.011921468190848827 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003637009998783469 + ], + "opposite_logprob_tokens": [ + -5.628636837005615 + ], + "logprob_reward": -0.003637009998783469, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003637009998783469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10098853707313538, + 0.05312666669487953, + 0.3111802935600281, + 0.06292551010847092, + 0.1402072310447693, + 0.04171918332576752, + 0.10960952937602997, + 0.04044061154127121, + 0.04959677532315254, + 0.09020575135946274 + ], + "credit_map": [ + 0.015322274528443813, + 0.011453662067651749, + 0.009038855321705341, + 0.01157759316265583, + 0.012453180737793446, + 0.013552390970289707, + 0.015678079798817635, + 0.011912494897842407, + 0.008171025663614273, + 0.005307902581989765, + 0.0041899243369698524, + 0.006362139247357845, + 0.006642065942287445, + 0.008565888740122318, + 0.006381693296134472, + 0.0075060222297906876, + 0.06507444381713867, + 0.03269898518919945, + 0.030542396008968353, + 0.03868373483419418, + 0.028542954474687576, + 0.04768214002251625, + 0.029529595747590065, + 0.038426030427217484, + 0.006389021873474121, + 0.006290128454566002, + 0.006532748695462942, + 0.0081641199067235, + 0.007932428270578384, + 0.013677062466740608, + 0.0061570219695568085, + 0.007782971486449242, + 0.011402665637433529, + 0.015520688146352768, + 0.013545584864914417, + 0.017140639945864677, + 0.01685888133943081, + 0.03674349933862686, + 0.012187839485704899, + 0.016807427629828453, + 0.008175631985068321, + 0.004248531069606543, + 0.006126753985881805, + 0.005342795979231596, + 0.0037001795135438442, + 0.00640094792470336, + 0.003316516289487481, + 0.004407822620123625, + 0.008626540191471577, + 0.02161088027060032, + 0.00888531282544136, + 0.019856590777635574, + 0.010682663880288601, + 0.017779670655727386, + 0.009031208232045174, + 0.013136656023561954, + 0.005891722626984119, + 0.004948650486767292, + 0.004578035324811935, + 0.006670873146504164, + 0.003936614375561476, + 0.006139886565506458, + 0.004509820602834225, + 0.0037650056183338165, + 0.010188700631260872, + 0.004574351944029331, + 0.011449165642261505, + 0.005959377158433199, + 0.0034318130929023027, + 0.005907592363655567, + 0.003642698982730508, + 0.0044430699199438095, + 0.00851423665881157, + 0.012901677750051022, + 0.010027593933045864, + 0.011840030550956726, + 0.008033573627471924, + 0.015337754972279072, + 0.009385930374264717, + 0.0141649404540658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.003637009998783469 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296907812356949 + ], + "opposite_logprob_tokens": [ + -2.237969160079956 + ], + "logprob_reward": -0.11296907812356949, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296907812356949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11932308971881866, + 0.11600954830646515, + 0.13318657875061035, + 0.11806514114141464, + 0.15330877900123596, + 0.07931115478277206, + 0.08191464841365814, + 0.05415017157793045, + 0.11406395584344864, + 0.03066694363951683 + ], + "credit_map": [ + 0.010913444682955742, + 0.0133352130651474, + 0.011840994469821453, + 0.012836270034313202, + 0.018326057121157646, + 0.014058146625757217, + 0.021663999184966087, + 0.016348961740732193, + 0.01906437985599041, + 0.010765745304524899, + 0.011179742403328419, + 0.01558830589056015, + 0.008735863491892815, + 0.015359560027718544, + 0.02093636430799961, + 0.014379582367837429, + 0.01449720747768879, + 0.01495535857975483, + 0.017303816974163055, + 0.024117540568113327, + 0.013483098708093166, + 0.024415750056505203, + 0.01032780110836029, + 0.014086020179092884, + 0.011173942126333714, + 0.013058595359325409, + 0.012986221350729465, + 0.019301043823361397, + 0.013618449680507183, + 0.020057277753949165, + 0.014918792061507702, + 0.012950819917023182, + 0.015359319746494293, + 0.017243506386876106, + 0.016462191939353943, + 0.01555714476853609, + 0.019268987700343132, + 0.03548155725002289, + 0.01486903615295887, + 0.019067030400037766, + 0.007459149695932865, + 0.00791230984032154, + 0.016408978030085564, + 0.006660839077085257, + 0.013318979181349277, + 0.01293259859085083, + 0.00639466242864728, + 0.008223637007176876, + 0.005649724043905735, + 0.0200826283544302, + 0.008257398381829262, + 0.013853086158633232, + 0.008892815560102463, + 0.010204346850514412, + 0.004973715171217918, + 0.010000931099057198, + 0.006202280521392822, + 0.006035433616489172, + 0.004991925787180662, + 0.00816994346678257, + 0.00986200850456953, + 0.007446247152984142, + 0.005894638132303953, + 0.005547693464905024, + 0.022017253562808037, + 0.009012540802359581, + 0.03579005226492882, + 0.014809996820986271, + 0.0060525150038301945, + 0.011037943884730339, + 0.005898823495954275, + 0.009444832801818848, + 0.003656727960333228, + 0.0035966788418591022, + 0.006047791801393032, + 0.004048580769449472, + 0.0022652100306004286, + 0.0032056556083261967, + 0.00410807179287076, + 0.0037382293958216906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.11296907812356949 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015136463334783912 + ], + "opposite_logprob_tokens": [ + -6.501513481140137 + ], + "logprob_reward": -0.0015136463334783912, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015136463334783912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1228332668542862, + 0.08375982940196991, + 0.09463600069284439, + 0.06531520932912827, + 0.16549696028232574, + 0.1464308798313141, + 0.07756287604570389, + 0.03874683007597923, + 0.05929660052061081, + 0.145921528339386 + ], + "credit_map": [ + 0.02751019410789013, + 0.013195891864597797, + 0.009555438533425331, + 0.01357749942690134, + 0.013468488119542599, + 0.01860332302749157, + 0.010642039589583874, + 0.016280392184853554, + 0.020126299932599068, + 0.0075264242477715015, + 0.007457911968231201, + 0.00694366917014122, + 0.00819481536746025, + 0.01392210554331541, + 0.007031306624412537, + 0.012557294219732285, + 0.018914764747023582, + 0.00911771785467863, + 0.010659325867891312, + 0.0108109712600708, + 0.008949582464993, + 0.017552342265844345, + 0.007971721701323986, + 0.010659581050276756, + 0.005462684202939272, + 0.0054156784899532795, + 0.008386029861867428, + 0.0119358841329813, + 0.006141840014606714, + 0.011580643244087696, + 0.009159000590443611, + 0.007233443204313517, + 0.015814991667866707, + 0.01915047876536846, + 0.017895806580781937, + 0.0236202385276556, + 0.018661698326468468, + 0.034058019518852234, + 0.01383141428232193, + 0.022464310750365257, + 0.020622096955776215, + 0.02186664193868637, + 0.01548671443015337, + 0.01737198606133461, + 0.0151055371388793, + 0.0227770134806633, + 0.016366668045520782, + 0.016834227368235588, + 0.006532388739287853, + 0.023521779105067253, + 0.004972174763679504, + 0.01421424001455307, + 0.008008843287825584, + 0.007892829366028309, + 0.0035818140022456646, + 0.008838814683258533, + 0.00575206009671092, + 0.004900586791336536, + 0.00341052352450788, + 0.0056819068267941475, + 0.0045234933495521545, + 0.006596003193408251, + 0.0032031398732215166, + 0.004679116886109114, + 0.007846440188586712, + 0.005835429299622774, + 0.016879772767424583, + 0.00645499536767602, + 0.0037782478611916304, + 0.012346677482128143, + 0.0031063384376466274, + 0.003048701211810112, + 0.015680866315960884, + 0.017420239746570587, + 0.019295817241072655, + 0.02036486193537712, + 0.011390184052288532, + 0.01935902237892151, + 0.020311901345849037, + 0.022098632529377937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.0015136463334783912 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6005810499191284 + ], + "opposite_logprob_tokens": [ + -0.22558103501796722 + ], + "logprob_reward": -1.6005810499191284, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005810499191284, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12478853762149811, + 0.042466651648283005, + 0.07501919567584991, + 0.0876193568110466, + 0.14781126379966736, + 0.0869276225566864, + 0.09266898036003113, + 0.12695366144180298, + 0.1329546719789505, + 0.08278999477624893 + ], + "credit_map": [ + 0.02290462516248226, + 0.014109326526522636, + 0.00805431418120861, + 0.013597465120255947, + 0.01995280385017395, + 0.01631808467209339, + 0.011244656518101692, + 0.018607264384627342, + 0.006367486901581287, + 0.004518597386777401, + 0.004058691672980785, + 0.006565900519490242, + 0.004359929356724024, + 0.0062055145390331745, + 0.005811197217553854, + 0.004579332657158375, + 0.00996390637010336, + 0.010598717257380486, + 0.006978126242756844, + 0.008155275136232376, + 0.00910500343888998, + 0.013842064887285233, + 0.008612299337983131, + 0.00776380579918623, + 0.007765599060803652, + 0.007827132008969784, + 0.00944285187870264, + 0.013460107147693634, + 0.009009464643895626, + 0.017041606828570366, + 0.01208951324224472, + 0.010983086191117764, + 0.016086090356111526, + 0.014640254899859428, + 0.015368703752756119, + 0.015029489062726498, + 0.019152730703353882, + 0.038172006607055664, + 0.014122634194791317, + 0.015239370055496693, + 0.013243076391518116, + 0.010752293281257153, + 0.008050600066781044, + 0.009025009348988533, + 0.013935103081166744, + 0.01538830529898405, + 0.005862434394657612, + 0.010670807212591171, + 0.006194534711539745, + 0.026489725336432457, + 0.006752867717295885, + 0.016945065930485725, + 0.009543388150632381, + 0.010126635432243347, + 0.005655001848936081, + 0.010961763560771942, + 0.010689329355955124, + 0.012428250163793564, + 0.009895057417452335, + 0.0170108824968338, + 0.02856196090579033, + 0.021874137222766876, + 0.013273375108838081, + 0.013220684602856636, + 0.029245436191558838, + 0.008693904615938663, + 0.05117001757025719, + 0.007582095451653004, + 0.005586104467511177, + 0.020694933831691742, + 0.004031933844089508, + 0.005950257182121277, + 0.007272507529705763, + 0.013733620755374432, + 0.007650885730981827, + 0.010903648100793362, + 0.007992261089384556, + 0.010440248064696789, + 0.010625922121107578, + 0.014170902781188488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -1.6005810499191284 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In a perfectly locked-off, medium close-up shot, a vibrant mantis shrimp hovers quietly in crystal-clear tropical waters. Its segmented carapace, clad in iridescent emerald, sapphire blue, and vibrant orange, shimmers like porcelain under the bright sunlight filtering through the water's surface, creating swaying caustics. Its body is slightly curled, its prominent, independently movable compound eyes warily observing the horizon. Its hammer-like raptors (predatory appendages) are tucked tightly beneath its body, like a fully wound spring, gathering terrifying power ready to explode. Suddenly, the entire scene explodes with Bones Studio's signature high-energy, fluid animation. With a speed barely perceptible to the naked eye, the raptor's limbs, rendered in a blur of colorful afterimages at the extreme frame rate, rip through the tranquil waters. Then, at the front of its impact path, a spherical shockwave, instantly formed from countless tiny, seething cavitation bubbles, violently spread outward—a silent \"howl\"—creating a visual feast in absolute silence. This piercing white ring of energy contrasted sharply with the vibrant neon colors of the mantis shrimp itself. Throughout this process, the completely still camera remained motionless, its extreme stability creating a thrilling dramatic tension with the creature's instantaneous, all-shattering burst of violent power, vividly conveying this momentary physical wonder through animated form.", + "seed": 25755, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_6/index_6544_seed_25755_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2876455429010093, + "vlm_logprob": -0.2876455429010093, + "vlm_yes_no_margin": 3.575000011920929, + "vlm_token_credit": -0.2876455429010093, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019724301993846893 + ], + "opposite_logprob_tokens": [ + -4.519724369049072 + ], + "logprob_reward": -0.019724301993846893, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.019724301993846893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11780493706464767, + 0.04510519281029701, + 0.0704827532172203, + 0.04347679764032364, + 0.026542596518993378, + 0.20638060569763184, + 0.03872641921043396, + 0.1852942705154419, + 0.08391492813825607, + 0.18227146565914154 + ], + "credit_map": [ + 0.024442456662654877, + 0.013905888423323631, + 0.007360528223216534, + 0.007373812608420849, + 0.008245034143328667, + 0.02890188619494438, + 0.010869261808693409, + 0.0167060736566782, + 0.004341823048889637, + 0.0048050046898424625, + 0.004311417695134878, + 0.005039258394390345, + 0.0038487573619931936, + 0.009823205880820751, + 0.006293173413723707, + 0.006642554886639118, + 0.007079369388520718, + 0.0054505858570337296, + 0.0041605764999985695, + 0.003329241182655096, + 0.006304573267698288, + 0.027261164039373398, + 0.0033551736269146204, + 0.013542067259550095, + 0.008774654008448124, + 0.007081341464072466, + 0.004661778919398785, + 0.003858734853565693, + 0.0045199207961559296, + 0.005266768392175436, + 0.004769737832248211, + 0.0045438590459525585, + 0.003355981782078743, + 0.0033581897150725126, + 0.0019040992483496666, + 0.0017439130460843444, + 0.0023978000972419977, + 0.006184434983879328, + 0.0024577765725553036, + 0.005140401888638735, + 0.02136731520295143, + 0.07297565042972565, + 0.006732981652021408, + 0.006913700606673956, + 0.011186077259480953, + 0.01839308999478817, + 0.009699481539428234, + 0.0591123104095459, + 0.004022636450827122, + 0.005357439164072275, + 0.006379322614520788, + 0.006597758736461401, + 0.0038446264807134867, + 0.003723924746736884, + 0.005684725474566221, + 0.003115987405180931, + 0.016562262549996376, + 0.025949666276574135, + 0.014931204728782177, + 0.02343342825770378, + 0.028543878346681595, + 0.02463100664317608, + 0.0256844162940979, + 0.025558393448591232, + 0.009409620426595211, + 0.014728505164384842, + 0.008000134490430355, + 0.0075399079360067844, + 0.01199235487729311, + 0.007134137209504843, + 0.016708001494407654, + 0.008402264676988125, + 0.01220401655882597, + 0.0232850369066, + 0.0363956056535244, + 0.011122290976345539, + 0.061157651245594025, + 0.012700030580163002, + 0.012665149755775928, + 0.012741691432893276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.019724301993846893 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002504785545170307 + ], + "opposite_logprob_tokens": [ + -6.002504825592041 + ], + "logprob_reward": -0.002504785545170307, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002504785545170307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1737508624792099, + 0.1566234529018402, + 0.0729459896683693, + 0.05524303764104843, + 0.06461714208126068, + 0.04884164035320282, + 0.14328667521476746, + 0.05221150442957878, + 0.1574317216873169, + 0.07504796981811523 + ], + "credit_map": [ + 0.03190261870622635, + 0.016498591750860214, + 0.021015606820583344, + 0.009552604518830776, + 0.01564546301960945, + 0.03681853413581848, + 0.021217316389083862, + 0.021100113168358803, + 0.015584923326969147, + 0.0139670604839921, + 0.009139721281826496, + 0.007351228501647711, + 0.0117301931604743, + 0.054010678082704544, + 0.010709470137953758, + 0.03413017466664314, + 0.009209791198372841, + 0.01132764108479023, + 0.006043136585503817, + 0.007882161065936089, + 0.010925997979938984, + 0.010758280754089355, + 0.009147745557129383, + 0.00765123451128602, + 0.009551501832902431, + 0.006810017861425877, + 0.005419331602752209, + 0.005539906211197376, + 0.005943452939391136, + 0.007700526621192694, + 0.008540201000869274, + 0.005738104227930307, + 0.011059940792620182, + 0.00921180471777916, + 0.0067151435650885105, + 0.004315914120525122, + 0.006912339013069868, + 0.010835115797817707, + 0.005301120225340128, + 0.010265766642987728, + 0.004541544243693352, + 0.013575109653174877, + 0.0017620297148823738, + 0.0019759752321988344, + 0.0027932077646255493, + 0.008045880123972893, + 0.002714065834879875, + 0.013433828018605709, + 0.019832132384181023, + 0.01889040134847164, + 0.021293506026268005, + 0.022156301885843277, + 0.01569666713476181, + 0.014554746448993683, + 0.019205279648303986, + 0.01165764220058918, + 0.005023731850087643, + 0.007686849217861891, + 0.0056540402583777905, + 0.005302529316395521, + 0.008152296766638756, + 0.009195971302688122, + 0.005679210182279348, + 0.005516872741281986, + 0.014764521270990372, + 0.027575939893722534, + 0.017545251175761223, + 0.015104899182915688, + 0.018995819613337517, + 0.016938403248786926, + 0.029368383809924126, + 0.01713850349187851, + 0.006544983945786953, + 0.0107735525816679, + 0.010253203101456165, + 0.007799212820827961, + 0.013997361995279789, + 0.007192813325673342, + 0.011775441467761993, + 0.006711400579661131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.002504785545170307 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034223645925522 + ], + "opposite_logprob_tokens": [ + -2.350342273712158 + ], + "logprob_reward": -0.10034223645925522, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034223645925522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10097330808639526, + 0.08968257904052734, + 0.08951502293348312, + 0.03994497284293175, + 0.10274527221918106, + 0.237165629863739, + 0.04500316455960274, + 0.04304729402065277, + 0.06184766814112663, + 0.1900751143693924 + ], + "credit_map": [ + 0.013843214139342308, + 0.01259384024888277, + 0.010249610990285873, + 0.00891108252108097, + 0.013232414610683918, + 0.018467890098690987, + 0.012457181699573994, + 0.01121806725859642, + 0.009163889102637768, + 0.012161670252680779, + 0.005942816846072674, + 0.006343989633023739, + 0.01246277242898941, + 0.02004143036901951, + 0.009120511822402477, + 0.014445493929088116, + 0.00986392516642809, + 0.007472823839634657, + 0.004836591891944408, + 0.004176036454737186, + 0.0075404890812933445, + 0.03384198993444443, + 0.004272988066077232, + 0.017510168254375458, + 0.007222101092338562, + 0.0040583377704024315, + 0.004751488100737333, + 0.005314415320754051, + 0.004145606886595488, + 0.005044321529567242, + 0.005117875058203936, + 0.004290821962058544, + 0.01500372402369976, + 0.012917875312268734, + 0.0063733612187206745, + 0.006405036896467209, + 0.00939651858061552, + 0.03044823370873928, + 0.0076156421564519405, + 0.014584871008992195, + 0.02434959076344967, + 0.08487562835216522, + 0.006054494995623827, + 0.0076672593131661415, + 0.012849102728068829, + 0.021389862522482872, + 0.010944783687591553, + 0.06903489679098129, + 0.004636629018932581, + 0.00895462092012167, + 0.00626318110153079, + 0.0038980538956820965, + 0.006124209146946669, + 0.0057772304862737656, + 0.005695856176316738, + 0.003653382882475853, + 0.0030679628252983093, + 0.00781578291207552, + 0.005224371328949928, + 0.0035338448360562325, + 0.009090246632695198, + 0.004940018057823181, + 0.00510915694758296, + 0.004265905357897282, + 0.007979415357112885, + 0.010735736228525639, + 0.006193847861140966, + 0.0064110103994607925, + 0.010309859178960323, + 0.00628179544582963, + 0.006537452805787325, + 0.007398545742034912, + 0.013041837140917778, + 0.0247871782630682, + 0.04217373579740524, + 0.01044053491204977, + 0.06613343954086304, + 0.014346019364893436, + 0.011794834397733212, + 0.007357506081461906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.10034223645925522 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002197947818785906 + ], + "opposite_logprob_tokens": [ + -6.127197742462158 + ], + "logprob_reward": -0.002197947818785906, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002197947818785906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043837569653987885, + 0.036029037088155746, + 0.05069122463464737, + 0.017964988946914673, + 0.019473865628242493, + 0.07294531911611557, + 0.0467536598443985, + 0.029641466215252876, + 0.11145240068435669, + 0.571210503578186 + ], + "credit_map": [ + 0.00874300580471754, + 0.0038377284072339535, + 0.0032095585484057665, + 0.002930563874542713, + 0.004607377573847771, + 0.010706677101552486, + 0.0034264451824128628, + 0.006376206409186125, + 0.0037183845415711403, + 0.002815880114212632, + 0.0021136163268238306, + 0.0021019147243350744, + 0.002435636008158326, + 0.012546258047223091, + 0.00221281498670578, + 0.008084527216851711, + 0.00553157040849328, + 0.0053387912921607494, + 0.004231397993862629, + 0.003278194461017847, + 0.0037814609240740538, + 0.016588278114795685, + 0.002769988728687167, + 0.00917153898626566, + 0.003125281073153019, + 0.0025614676997065544, + 0.0021068912465125322, + 0.0022681071422994137, + 0.0013301930157467723, + 0.002054196549579501, + 0.002241251291707158, + 0.002277596853673458, + 0.0034836470149457455, + 0.0030893448274582624, + 0.001683420385234058, + 0.0018956304993480444, + 0.0023647365160286427, + 0.0031319845002144575, + 0.0016688320320099592, + 0.0021562688052654266, + 0.006889361422508955, + 0.025989139452576637, + 0.0015877910191193223, + 0.0024519970174878836, + 0.003924421034753323, + 0.006432843394577503, + 0.0036231984850019217, + 0.022046558558940887, + 0.0064588249661028385, + 0.007204058114439249, + 0.0040577491745352745, + 0.005551939830183983, + 0.006192430388182402, + 0.00693811709061265, + 0.005899774841964245, + 0.004450760781764984, + 0.0025696412194520235, + 0.004573277197778225, + 0.003255404531955719, + 0.003495018230751157, + 0.00422076927497983, + 0.004646096378564835, + 0.0038661703001707792, + 0.0030150844249874353, + 0.019898878410458565, + 0.016896666958928108, + 0.011825557798147202, + 0.010789004154503345, + 0.014665004797279835, + 0.010063216090202332, + 0.014703473076224327, + 0.012610580772161484, + 0.03545119985938072, + 0.07588951289653778, + 0.12253236025571823, + 0.034366291016340256, + 0.19744013249874115, + 0.03990248963236809, + 0.04170738160610199, + 0.023921096697449684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.002197947818785906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a mantis shrimp in crystal-clear tropical waters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrimp display a segmented carapace with iridescent emerald, sapphire blue, and vibrant orange colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the shrimp's hammer-like raptor appendages initially tucked tightly beneath its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shrimp rapidly strike the water, creating a spherical shockwave of tiny cavitation bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely locked-off and motionless throughout the entire sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134584426879883 + ], + "opposite_logprob_tokens": [ + -0.31345850229263306 + ], + "logprob_reward": -1.3134584426879883, + "yes_no_margin_reward": -0.9999999403953552, + "token_credit_reward": -1.3134584426879883, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05476231127977371, + 0.030756384134292603, + 0.019193200394511223, + 0.02244037203490734, + 0.01910223811864853, + 0.06542188674211502, + 0.06861412525177002, + 0.023190705105662346, + 0.053416717797517776, + 0.6431020498275757 + ], + "credit_map": [ + 0.010332152247428894, + 0.004379135090857744, + 0.004127766937017441, + 0.003534819232299924, + 0.004351365379989147, + 0.01558644324541092, + 0.003867106279358268, + 0.008583525195717812, + 0.002882891334593296, + 0.003014626679942012, + 0.0021198843605816364, + 0.0019801135640591383, + 0.0031836337875574827, + 0.008913285098969936, + 0.002425141166895628, + 0.00623680604621768, + 0.002584573347121477, + 0.0029384419322013855, + 0.0013914895243942738, + 0.0018429378978908062, + 0.0014301640912890434, + 0.004877373576164246, + 0.0013680884148925543, + 0.0027601311448961496, + 0.005368873011320829, + 0.0034332112409174442, + 0.0018956614658236504, + 0.0016946365358307958, + 0.002151321619749069, + 0.002641690196469426, + 0.002267338801175356, + 0.002987638581544161, + 0.003210226073861122, + 0.0031079468317329884, + 0.0022176685743033886, + 0.0014139509294182062, + 0.0019564114045351744, + 0.0025897084269672632, + 0.0020188898779451847, + 0.0025874341372400522, + 0.007126744836568832, + 0.021847685799002647, + 0.0020406735129654408, + 0.0027145191561430693, + 0.003786433720961213, + 0.006178538780659437, + 0.0036458326503634453, + 0.018081452697515488, + 0.01051962561905384, + 0.008226973004639149, + 0.009447709657251835, + 0.007290471810847521, + 0.009151107631623745, + 0.008972507901489735, + 0.008600519970059395, + 0.006405213847756386, + 0.002139660296961665, + 0.0029029676225036383, + 0.00278019392862916, + 0.0027866049204021692, + 0.00259271333925426, + 0.0046953847631812096, + 0.0023144728038460016, + 0.0029787083622068167, + 0.00812947005033493, + 0.00845094583928585, + 0.005408328026533127, + 0.004786945879459381, + 0.006784655619412661, + 0.005258835386484861, + 0.007747076451778412, + 0.006850462406873703, + 0.03917228803038597, + 0.08593437820672989, + 0.1405222862958908, + 0.03703110292553902, + 0.22656895220279694, + 0.0443829707801342, + 0.043041523545980453, + 0.026448562741279602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -1.3134584426879883 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_7/index_tensor([6248, 6248])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_7/index_tensor([6248, 6248])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7ac769da585b4791521ade8aa531ba6164dd06b1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_7/index_tensor([6248, 6248])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a classic 16-bit pixel art style, a shaky overhead shot closely follows a character equipped with a massive rocket arm as they soar vertically at breakneck speed. Clad in tight armor composed of blue and gray pixels, the character's massive metal rocket arm blasts downward from the elbow with a massive cone of fiery white, yellow, and orange pixels. This pulsing flame becomes the sole source of light, its violent flickering light instantly illuminating the surrounding narrow passage walls, composed of dark brown and gray pixels, casting a constantly dancing, high-contrast play of light and shadow on the rough pixel surfaces. As the character soars, the unstable camera feed trembles violently from the immense propulsive force, and blurred pixel particles zip past beneath them, perfectly embodying the tense, powerful, and dynamic atmosphere of vertical advancement typical of retro games.", + "seed": 27864, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_7/index_6248_seed_27864_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08912386633455753, + "vlm_logprob": -0.08912386633455753, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.08912386633455753, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32000112533569336 + ], + "opposite_logprob_tokens": [ + -1.3200011253356934 + ], + "logprob_reward": -0.32000112533569336, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32000112533569336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11482813209295273, + 0.08812735229730606, + 0.11432827264070511, + 0.0840500146150589, + 0.07048949599266052, + 0.03584814816713333, + 0.08102376013994217, + 0.14651831984519958, + 0.03913816437125206, + 0.22564835846424103 + ], + "credit_map": [ + 0.01478145644068718, + 0.012261767871677876, + 0.011170282028615475, + 0.012952418066561222, + 0.013006028719246387, + 0.011943173594772816, + 0.010029632598161697, + 0.0286833755671978, + 0.015063120052218437, + 0.012501495890319347, + 0.015538978390395641, + 0.006353297736495733, + 0.010744134895503521, + 0.011587952263653278, + 0.004690582398325205, + 0.011647791601717472, + 0.019275683909654617, + 0.037357959896326065, + 0.007845642976462841, + 0.005565807223320007, + 0.007773071061819792, + 0.024048248305916786, + 0.0037839266005903482, + 0.008677934296429157, + 0.013555418699979782, + 0.015825890004634857, + 0.012893365696072578, + 0.006320721935480833, + 0.006910581141710281, + 0.013397440314292908, + 0.003741459222510457, + 0.011405129916965961, + 0.008356777019798756, + 0.009545497596263885, + 0.008588063530623913, + 0.006326933391392231, + 0.009082434698939323, + 0.012580793350934982, + 0.006430245004594326, + 0.009578748606145382, + 0.003707109484821558, + 0.004460716620087624, + 0.004849027842283249, + 0.004258396103978157, + 0.00512170372530818, + 0.007014477159827948, + 0.002077954588457942, + 0.004358764737844467, + 0.006892774254083633, + 0.006948626600205898, + 0.008342336863279343, + 0.002302093431353569, + 0.0039235008880496025, + 0.028728069737553596, + 0.001945273019373417, + 0.021941084414720535, + 0.017042092978954315, + 0.01613970287144184, + 0.031172797083854675, + 0.00847515370696783, + 0.01201311033219099, + 0.026730753481388092, + 0.0082413200289011, + 0.026703398674726486, + 0.00832323543727398, + 0.006032589357346296, + 0.005997073836624622, + 0.0021245914977043867, + 0.003891517175361514, + 0.00557123264297843, + 0.0020431375596672297, + 0.00515478802844882, + 0.015791088342666626, + 0.009085035882890224, + 0.007192105520516634, + 0.005557393189519644, + 0.009458030574023724, + 0.09501547366380692, + 0.004169566556811333, + 0.07937966287136078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.32000112533569336 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297333985567093 + ], + "opposite_logprob_tokens": [ + -2.23797345161438 + ], + "logprob_reward": -0.11297333985567093, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297333985567093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6145347952842712, + 0.06378781795501709, + 0.09999797493219376, + 0.02618875354528427, + 0.04369499906897545, + 0.02131381630897522, + 0.020787963643670082, + 0.06426052749156952, + 0.018116125836968422, + 0.027317244559526443 + ], + "credit_map": [ + 0.12851746380329132, + 0.03719240427017212, + 0.04692952707409859, + 0.029671765863895416, + 0.034312304109334946, + 0.06284552812576294, + 0.03157304972410202, + 0.2434927523136139, + 0.009622594341635704, + 0.00942368432879448, + 0.011033480986952782, + 0.004324787762016058, + 0.008349324576556683, + 0.008260802365839481, + 0.003426429582759738, + 0.009346711449325085, + 0.018382182344794273, + 0.015977777540683746, + 0.01671130396425724, + 0.008820057846605778, + 0.008893890306353569, + 0.014241493307054043, + 0.006502571981400251, + 0.010468700900673866, + 0.003430616110563278, + 0.003969787620007992, + 0.003645643824711442, + 0.002345743589103222, + 0.003403550246730447, + 0.003789076814427972, + 0.0015573397977277637, + 0.004046996124088764, + 0.004661239217966795, + 0.006231043953448534, + 0.004707068204879761, + 0.003844562452286482, + 0.005844061728566885, + 0.00683580944314599, + 0.0047745355404913425, + 0.006796682253479958, + 0.002180911134928465, + 0.00242947437800467, + 0.0032459981739521027, + 0.0018278352217748761, + 0.003338991664350033, + 0.00448295334354043, + 0.0012809245381504297, + 0.002526728203520179, + 0.0040160492062568665, + 0.0035825108643621206, + 0.003242833074182272, + 0.0015209150733426213, + 0.0020731922704726458, + 0.002646347275003791, + 0.001283849822357297, + 0.0024222650099545717, + 0.008223159238696098, + 0.008157270960509777, + 0.015806790441274643, + 0.003376120701432228, + 0.00402170279994607, + 0.007845068350434303, + 0.002725476399064064, + 0.014104936271905899, + 0.003302425378933549, + 0.002356388606131077, + 0.003147940617054701, + 0.0016308934427797794, + 0.0018639429472386837, + 0.0024569244123995304, + 0.0012545083882287145, + 0.002103102160617709, + 0.004174106754362583, + 0.0020582983270287514, + 0.0020740246400237083, + 0.00132096151355654, + 0.0019375592237338424, + 0.009405050426721573, + 0.0009374679066240788, + 0.005409775301814079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.11297333985567093 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022031813859939575 + ], + "opposite_logprob_tokens": [ + -6.127202987670898 + ], + "logprob_reward": -0.0022031813859939575, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022031813859939575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30081605911254883, + 0.07754167914390564, + 0.16915006935596466, + 0.04556763544678688, + 0.04883752390742302, + 0.03150679171085358, + 0.0383543036878109, + 0.15002253651618958, + 0.015102077275514603, + 0.1231013610959053 + ], + "credit_map": [ + 0.056129660457372665, + 0.019782505929470062, + 0.023681407794356346, + 0.017846496775746346, + 0.014798295684158802, + 0.032325685024261475, + 0.015461877919733524, + 0.12079013139009476, + 0.010257509537041187, + 0.010224651545286179, + 0.010010933503508568, + 0.00961050484329462, + 0.010317768901586533, + 0.00852177944034338, + 0.009084411896765232, + 0.009514118544757366, + 0.02138485759496689, + 0.06597395241260529, + 0.011311359703540802, + 0.007611572276800871, + 0.010300417430698872, + 0.039525941014289856, + 0.005857367999851704, + 0.007184583693742752, + 0.005688594654202461, + 0.005970852915197611, + 0.00682660099118948, + 0.005264222156256437, + 0.005858418066054583, + 0.005925516132265329, + 0.003712079953402281, + 0.006321347318589687, + 0.006857249420136213, + 0.0067202141508460045, + 0.008169228211045265, + 0.002604997716844082, + 0.005607328377664089, + 0.007181711029261351, + 0.002491934923455119, + 0.009204860776662827, + 0.0031538920011371374, + 0.0038314613047987223, + 0.004065311513841152, + 0.003929679282009602, + 0.004148958716541529, + 0.006613754667341709, + 0.0021328870207071304, + 0.00363084371201694, + 0.00426807627081871, + 0.004984443075954914, + 0.004705397412180901, + 0.002757619833573699, + 0.003756638616323471, + 0.00692345155403018, + 0.0020079913083463907, + 0.00895068608224392, + 0.017998969182372093, + 0.01619398593902588, + 0.03858114778995514, + 0.008271408267319202, + 0.00899134948849678, + 0.02063198760151863, + 0.006005721632391214, + 0.0333479568362236, + 0.002262961817905307, + 0.0025284846778959036, + 0.0017738341121003032, + 0.0015850058989599347, + 0.0017314278520643711, + 0.0021304243709892035, + 0.0013187967706471682, + 0.0017711413092911243, + 0.00933010783046484, + 0.0054206931963562965, + 0.0038712655659765005, + 0.003070537466555834, + 0.005562145262956619, + 0.052041999995708466, + 0.002354069845750928, + 0.04145053029060364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0022031813859939575 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028165215626358986 + ], + "opposite_logprob_tokens": [ + -5.877816677093506 + ], + "logprob_reward": -0.0028165215626358986, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028165215626358986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23779024183750153, + 0.09905966371297836, + 0.07566272467374802, + 0.06540505588054657, + 0.07400023192167282, + 0.022031033411622047, + 0.055073998868465424, + 0.25778862833976746, + 0.015551019459962845, + 0.09763747453689575 + ], + "credit_map": [ + 0.050529442727565765, + 0.011444192379713058, + 0.017622841522097588, + 0.010555007494986057, + 0.010283335112035275, + 0.025561314076185226, + 0.011723075993359089, + 0.10007103532552719, + 0.02194243296980858, + 0.014490406028926373, + 0.015645360574126244, + 0.005694591905921698, + 0.008226235397160053, + 0.013782200403511524, + 0.004633663222193718, + 0.014644774608314037, + 0.01426096260547638, + 0.022824378684163094, + 0.0063246083445847034, + 0.004983965773135424, + 0.005816715769469738, + 0.013945399783551693, + 0.0030037963297218084, + 0.004502892028540373, + 0.009357224218547344, + 0.01320516038686037, + 0.009977071546018124, + 0.0042487396858632565, + 0.0051054572686553, + 0.012206935323774815, + 0.0021608276292681694, + 0.009143641218543053, + 0.008553802967071533, + 0.011964353732764721, + 0.008492931723594666, + 0.005996923893690109, + 0.009796231985092163, + 0.013286363333463669, + 0.00769408792257309, + 0.008215534500777721, + 0.0043943701311945915, + 0.0028429534286260605, + 0.003524160711094737, + 0.001779603655450046, + 0.002411849098280072, + 0.003434400772675872, + 0.0011567013571038842, + 0.002486994257196784, + 0.006217747926712036, + 0.004609180148690939, + 0.007077184971421957, + 0.001939951442182064, + 0.0030390063766390085, + 0.01859322190284729, + 0.001537349191494286, + 0.012060358189046383, + 0.03371483460068703, + 0.02757692150771618, + 0.06807710975408554, + 0.017873572185635567, + 0.011368460953235626, + 0.035057105123996735, + 0.007914293557405472, + 0.056206297129392624, + 0.003005372826009989, + 0.002150068525224924, + 0.002077706390991807, + 0.001486727618612349, + 0.0014255542773753405, + 0.0020341805648058653, + 0.0010905181989073753, + 0.002280892338603735, + 0.010504708625376225, + 0.00535122025758028, + 0.004853685852140188, + 0.0030926240142434835, + 0.005462962202727795, + 0.03930894657969475, + 0.001964992843568325, + 0.027098331600427628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.0028165215626358986 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625163532793522 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.007625163532793522, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625163532793522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28867220878601074, + 0.03099445253610611, + 0.02641027420759201, + 0.07297370582818985, + 0.03641361743211746, + 0.01649388298392296, + 0.07622653990983963, + 0.05967330560088158, + 0.017100326716899872, + 0.3750416338443756 + ], + "credit_map": [ + 0.06510297954082489, + 0.014429569244384766, + 0.019312219694256783, + 0.010287579149007797, + 0.012361795641481876, + 0.033053550869226456, + 0.013745258562266827, + 0.12037930637598038, + 0.00629415875300765, + 0.004048926755785942, + 0.00477480236440897, + 0.002294120378792286, + 0.0027075784746557474, + 0.004074284806847572, + 0.0011595884570851922, + 0.005640995688736439, + 0.0031197736971080303, + 0.0077205076813697815, + 0.002091076457872987, + 0.0020586145110428333, + 0.002394709037616849, + 0.005014469381421804, + 0.002099418081343174, + 0.0019117090851068497, + 0.014563897624611855, + 0.016433613374829292, + 0.010245732963085175, + 0.005149083212018013, + 0.0070927999913692474, + 0.009416680783033371, + 0.0026341904886066914, + 0.007437715772539377, + 0.004838872700929642, + 0.004115980118513107, + 0.00537102110683918, + 0.002698478288948536, + 0.004179008770734072, + 0.0087473439052701, + 0.001966224517673254, + 0.004496688488870859, + 0.001759379985742271, + 0.0022567925043404102, + 0.002448972314596176, + 0.0013499085325747728, + 0.001983026275411248, + 0.00400410545989871, + 0.0008544297306798398, + 0.0018372698687016964, + 0.007813951000571251, + 0.006054745987057686, + 0.008223628625273705, + 0.001967191928997636, + 0.0047470321878790855, + 0.02402365207672119, + 0.001906643039546907, + 0.02148970402777195, + 0.008130955509841442, + 0.006283361930400133, + 0.014545547775924206, + 0.00408058799803257, + 0.005426560994237661, + 0.006048622541129589, + 0.003682695096358657, + 0.011474980972707272, + 0.0038916817866265774, + 0.002617544960230589, + 0.0030300000216811895, + 0.0011376288020983338, + 0.0010939576895907521, + 0.0023081880062818527, + 0.0008012878824956715, + 0.0022200383245944977, + 0.02312598191201687, + 0.014311889186501503, + 0.010441118851304054, + 0.008811150677502155, + 0.015202897600829601, + 0.1605343520641327, + 0.006964621599763632, + 0.1356496512889862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.007625163532793522 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a classic 16-bit pixel art style, a shaky overhead shot closely follows a character equipped with a massive rocket arm as they soar vertically at breakneck speed. Clad in tight armor composed of blue and gray pixels, the character's massive metal rocket arm blasts downward from the elbow with a massive cone of fiery white, yellow, and orange pixels. This pulsing flame becomes the sole source of light, its violent flickering light instantly illuminating the surrounding narrow passage walls, composed of dark brown and gray pixels, casting a constantly dancing, high-contrast play of light and shadow on the rough pixel surfaces. As the character soars, the unstable camera feed trembles violently from the immense propulsive force, and blurred pixel particles zip past beneath them, perfectly embodying the tense, powerful, and dynamic atmosphere of vertical advancement typical of retro games.", + "seed": 27865, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_7/index_6248_seed_27865_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024452028051018714, + "vlm_logprob": -0.024452028051018714, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.024452028051018714, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0624944306910038 + ], + "opposite_logprob_tokens": [ + -2.9374945163726807 + ], + "logprob_reward": -0.0624944306910038, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0624944306910038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07708236575126648, + 0.051421962678432465, + 0.04683798924088478, + 0.07449230551719666, + 0.583303689956665, + 0.019128037616610527, + 0.03540225327014923, + 0.0424935519695282, + 0.052604544907808304, + 0.017233289778232574 + ], + "credit_map": [ + 0.009285543113946915, + 0.01149270310997963, + 0.009822841733694077, + 0.009549655020236969, + 0.011333919130265713, + 0.010366865433752537, + 0.007182867731899023, + 0.008047974668443203, + 0.006991862785071135, + 0.006204165518283844, + 0.01092600915580988, + 0.0035096541978418827, + 0.006164201069623232, + 0.006592147052288055, + 0.004328783135861158, + 0.006705138366669416, + 0.007659422233700752, + 0.0034535378217697144, + 0.007997645065188408, + 0.00579441525042057, + 0.002902607200667262, + 0.008125701919198036, + 0.004519911482930183, + 0.006384749431163073, + 0.011649561114609241, + 0.007912090048193932, + 0.011525207199156284, + 0.009016910567879677, + 0.005056964233517647, + 0.0074430531822144985, + 0.009962392970919609, + 0.011926131322979927, + 0.057570651173591614, + 0.08489032834768295, + 0.05563952773809433, + 0.10339848697185516, + 0.07327974587678909, + 0.08523005247116089, + 0.09008313715457916, + 0.03321176394820213, + 0.0021195830777287483, + 0.00264843855984509, + 0.0019228917080909014, + 0.0019393042894080281, + 0.0026505040004849434, + 0.0024838370736688375, + 0.0036023107822984457, + 0.001761166611686349, + 0.004831757862120867, + 0.002056090859696269, + 0.011019994504749775, + 0.003243159269914031, + 0.002424929291009903, + 0.003614691086113453, + 0.003967840690165758, + 0.004243792500346899, + 0.005280326586216688, + 0.0041954717598855495, + 0.006235216278582811, + 0.005824989639222622, + 0.0033845584839582443, + 0.0066438824869692326, + 0.006263348739594221, + 0.00466575613245368, + 0.004190908279269934, + 0.005628757178783417, + 0.008068695664405823, + 0.006982672959566116, + 0.006056360434740782, + 0.007588617969304323, + 0.00716131366789341, + 0.006927217822521925, + 0.0022829254157841206, + 0.0022108019329607487, + 0.002947374014183879, + 0.00113603041972965, + 0.0022318426053971052, + 0.0022258241660892963, + 0.0017086551524698734, + 0.0024898347910493612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0624944306910038 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089595966041088 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089595966041088, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089595966041088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07394241541624069, + 0.02034245803952217, + 0.02302917279303074, + 0.057208143174648285, + 0.5035949349403381, + 0.02940780110657215, + 0.024434769526124, + 0.04826631397008896, + 0.20370735228061676, + 0.016066646203398705 + ], + "credit_map": [ + 0.00906518753618002, + 0.01357520092278719, + 0.007206915412098169, + 0.010554446838796139, + 0.012182818725705147, + 0.007468680385500193, + 0.006871264893561602, + 0.007017898373305798, + 0.0030056715477257967, + 0.0021055624820291996, + 0.005244571715593338, + 0.0015263075474649668, + 0.0019137415802106261, + 0.0021532434038817883, + 0.0015185303054749966, + 0.002874828176572919, + 0.0045555694960057735, + 0.0017532890196889639, + 0.004166575614362955, + 0.00279145035892725, + 0.0015812820056453347, + 0.003251505084335804, + 0.002059811493381858, + 0.002869688207283616, + 0.005551778711378574, + 0.004610009491443634, + 0.010734844952821732, + 0.00946810469031334, + 0.004523403011262417, + 0.005255708005279303, + 0.011056147515773773, + 0.006008146796375513, + 0.05053849518299103, + 0.07291002571582794, + 0.04983801022171974, + 0.0860941931605339, + 0.06084263324737549, + 0.07709071040153503, + 0.07607535272836685, + 0.030205506831407547, + 0.003145029302686453, + 0.005163582973182201, + 0.0035071163438260555, + 0.002185041317716241, + 0.004701994359493256, + 0.003326760372146964, + 0.004231685772538185, + 0.003146591829136014, + 0.003577844239771366, + 0.0014258295996114612, + 0.0055249216966331005, + 0.0017350739799439907, + 0.0016252562636509538, + 0.004641920793801546, + 0.002287107054144144, + 0.0036168163642287254, + 0.004884846042841673, + 0.004444332793354988, + 0.006693607661873102, + 0.009095045737922192, + 0.004293530248105526, + 0.006434760056436062, + 0.007291974034160376, + 0.0051282173953950405, + 0.014538311399519444, + 0.020906412973999977, + 0.053198639303445816, + 0.028317047283053398, + 0.015894360840320587, + 0.032878972589969635, + 0.01775285229086876, + 0.020220763981342316, + 0.0018270249711349607, + 0.002487208228558302, + 0.0025504999794065952, + 0.0010062141809612513, + 0.002864293986931443, + 0.0019071127753704786, + 0.0015041774604469538, + 0.0019201146205887198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.011089595966041088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952369887381792 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952369887381792, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952369887381792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03500315919518471, + 0.021363086998462677, + 0.02156008780002594, + 0.035217031836509705, + 0.7600275874137878, + 0.01646575890481472, + 0.016177456825971603, + 0.02863624319434166, + 0.05364201217889786, + 0.011907598935067654 + ], + "credit_map": [ + 0.0041291979141533375, + 0.006214516703039408, + 0.004008126445114613, + 0.0051157656125724316, + 0.004888644441962242, + 0.0037128203548491, + 0.0032156999222934246, + 0.0037183882668614388, + 0.002886892529204488, + 0.0032643787562847137, + 0.003104462521150708, + 0.001756426878273487, + 0.003012158442288637, + 0.002540489425882697, + 0.0022735316306352615, + 0.002524747746065259, + 0.0030073204543441534, + 0.0012595017906278372, + 0.00466645834967494, + 0.0020342119969427586, + 0.0012523629702627659, + 0.00283754151314497, + 0.0028803900349885225, + 0.003622302319854498, + 0.0028736942913383245, + 0.002694298280403018, + 0.005368053913116455, + 0.004162669647485018, + 0.0027308552525937557, + 0.005229116417467594, + 0.004623666871339083, + 0.007534677628427744, + 0.07601066678762436, + 0.11693945527076721, + 0.07060008496046066, + 0.12979228794574738, + 0.09628748148679733, + 0.1110469326376915, + 0.116249218583107, + 0.04310144484043121, + 0.0018032436491921544, + 0.002757876645773649, + 0.002397531410679221, + 0.0015149788232520223, + 0.0028101846110075712, + 0.001261710305698216, + 0.0023906328715384007, + 0.0015296004712581635, + 0.002905159955844283, + 0.0009481961606070399, + 0.0031692390330135822, + 0.0012268814025446773, + 0.0011364484671503305, + 0.0024244249798357487, + 0.0018116958672180772, + 0.002555410610511899, + 0.003584712976589799, + 0.003350816899910569, + 0.0032051478046923876, + 0.00473740603774786, + 0.002966011641547084, + 0.003768108319491148, + 0.0041380031034350395, + 0.0028860350139439106, + 0.00498077180236578, + 0.003977792803198099, + 0.018084969371557236, + 0.003414720995351672, + 0.0028318529948592186, + 0.011380840092897415, + 0.002137048402801156, + 0.006834017112851143, + 0.0018831421621143818, + 0.0018017542315647006, + 0.0018178947502747178, + 0.0006869611097499728, + 0.0021476782858371735, + 0.0012053068494424224, + 0.0010455737356096506, + 0.0013192882761359215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.005952369887381792 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634591285139322 + ], + "opposite_logprob_tokens": [ + -5.379634380340576 + ], + "logprob_reward": -0.004634591285139322, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634591285139322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051428284496068954, + 0.08367669582366943, + 0.08714953809976578, + 0.06339617073535919, + 0.4055551588535309, + 0.02871542051434517, + 0.1384977102279663, + 0.02553226239979267, + 0.09590236842632294, + 0.020146455615758896 + ], + "credit_map": [ + 0.006793785374611616, + 0.007118551060557365, + 0.005873479414731264, + 0.006911488249897957, + 0.006554031744599342, + 0.006383015774190426, + 0.005810603965073824, + 0.005983325652778149, + 0.009695740416646004, + 0.010201828554272652, + 0.021143030375242233, + 0.005574374925345182, + 0.009523692540824413, + 0.009767609648406506, + 0.006011114455759525, + 0.011759299784898758, + 0.01707175374031067, + 0.0073653943836688995, + 0.01669071987271309, + 0.008372700773179531, + 0.004882900509983301, + 0.01392304990440607, + 0.007162549998611212, + 0.011680463328957558, + 0.006325682159513235, + 0.005223918706178665, + 0.009202002547681332, + 0.009007005020976067, + 0.004606037866324186, + 0.008265741169452667, + 0.010601603426039219, + 0.010164172388613224, + 0.04130832105875015, + 0.06104172021150589, + 0.036341723054647446, + 0.06760351359844208, + 0.05543840304017067, + 0.057631660252809525, + 0.06162210926413536, + 0.024567676708102226, + 0.003147441428154707, + 0.00411986792460084, + 0.003346665296703577, + 0.0033944558817893267, + 0.004052003845572472, + 0.0021926136687397957, + 0.006011354736983776, + 0.0024510149378329515, + 0.016231203451752663, + 0.005246346350759268, + 0.047426339238882065, + 0.008483349345624447, + 0.005422850605100393, + 0.019366586580872536, + 0.009930291213095188, + 0.026390736922621727, + 0.003500513033941388, + 0.0027484965976327658, + 0.0030093241948634386, + 0.003455096622928977, + 0.00271306443028152, + 0.002958361292257905, + 0.004467995837330818, + 0.002679410157725215, + 0.008129450492560863, + 0.006445430219173431, + 0.031574346125125885, + 0.006428336258977652, + 0.005228973459452391, + 0.02118656225502491, + 0.005104329437017441, + 0.011804936453700066, + 0.002700381213799119, + 0.0027093421667814255, + 0.0031397335696965456, + 0.0012710875598713756, + 0.0037170331925153732, + 0.0021115834824740887, + 0.00224488927051425, + 0.002252403646707535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.004634591285139322 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808915242552757 + ], + "opposite_logprob_tokens": [ + -3.2880890369415283 + ], + "logprob_reward": -0.03808915242552757, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808915242552757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10943687707185745, + 0.04861605167388916, + 0.313914030790329, + 0.04977480694651604, + 0.16659995913505554, + 0.03261224180459976, + 0.09032049030065536, + 0.06725228577852249, + 0.07953153550624847, + 0.04194175451993942 + ], + "credit_map": [ + 0.01283700205385685, + 0.017828846350312233, + 0.016326580196619034, + 0.01478180754929781, + 0.013391649350523949, + 0.0117138447239995, + 0.01259821467101574, + 0.009958925656974316, + 0.009543134830892086, + 0.0038794544525444508, + 0.01177692599594593, + 0.0036397825460880995, + 0.004165761172771454, + 0.006354562472552061, + 0.003881563898175955, + 0.005374867934733629, + 0.05645616725087166, + 0.021347513422369957, + 0.0714719220995903, + 0.030774332582950592, + 0.014809352345764637, + 0.04757291078567505, + 0.03080707974731922, + 0.04067474603652954, + 0.007696591317653656, + 0.0035813467111438513, + 0.010854797437787056, + 0.004767017904669046, + 0.004399606958031654, + 0.005639221053570509, + 0.006054507102817297, + 0.006781714968383312, + 0.0156581848859787, + 0.025508975610136986, + 0.014945208095014095, + 0.0299359317868948, + 0.023749826475977898, + 0.018311064690351486, + 0.024712752550840378, + 0.013778012245893478, + 0.004441559314727783, + 0.00379308988340199, + 0.006526387762278318, + 0.0032324446365237236, + 0.004093771334737539, + 0.0029253773391246796, + 0.004610556177794933, + 0.0029890555888414383, + 0.01148923672735691, + 0.004273287486284971, + 0.02919171378016472, + 0.007335105910897255, + 0.004674774594604969, + 0.012054327875375748, + 0.006932401563972235, + 0.014369643293321133, + 0.008369720540940762, + 0.005254002287983894, + 0.007863475009799004, + 0.011287611909210682, + 0.005627135746181011, + 0.007640907075256109, + 0.013501187786459923, + 0.007708242163062096, + 0.007991345599293709, + 0.0067758988589048386, + 0.011916949413716793, + 0.007154212798923254, + 0.0078021869994699955, + 0.016144338995218277, + 0.009490391239523888, + 0.012256214395165443, + 0.005078604910522699, + 0.007054042536765337, + 0.005837107542902231, + 0.002157466486096382, + 0.006285140290856361, + 0.00567896431311965, + 0.0038525620475411415, + 0.005997867323458195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.03808915242552757 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a classic 16-bit pixel art style, a shaky overhead shot closely follows a character equipped with a massive rocket arm as they soar vertically at breakneck speed. Clad in tight armor composed of blue and gray pixels, the character's massive metal rocket arm blasts downward from the elbow with a massive cone of fiery white, yellow, and orange pixels. This pulsing flame becomes the sole source of light, its violent flickering light instantly illuminating the surrounding narrow passage walls, composed of dark brown and gray pixels, casting a constantly dancing, high-contrast play of light and shadow on the rough pixel surfaces. As the character soars, the unstable camera feed trembles violently from the immense propulsive force, and blurred pixel particles zip past beneath them, perfectly embodying the tense, powerful, and dynamic atmosphere of vertical advancement typical of retro games.", + "seed": 27866, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_7/index_6248_seed_27866_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08229012070223689, + "vlm_logprob": -0.08229012070223689, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.08229012070223689, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3308263123035431 + ], + "opposite_logprob_tokens": [ + -1.3308262825012207 + ], + "logprob_reward": -0.3308263123035431, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3308263123035431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3006713390350342, + 0.10490024089813232, + 0.053419359028339386, + 0.10844637453556061, + 0.07013709843158722, + 0.06050245463848114, + 0.06272506713867188, + 0.1026076152920723, + 0.04392801597714424, + 0.09266246110200882 + ], + "credit_map": [ + 0.04810564965009689, + 0.03616741672158241, + 0.06146150082349777, + 0.028340140357613564, + 0.03225112706422806, + 0.030298518016934395, + 0.023593489080667496, + 0.04045350104570389, + 0.013846142217516899, + 0.01372931431978941, + 0.01695670373737812, + 0.010965039022266865, + 0.013357465155422688, + 0.010027632117271423, + 0.01432269811630249, + 0.011695252731442451, + 0.006364045664668083, + 0.005434569902718067, + 0.0073415907099843025, + 0.0084531856700778, + 0.006833143066614866, + 0.00501276133581996, + 0.007111658807843924, + 0.006868405733257532, + 0.013859378173947334, + 0.0161451306194067, + 0.015561032108962536, + 0.01094161905348301, + 0.010436683893203735, + 0.009854929521679878, + 0.009306596592068672, + 0.022341007366776466, + 0.011482602916657925, + 0.009439238347113132, + 0.012623226270079613, + 0.0070188893005251884, + 0.009588683024048805, + 0.005711670964956284, + 0.006365617271512747, + 0.007907163351774216, + 0.009538950398564339, + 0.007045242935419083, + 0.008664561435580254, + 0.008907794021070004, + 0.006380169186741114, + 0.00579491863027215, + 0.004861678462475538, + 0.009309139102697372, + 0.007953193970024586, + 0.0059841410256922245, + 0.008115814998745918, + 0.008921978063881397, + 0.010296547785401344, + 0.005864895414561033, + 0.007422169204801321, + 0.008166329935193062, + 0.014383687637746334, + 0.0188769344240427, + 0.012744984589517117, + 0.008768090978264809, + 0.012012908235192299, + 0.010256645269691944, + 0.01448887214064598, + 0.011075488291680813, + 0.007378720212727785, + 0.005501600913703442, + 0.006842784117907286, + 0.004484014585614204, + 0.003760836785659194, + 0.004010471981018782, + 0.004535339307039976, + 0.007414249703288078, + 0.009635727852582932, + 0.01784241758286953, + 0.011814833618700504, + 0.012782466597855091, + 0.009167803451418877, + 0.009686794131994247, + 0.007788931950926781, + 0.013943477533757687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.3308263123035431 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499119311571121 + ], + "opposite_logprob_tokens": [ + -2.9299912452697754 + ], + "logprob_reward": -0.05499119311571121, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499119311571121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17303645610809326, + 0.11205479502677917, + 0.1168571263551712, + 0.09171054512262344, + 0.06464552879333496, + 0.07849013060331345, + 0.1196032389998436, + 0.1491643190383911, + 0.050481826066970825, + 0.043956026434898376 + ], + "credit_map": [ + 0.03301117941737175, + 0.023768160492181778, + 0.041196487843990326, + 0.0175225380808115, + 0.012380806729197502, + 0.012512480840086937, + 0.014564028941094875, + 0.01808076724410057, + 0.012604596093297005, + 0.016279155388474464, + 0.01641148515045643, + 0.0141886156052351, + 0.013928631320595741, + 0.011529584415256977, + 0.015409902669489384, + 0.01170282531529665, + 0.012093831785023212, + 0.011262058280408382, + 0.014526044949889183, + 0.016645586118102074, + 0.012005993165075779, + 0.015856793150305748, + 0.018739977851510048, + 0.015726841986179352, + 0.012057946063578129, + 0.011788581497967243, + 0.011411809362471104, + 0.010183078236877918, + 0.012608147226274014, + 0.007220571860671043, + 0.013581255450844765, + 0.012859160080552101, + 0.009661774151027203, + 0.008957669138908386, + 0.011701707728207111, + 0.005293742753565311, + 0.009763692505657673, + 0.0059422883205115795, + 0.005993187893182039, + 0.007331463973969221, + 0.00986310001462698, + 0.008649485185742378, + 0.009483005851507187, + 0.010707579553127289, + 0.009324234910309315, + 0.012253559194505215, + 0.007864469662308693, + 0.010344690643250942, + 0.011888805776834488, + 0.014994346536695957, + 0.013206087052822113, + 0.019149361178278923, + 0.019345765933394432, + 0.010007087141275406, + 0.01668470911681652, + 0.01432708092033863, + 0.014060822315514088, + 0.018050992861390114, + 0.053856004029512405, + 0.007772533688694239, + 0.0056398301385343075, + 0.0060653588734567165, + 0.010223534889519215, + 0.03349524736404419, + 0.00687182554975152, + 0.007825145497918129, + 0.006758385803550482, + 0.00610960042104125, + 0.0047492533922195435, + 0.005825027357786894, + 0.004638313315808773, + 0.007704275660216808, + 0.003455051453784108, + 0.0092484625056386, + 0.00563086848706007, + 0.005430411547422409, + 0.004643094260245562, + 0.00613565556704998, + 0.0037103660870343447, + 0.005702116061002016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.05499119311571121 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036214501596987247 + ], + "opposite_logprob_tokens": [ + -5.628621578216553 + ], + "logprob_reward": -0.0036214501596987247, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036214501596987247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3064587712287903, + 0.07080908864736557, + 0.07529070973396301, + 0.08478530496358871, + 0.0352696068584919, + 0.09884297102689743, + 0.10790350288152695, + 0.13853541016578674, + 0.032436270266771317, + 0.0496683195233345 + ], + "credit_map": [ + 0.04105003550648689, + 0.041637785732746124, + 0.049256909638643265, + 0.030178379267454147, + 0.040635835379362106, + 0.04046788439154625, + 0.03299623727798462, + 0.030235694721341133, + 0.010665650479495525, + 0.00994886551052332, + 0.012699305079877377, + 0.007768536917865276, + 0.008963783271610737, + 0.007454643491655588, + 0.006312709767371416, + 0.006995597388595343, + 0.009979169815778732, + 0.007965472526848316, + 0.014132337644696236, + 0.011770970188081264, + 0.008643768727779388, + 0.006584467366337776, + 0.006388664245605469, + 0.009825858287513256, + 0.008958183228969574, + 0.010220348834991455, + 0.006927346345037222, + 0.00985129363834858, + 0.0105748251080513, + 0.007155371364206076, + 0.013018187135457993, + 0.01807975023984909, + 0.005004371982067823, + 0.004945963155478239, + 0.006247845012694597, + 0.004047588910907507, + 0.003955175168812275, + 0.0035417077597230673, + 0.003518733661621809, + 0.004008222371339798, + 0.00999078806489706, + 0.011768236756324768, + 0.01502431370317936, + 0.009546168148517609, + 0.012655286118388176, + 0.015369926579296589, + 0.00908613484352827, + 0.01540212519466877, + 0.01245745737105608, + 0.014800826087594032, + 0.015182766132056713, + 0.014314827509224415, + 0.015723824501037598, + 0.009893790818750858, + 0.011066414415836334, + 0.01446359883993864, + 0.012816541828215122, + 0.018617883324623108, + 0.03704354166984558, + 0.00896826945245266, + 0.010018540546298027, + 0.008284127339720726, + 0.014430847950279713, + 0.028355663642287254, + 0.0042961761355400085, + 0.004375990480184555, + 0.004203231073915958, + 0.004159478470683098, + 0.0034011697862297297, + 0.002634056843817234, + 0.0031619800720363855, + 0.006204187870025635, + 0.0055245161056518555, + 0.008332004770636559, + 0.007387073710560799, + 0.006175046321004629, + 0.006658631842583418, + 0.005257369484752417, + 0.004288105759769678, + 0.006045570597052574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0036214501596987247 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064966097474098 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016064966097474098, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064966097474098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23445679247379303, + 0.15027329325675964, + 0.06225348263978958, + 0.0708780363202095, + 0.05713783949613571, + 0.17570751905441284, + 0.029923420399427414, + 0.07550384849309921, + 0.0657937079668045, + 0.07807207107543945 + ], + "credit_map": [ + 0.03845610469579697, + 0.03136875107884407, + 0.04190300032496452, + 0.01967238076031208, + 0.02750633843243122, + 0.022829322144389153, + 0.02341338060796261, + 0.029307516291737556, + 0.01449536718428135, + 0.020477402955293655, + 0.020509446039795876, + 0.017108548432588577, + 0.024803636595606804, + 0.02007303014397621, + 0.01793241687119007, + 0.014873440377414227, + 0.007650080602616072, + 0.0066482555121183395, + 0.009660246782004833, + 0.01009158045053482, + 0.007804002612829208, + 0.00584942614659667, + 0.006052032113075256, + 0.008497858420014381, + 0.009243390522897243, + 0.008712097071111202, + 0.006771563086658716, + 0.006898717954754829, + 0.0074665602296590805, + 0.007054563611745834, + 0.008005386218428612, + 0.01672576181590557, + 0.00942762941122055, + 0.007323738653212786, + 0.012574070133268833, + 0.005123607814311981, + 0.006425453815609217, + 0.005738013423979282, + 0.005109085701406002, + 0.005416239611804485, + 0.016723070293664932, + 0.02175690233707428, + 0.024970559403300285, + 0.02214237116277218, + 0.017137041315436363, + 0.02285483293235302, + 0.022864453494548798, + 0.027258291840553284, + 0.0030230970587581396, + 0.0031899460591375828, + 0.003748406656086445, + 0.004251448437571526, + 0.004909561015665531, + 0.00301552377641201, + 0.0034581124782562256, + 0.004327322356402874, + 0.007576686795800924, + 0.008436220698058605, + 0.024279562756419182, + 0.004887539427727461, + 0.004801617003977299, + 0.003562386380508542, + 0.0062380339950323105, + 0.01572180725634098, + 0.007751959841698408, + 0.00902523472905159, + 0.01066849660128355, + 0.009335752576589584, + 0.00877812597900629, + 0.004629913717508316, + 0.007519373204559088, + 0.008084855042397976, + 0.008834784850478172, + 0.015391487628221512, + 0.010473798029124737, + 0.01108547393232584, + 0.009817419573664665, + 0.00755843473598361, + 0.006621527019888163, + 0.008289145305752754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.016064966097474098 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946681834757328 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946681834757328, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946681834757328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19440943002700806, + 0.2678855061531067, + 0.02766614966094494, + 0.053716592490673065, + 0.0737420991063118, + 0.09574232995510101, + 0.037647794932127, + 0.18176697194576263, + 0.029454896226525307, + 0.03796821087598801 + ], + "credit_map": [ + 0.028592566028237343, + 0.02404051460325718, + 0.05213293805718422, + 0.020639052614569664, + 0.016100728884339333, + 0.013906332664191723, + 0.0152309350669384, + 0.02376635931432247, + 0.03471203148365021, + 0.03737040236592293, + 0.040200140327215195, + 0.029006626456975937, + 0.0395471528172493, + 0.03154062479734421, + 0.03204372525215149, + 0.023464815691113472, + 0.003968397155404091, + 0.003190575400367379, + 0.006026791874319315, + 0.0031904487404972315, + 0.0026137372478842735, + 0.0026857624761760235, + 0.002382182516157627, + 0.003608253551647067, + 0.007459585089236498, + 0.008192701265215874, + 0.008588182739913464, + 0.004917493090033531, + 0.005067876540124416, + 0.0044555459171533585, + 0.004271460697054863, + 0.01076374389231205, + 0.011495810002088547, + 0.00881593395024538, + 0.015063519589602947, + 0.006578030996024609, + 0.008805638179183006, + 0.007414660882204771, + 0.007000393234193325, + 0.008568113669753075, + 0.014399632811546326, + 0.010316134430468082, + 0.01135060004889965, + 0.012707886286079884, + 0.012053078971803188, + 0.008800934068858624, + 0.01023842766880989, + 0.01587563380599022, + 0.004721234552562237, + 0.0046708243899047375, + 0.0058783977292478085, + 0.00405581621453166, + 0.005632850807160139, + 0.0038991670589894056, + 0.003481485415250063, + 0.005308021791279316, + 0.016558432951569557, + 0.010712300427258015, + 0.07157012820243835, + 0.006122693885117769, + 0.004595276899635792, + 0.005791745614260435, + 0.005054751876741648, + 0.061361636966466904, + 0.003490916918963194, + 0.004611711483448744, + 0.0036754668690264225, + 0.0031352383084595203, + 0.004071149043738842, + 0.002920253435149789, + 0.0036598066799342632, + 0.0038903539534658194, + 0.0056129079312086105, + 0.0053729526698589325, + 0.004349451046437025, + 0.004613442346453667, + 0.004994944203644991, + 0.00424716854467988, + 0.004311488475650549, + 0.004465856123715639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.005946681834757328 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a classic 16-bit pixel art style, a shaky overhead shot closely follows a character equipped with a massive rocket arm as they soar vertically at breakneck speed. Clad in tight armor composed of blue and gray pixels, the character's massive metal rocket arm blasts downward from the elbow with a massive cone of fiery white, yellow, and orange pixels. This pulsing flame becomes the sole source of light, its violent flickering light instantly illuminating the surrounding narrow passage walls, composed of dark brown and gray pixels, casting a constantly dancing, high-contrast play of light and shadow on the rough pixel surfaces. As the character soars, the unstable camera feed trembles violently from the immense propulsive force, and blurred pixel particles zip past beneath them, perfectly embodying the tense, powerful, and dynamic atmosphere of vertical advancement typical of retro games.", + "seed": 27867, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_7/index_6248_seed_27867_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018219967698678376, + "vlm_logprob": -0.018219967698678376, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.018219967698678376, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022021062672138214 + ], + "opposite_logprob_tokens": [ + -4.147021293640137 + ], + "logprob_reward": -0.022021062672138214, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022021062672138214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43167856335639954, + 0.06975316256284714, + 0.060805127024650574, + 0.038670942187309265, + 0.025296170264482498, + 0.020933760330080986, + 0.037264615297317505, + 0.1501826047897339, + 0.034884750843048096, + 0.13053035736083984 + ], + "credit_map": [ + 0.036731936037540436, + 0.019058892503380775, + 0.12756115198135376, + 0.014525385573506355, + 0.010800606571137905, + 0.032963063567876816, + 0.01735210232436657, + 0.1726853847503662, + 0.011450733989477158, + 0.0031078362371772528, + 0.015289291739463806, + 0.004176023416221142, + 0.0025592688471078873, + 0.007839699275791645, + 0.00474823173135519, + 0.020582083612680435, + 0.0066862585954368114, + 0.002668623114004731, + 0.015454535372555256, + 0.0032024008687585592, + 0.0022515705786645412, + 0.007711907848715782, + 0.0037403185851871967, + 0.019089510664343834, + 0.007068617735058069, + 0.0021243931259959936, + 0.00520562008023262, + 0.004044323693960905, + 0.004996918141841888, + 0.006277040578424931, + 0.003267450723797083, + 0.005686579737812281, + 0.003591001033782959, + 0.003126920899376273, + 0.0017812910955399275, + 0.0014044103445485234, + 0.004206849727779627, + 0.004973581060767174, + 0.0017738755559548736, + 0.004438241943717003, + 0.0029195488896220922, + 0.0017463635886088014, + 0.0015062635065987706, + 0.0016991342417895794, + 0.002260490320622921, + 0.003487280337139964, + 0.002387415850535035, + 0.004927263129502535, + 0.00500686839222908, + 0.004698743112385273, + 0.0043104770593345165, + 0.003476268146187067, + 0.005672958213835955, + 0.004841241519898176, + 0.004561276640743017, + 0.004696781747043133, + 0.02114608697593212, + 0.03426184877753258, + 0.024179397150874138, + 0.00886828638613224, + 0.015457343310117722, + 0.011508765630424023, + 0.013456961140036583, + 0.021303927525877953, + 0.005799018777906895, + 0.0038656441029161215, + 0.0055525219067931175, + 0.005030347034335136, + 0.0030393002089112997, + 0.002842287067323923, + 0.003676465479657054, + 0.005079165101051331, + 0.019214555621147156, + 0.014040607959032059, + 0.018347641453146935, + 0.020418206229805946, + 0.00952113326638937, + 0.016834022477269173, + 0.01539693959057331, + 0.016757255420088768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.022021062672138214 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005269447807222605 + ], + "opposite_logprob_tokens": [ + -5.255269527435303 + ], + "logprob_reward": -0.005269447807222605, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005269447807222605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1597108393907547, + 0.06800901144742966, + 0.04790448024868965, + 0.03524439036846161, + 0.029813064262270927, + 0.056715212762355804, + 0.02518562041223049, + 0.14858713746070862, + 0.0466744489967823, + 0.38215577602386475 + ], + "credit_map": [ + 0.019720114767551422, + 0.0269926767796278, + 0.01738191954791546, + 0.016615211963653564, + 0.019731659442186356, + 0.02175423502922058, + 0.01916591450572014, + 0.018349120393395424, + 0.011932093650102615, + 0.00406999746337533, + 0.017605025321245193, + 0.004761821590363979, + 0.002720295451581478, + 0.006275113672018051, + 0.004524656105786562, + 0.01612001098692417, + 0.007430588360875845, + 0.002480483613908291, + 0.008202231489121914, + 0.0026188951451331377, + 0.001733678043819964, + 0.011798583902418613, + 0.004021587781608105, + 0.009618429467082024, + 0.009733080863952637, + 0.001717297825962305, + 0.005704707931727171, + 0.00279626646079123, + 0.0019989903084933758, + 0.005491999443620443, + 0.002175613772124052, + 0.005626436322927475, + 0.0037394720129668713, + 0.003848354099318385, + 0.0025596958585083485, + 0.0018285199766978621, + 0.0047888499684631824, + 0.006482300348579884, + 0.002747567603364587, + 0.0038183042779564857, + 0.008029399439692497, + 0.006135368254035711, + 0.004170658066868782, + 0.003895528381690383, + 0.006157641764730215, + 0.010220224037766457, + 0.006015399005264044, + 0.012090994976460934, + 0.0040443530306220055, + 0.002790536265820265, + 0.0028816317208111286, + 0.0021166589576750994, + 0.002560393651947379, + 0.003408091375604272, + 0.002141371136531234, + 0.005242583807557821, + 0.02529522404074669, + 0.026460951194167137, + 0.022306079044938087, + 0.011277753859758377, + 0.017060691490769386, + 0.01100875809788704, + 0.016911162063479424, + 0.01826651208102703, + 0.00842155423015356, + 0.00502793351188302, + 0.0067126331850886345, + 0.0052547212690114975, + 0.006531350314617157, + 0.0038845445960760117, + 0.005284925922751427, + 0.005556787829846144, + 0.06015045940876007, + 0.031184706836938858, + 0.07837165892124176, + 0.022117270156741142, + 0.04188935086131096, + 0.05331399664282799, + 0.03406742215156555, + 0.06106092408299446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.005269447807222605 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003617292968556285 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.003617292968556285, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003617292968556285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11966216564178467, + 0.11179143190383911, + 0.059670817106962204, + 0.04746489226818085, + 0.03872278705239296, + 0.13284443318843842, + 0.06344465166330338, + 0.18657033145427704, + 0.09970331937074661, + 0.14012514054775238 + ], + "credit_map": [ + 0.01581072434782982, + 0.015215592458844185, + 0.020113425329327583, + 0.009682543575763702, + 0.010736692696809769, + 0.009771245531737804, + 0.010302533395588398, + 0.028029412031173706, + 0.017352933064103127, + 0.007161505986005068, + 0.03534318879246712, + 0.0066679236479103565, + 0.0044074151664972305, + 0.009597418829798698, + 0.006525143980979919, + 0.024735914543271065, + 0.009314334020018578, + 0.0033867876045405865, + 0.015032262541353703, + 0.003270830260589719, + 0.0027913080994039774, + 0.005244569852948189, + 0.0034787089098244905, + 0.01715202070772648, + 0.012604877352714539, + 0.00236577563919127, + 0.006878126412630081, + 0.003058220725506544, + 0.0044179498217999935, + 0.007867933250963688, + 0.0025277917739003897, + 0.007744221482425928, + 0.006233177613466978, + 0.00403648940846324, + 0.004020103719085455, + 0.0026721348986029625, + 0.006023648660629988, + 0.005994528066366911, + 0.002676107222214341, + 0.0070665995590388775, + 0.019795091822743416, + 0.01465024333447218, + 0.009491873905062675, + 0.00987693015486002, + 0.01310802809894085, + 0.028131209313869476, + 0.016705626621842384, + 0.021085456013679504, + 0.009590012952685356, + 0.007732948288321495, + 0.00597824901342392, + 0.006348862778395414, + 0.006529026199132204, + 0.009690440259873867, + 0.008977965451776981, + 0.008597154170274734, + 0.03431505337357521, + 0.041500456631183624, + 0.023273980244994164, + 0.012147238478064537, + 0.017519768327474594, + 0.017411302775144577, + 0.016210218891501427, + 0.024192335084080696, + 0.017605755478143692, + 0.011279759928584099, + 0.016394924372434616, + 0.009407016448676586, + 0.009710774756968021, + 0.008301131427288055, + 0.012440534308552742, + 0.014563428238034248, + 0.022604404017329216, + 0.016467472538352013, + 0.02497437596321106, + 0.009918605908751488, + 0.014559944160282612, + 0.01709526777267456, + 0.012135778553783894, + 0.02236931584775448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.003617292968556285 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248458590358496 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248458590358496, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248458590358496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18221592903137207, + 0.03392639011144638, + 0.04507248103618622, + 0.04440344497561455, + 0.04144910350441933, + 0.07004455476999283, + 0.028724143281579018, + 0.05423566326498985, + 0.07564219832420349, + 0.4242861866950989 + ], + "credit_map": [ + 0.01774967461824417, + 0.008921468630433083, + 0.048770830035209656, + 0.0071688126772642136, + 0.005911760963499546, + 0.015011626295745373, + 0.007427703123539686, + 0.07125404477119446, + 0.006053579039871693, + 0.0021568117663264275, + 0.008979675360023975, + 0.0023750157561153173, + 0.0014095958322286606, + 0.00264762993901968, + 0.002672805916517973, + 0.007631274871528149, + 0.005812402348965406, + 0.002519322792068124, + 0.009723452851176262, + 0.002633053809404373, + 0.0017115396913141012, + 0.007289147470146418, + 0.002773531712591648, + 0.01261003315448761, + 0.011347741819918156, + 0.001737403217703104, + 0.007451778277754784, + 0.002564982743933797, + 0.0024614876601845026, + 0.007726577110588551, + 0.0021668854169547558, + 0.00894659198820591, + 0.005213548429310322, + 0.006327909417450428, + 0.0032487420830875635, + 0.0028730141930282116, + 0.006591299548745155, + 0.006415640935301781, + 0.0032193725928664207, + 0.007559578865766525, + 0.011338604614138603, + 0.006297426298260689, + 0.004380345810204744, + 0.004443669691681862, + 0.00586532149463892, + 0.013921473175287247, + 0.01089548971503973, + 0.012902230955660343, + 0.004124361090362072, + 0.0037146522663533688, + 0.0031857057474553585, + 0.0023686906788498163, + 0.0028332883957773447, + 0.0033579403534531593, + 0.003320388263091445, + 0.005819117650389671, + 0.008745240978896618, + 0.009496595710515976, + 0.009378806687891483, + 0.003851216286420822, + 0.0035604601725935936, + 0.005100558511912823, + 0.003558996831998229, + 0.010543784126639366, + 0.015677744522690773, + 0.0075479065999388695, + 0.013676554895937443, + 0.006769286934286356, + 0.006748710293322802, + 0.007112209685146809, + 0.00672096386551857, + 0.011388820596039295, + 0.10824451595544815, + 0.0335809700191021, + 0.11843041330575943, + 0.017899109050631523, + 0.015543977729976177, + 0.027979081496596336, + 0.01836046390235424, + 0.08424762636423111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.005248458590358496 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a classic 16-bit pixel art style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character clad in tight armor composed of blue and gray pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character possess a massive metal rocket arm that blasts downward from the elbow with a cone of fiery white, yellow, and orange pixels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a narrow passage with walls composed of dark brown and gray pixels illuminated by the pulsing flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and overhead, trembling violently as the character soars vertically?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494357645511627 + ], + "opposite_logprob_tokens": [ + -2.929943561553955 + ], + "logprob_reward": -0.05494357645511627, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494357645511627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19605684280395508, + 0.05827277898788452, + 0.0729137659072876, + 0.06342286616563797, + 0.04709772393107414, + 0.08463748544454575, + 0.09148640185594559, + 0.12756063044071198, + 0.11075334250926971, + 0.14779816567897797 + ], + "credit_map": [ + 0.01835642382502556, + 0.011236945167183876, + 0.05381555110216141, + 0.008622517809271812, + 0.006156022660434246, + 0.01686958782374859, + 0.009052613750100136, + 0.07194717228412628, + 0.009810072369873524, + 0.004812262021005154, + 0.008194134570658207, + 0.005121751222759485, + 0.0029356079176068306, + 0.008630175143480301, + 0.008040372282266617, + 0.010728402994573116, + 0.014004899188876152, + 0.00418020598590374, + 0.014164893887937069, + 0.00475816335529089, + 0.003186700399965048, + 0.01654907688498497, + 0.005492693744599819, + 0.010577131062746048, + 0.01861507259309292, + 0.002127348445355892, + 0.009598812088370323, + 0.0027086185291409492, + 0.0029633883386850357, + 0.011630079708993435, + 0.003379140980541706, + 0.012400403618812561, + 0.006950438022613525, + 0.005164850037544966, + 0.004154062829911709, + 0.0028372181113809347, + 0.006590993609279394, + 0.009196106344461441, + 0.004191513638943434, + 0.008012539707124233, + 0.013779115863144398, + 0.009514395147562027, + 0.005190063267946243, + 0.004006002098321915, + 0.007376696448773146, + 0.01440590713173151, + 0.012468932196497917, + 0.017896370962262154, + 0.014951424673199654, + 0.01174518559128046, + 0.013486461713910103, + 0.013293550349771976, + 0.004455935675650835, + 0.00879612099379301, + 0.009021854028105736, + 0.015735864639282227, + 0.019465815275907516, + 0.0328153632581234, + 0.02201816812157631, + 0.008099393919110298, + 0.009424660354852676, + 0.008843954652547836, + 0.007698137778788805, + 0.019195139408111572, + 0.01898997090756893, + 0.011790025979280472, + 0.022227149456739426, + 0.009862788021564484, + 0.008200337179005146, + 0.008566823787987232, + 0.012360486201941967, + 0.018755760043859482, + 0.022568916901946068, + 0.012565742246806622, + 0.023900507017970085, + 0.01413533091545105, + 0.016534993425011635, + 0.01959327980875969, + 0.018887188285589218, + 0.01961219683289528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.05494357645511627 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_8/index_tensor([4689, 4689])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_8/index_tensor([4689, 4689])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0a644ec2f64b723d972b52a10455d7e08799b96b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_8/index_tensor([4689, 4689])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In Makoto Shinkai's distinctive artistic style, a steady and deliberate push-in slowly pans across a narrow, rain-slicked concrete balcony ledge, perched high above a vast, futuristic cityscape. Below, a dizzying chasm of towering skyscrapers crisscrosses, lined with glowing skytrain tracks, their glass and steel facades drenched in the ethereal glow of giant, flickering holographic advertisements. The scene is dominated by a cool, nocturnal palette of deep blues and purples, punctuated by vibrant neon pinks and cyans that reflect with surreal clarity off every wet surface, casting long, dramatic shadows that dance with the shifting light. On the ledge, two cats are the focus of an intimate courtship ritual. One of the male cats, with sleek black fur that gleams in the multicolored light, cautiously extends its paw and delicately pushes forward a brightly colored fish-shaped kitten as an offering. Its posture is low, its gaze unwavering, sincere and hopeful. The recipient, a calico cat with soft, fluffy white and ginger fur, hesitates, its large, inquisitive green eyes wide, leaning forward slightly, its whiskers twitching in the misty air. The entire scene is imbued with a quiet, romantic melancholy. The gentle, small-scale interaction between the two animals stands in stark contrast to the overwhelming scale and impersonal beauty of this vibrant metropolis, capturing fleeting connections in a vast and indifferent world.", + "seed": 17533, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_8/index_4689_seed_17533_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.017181041464209557, + "vlm_logprob": -0.017181041464209557, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.017181041464209557, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01575709879398346 + ], + "opposite_logprob_tokens": [ + -4.515757083892822 + ], + "logprob_reward": -0.01575709879398346, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01575709879398346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1866840273141861, + 0.18935422599315643, + 0.11644664406776428, + 0.0697544515132904, + 0.1333371251821518, + 0.048044200986623764, + 0.10257948935031891, + 0.049554046243429184, + 0.09252225607633591, + 0.01172348391264677 + ], + "credit_map": [ + 0.021639274433255196, + 0.032305262982845306, + 0.03660324588418007, + 0.021828599274158478, + 0.018824037164449692, + 0.020077185705304146, + 0.017865808680653572, + 0.01754063367843628, + 0.008846071548759937, + 0.009479563683271408, + 0.005079157184809446, + 0.006177433766424656, + 0.07190189510583878, + 0.07411044090986252, + 0.0066915238276124, + 0.007068153005093336, + 0.019440405070781708, + 0.024853169918060303, + 0.01237883698195219, + 0.017373614013195038, + 0.007685265503823757, + 0.009311954490840435, + 0.01167507003992796, + 0.013728341087698936, + 0.006150497123599052, + 0.005068425089120865, + 0.018947428092360497, + 0.007166911382228136, + 0.005157460458576679, + 0.006754804868251085, + 0.004161206539720297, + 0.016347726806998253, + 0.015697835013270378, + 0.009389009326696396, + 0.018244272097945213, + 0.009297510609030724, + 0.014225083403289318, + 0.040215201675891876, + 0.008410153910517693, + 0.017858080565929413, + 0.005122171249240637, + 0.004544886760413647, + 0.006258797831833363, + 0.008802695199847221, + 0.0044385953806340694, + 0.0053025055676698685, + 0.007921317592263222, + 0.005653237458318472, + 0.011857596226036549, + 0.014489363878965378, + 0.01335857342928648, + 0.016268707811832428, + 0.012115736491978168, + 0.009455259889364243, + 0.010582285933196545, + 0.014451972208917141, + 0.007002315949648619, + 0.0054647475481033325, + 0.004046634770929813, + 0.0038131901528686285, + 0.008652925491333008, + 0.008106465451419353, + 0.0043700458481907845, + 0.008097724057734013, + 0.013955564238131046, + 0.00504772225394845, + 0.03020533174276352, + 0.005872646812349558, + 0.0043773045763373375, + 0.005910936277359724, + 0.019269369542598724, + 0.007883393205702305, + 0.0015063175233080983, + 0.001593808294273913, + 0.0010392967378720641, + 0.0011781268985942006, + 0.0019916456658393145, + 0.0013786079362034798, + 0.001403822680003941, + 0.0016318599227815866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.01575709879398346 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002498839981853962 + ], + "opposite_logprob_tokens": [ + -6.002498626708984 + ], + "logprob_reward": -0.002498839981853962, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002498839981853962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11644633114337921, + 0.208634153008461, + 0.11475742608308792, + 0.1209736317396164, + 0.08225038647651672, + 0.037484414875507355, + 0.12339093536138535, + 0.10042035579681396, + 0.041568707674741745, + 0.05407366529107094 + ], + "credit_map": [ + 0.01070772297680378, + 0.013901001773774624, + 0.011453509330749512, + 0.01458656694740057, + 0.01089802011847496, + 0.015220356173813343, + 0.013518987223505974, + 0.026160171255469322, + 0.010703499428927898, + 0.014246124774217606, + 0.04075438156723976, + 0.010159449651837349, + 0.07940134406089783, + 0.03486594185233116, + 0.011587036773562431, + 0.006916363723576069, + 0.009197951294481754, + 0.019630417227745056, + 0.03285019099712372, + 0.012844868004322052, + 0.006435239687561989, + 0.01124603021889925, + 0.012220434844493866, + 0.010332290083169937, + 0.011115706525743008, + 0.012593147344887257, + 0.033717382699251175, + 0.01173356268554926, + 0.012728052213788033, + 0.007803373970091343, + 0.007122371811419725, + 0.024160027503967285, + 0.011732993647456169, + 0.00815568957477808, + 0.011930383741855621, + 0.010242505930364132, + 0.011101474985480309, + 0.013057057745754719, + 0.007363129872828722, + 0.008667148649692535, + 0.0041088867001235485, + 0.0030866716988384724, + 0.005524441599845886, + 0.006140772718936205, + 0.004390182439237833, + 0.004902273882180452, + 0.004911802243441343, + 0.0044193812645971775, + 0.010077424347400665, + 0.007987769320607185, + 0.0163615383207798, + 0.04197017848491669, + 0.0070484415628015995, + 0.009545406326651573, + 0.017693404108285904, + 0.012706772424280643, + 0.013521634973585606, + 0.007514616008847952, + 0.012816098518669605, + 0.0077735078521072865, + 0.017389440909028053, + 0.014791551977396011, + 0.011229797266423702, + 0.01538370456546545, + 0.006378596648573875, + 0.00515646580606699, + 0.004989278502762318, + 0.0032627710606902838, + 0.006768250372260809, + 0.004642578307539225, + 0.004832442384213209, + 0.005538326688110828, + 0.01602327637374401, + 0.00612558051943779, + 0.0030899853445589542, + 0.0037870893720537424, + 0.003982082474976778, + 0.004605459980666637, + 0.01141357421875, + 0.00504661537706852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.002498839981853962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336938351392746 + ], + "opposite_logprob_tokens": [ + -3.408693790435791 + ], + "logprob_reward": -0.0336938351392746, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336938351392746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09093301743268967, + 0.2427559345960617, + 0.0731574147939682, + 0.18564671277999878, + 0.07504157721996307, + 0.0878577008843422, + 0.07530267536640167, + 0.049398574978113174, + 0.07548799365758896, + 0.04441843554377556 + ], + "credit_map": [ + 0.008513585664331913, + 0.008815765380859375, + 0.01401084940880537, + 0.012856613844633102, + 0.008545184507966042, + 0.011496254242956638, + 0.01395420916378498, + 0.01274054404348135, + 0.010726031847298145, + 0.01424463465809822, + 0.03732474893331528, + 0.0077260080724954605, + 0.09782450646162033, + 0.05795090273022652, + 0.008461283519864082, + 0.008497822098433971, + 0.007659093476831913, + 0.008150817826390266, + 0.02224184013903141, + 0.006495621055364609, + 0.005861644167453051, + 0.00646072905510664, + 0.008604017086327076, + 0.00768365478143096, + 0.015371764078736305, + 0.011370128020644188, + 0.06839890033006668, + 0.016396895051002502, + 0.009387950412929058, + 0.010161645710468292, + 0.007251002360135317, + 0.04730841889977455, + 0.00998566672205925, + 0.008681682869791985, + 0.009477062150835991, + 0.010817460715770721, + 0.011135983280837536, + 0.008628609590232372, + 0.008696237578988075, + 0.007618879899382591, + 0.008847125805914402, + 0.009003553539514542, + 0.011230051517486572, + 0.011774192564189434, + 0.01188851147890091, + 0.008043820969760418, + 0.017039168626070023, + 0.010031278245151043, + 0.008394146338105202, + 0.011393815279006958, + 0.009376428090035915, + 0.01507179532200098, + 0.008336539380252361, + 0.005821064580231905, + 0.009968302212655544, + 0.006940580438822508, + 0.006884473375976086, + 0.004602185916155577, + 0.006088798400014639, + 0.005733453668653965, + 0.008283812552690506, + 0.006411293521523476, + 0.004747422877699137, + 0.006647137925028801, + 0.012404194101691246, + 0.005372122395783663, + 0.022207682952284813, + 0.004271994810551405, + 0.006285542622208595, + 0.0061331214383244514, + 0.011676077730953693, + 0.007137257605791092, + 0.008918552659451962, + 0.006458327639847994, + 0.004725096747279167, + 0.0035289411898702383, + 0.00508884759619832, + 0.004582608584314585, + 0.006025985348969698, + 0.0050900778733193874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0336938351392746 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076242173090577126 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.0076242173090577126, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076242173090577126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1496949940919876, + 0.2546829581260681, + 0.07387930899858475, + 0.10749374330043793, + 0.059685882180929184, + 0.05357397720217705, + 0.03220869600772858, + 0.030957140028476715, + 0.19809924066066742, + 0.03972398862242699 + ], + "credit_map": [ + 0.01292905118316412, + 0.024587957188487053, + 0.019142184406518936, + 0.02064412459731102, + 0.022585254162549973, + 0.010795370675623417, + 0.020511237904429436, + 0.018499813973903656, + 0.019791478291153908, + 0.023158907890319824, + 0.07507731765508652, + 0.012319565750658512, + 0.03540492430329323, + 0.06725288927555084, + 0.011218021623790264, + 0.01045986544340849, + 0.011827307753264904, + 0.016198892146348953, + 0.008588597178459167, + 0.00897988397628069, + 0.005479247774928808, + 0.0070154620334506035, + 0.006797484587877989, + 0.008992435410618782, + 0.00872926414012909, + 0.006420583929866552, + 0.03784649446606636, + 0.010722462087869644, + 0.007155028637498617, + 0.005781924352049828, + 0.004867095500230789, + 0.02597089298069477, + 0.005799474660307169, + 0.0034667826257646084, + 0.00718699861317873, + 0.005585792008787394, + 0.006177210249006748, + 0.019707931205630302, + 0.003844472812488675, + 0.007917221635580063, + 0.005459384992718697, + 0.004525112919509411, + 0.005352644715458155, + 0.0076209185644984245, + 0.006509872619062662, + 0.005664474330842495, + 0.01125911995768547, + 0.007182449568063021, + 0.003947759047150612, + 0.00490393815562129, + 0.004211433231830597, + 0.0040570893324911594, + 0.004116986878216267, + 0.002522109542042017, + 0.004870887845754623, + 0.003578490810468793, + 0.004108039662241936, + 0.0037674333434551954, + 0.0021318108774721622, + 0.002980150282382965, + 0.005007531028240919, + 0.004943516105413437, + 0.0031727240420877934, + 0.0048459367826581, + 0.036989662796258926, + 0.011050580069422722, + 0.0729922354221344, + 0.012119279243052006, + 0.009774245321750641, + 0.012249475345015526, + 0.030197907239198685, + 0.01272586826235056, + 0.012606600299477577, + 0.003284215461462736, + 0.003442128887400031, + 0.0027355824131518602, + 0.003525955369696021, + 0.003641195595264435, + 0.006976200267672539, + 0.0035121093969792128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.0076242173090577126 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026331216096878052 + ], + "opposite_logprob_tokens": [ + -3.6513311862945557 + ], + "logprob_reward": -0.026331216096878052, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026331216096878052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1794765144586563, + 0.08522600680589676, + 0.08309456706047058, + 0.1930568963289261, + 0.06374801695346832, + 0.0765838623046875, + 0.09104844182729721, + 0.07933671027421951, + 0.11912950128316879, + 0.029299510642886162 + ], + "credit_map": [ + 0.01225995272397995, + 0.024165455251932144, + 0.02453707344830036, + 0.02301701530814171, + 0.029016969725489616, + 0.017279021441936493, + 0.024297084659337997, + 0.024903951212763786, + 0.00830754078924656, + 0.012626790441572666, + 0.0202649999409914, + 0.005762409884482622, + 0.019789785146713257, + 0.00903624203056097, + 0.005908108316361904, + 0.003530126763507724, + 0.007955702021718025, + 0.011850059032440186, + 0.027879538014531136, + 0.006409427151083946, + 0.004828695673495531, + 0.004126554820686579, + 0.010206510312855244, + 0.009838081896305084, + 0.016936378553509712, + 0.009057623334228992, + 0.07568727433681488, + 0.013035211712121964, + 0.005690510850399733, + 0.009464125148952007, + 0.007796715013682842, + 0.05538904294371605, + 0.007423446048051119, + 0.0041027311235666275, + 0.007607008330523968, + 0.0064234607852995396, + 0.008344213478267193, + 0.018087703734636307, + 0.004429217427968979, + 0.007330236025154591, + 0.0069513143971562386, + 0.007950975559651852, + 0.008141497150063515, + 0.008266839198768139, + 0.011249004863202572, + 0.007651007734239101, + 0.01404796727001667, + 0.012325252406299114, + 0.01121758297085762, + 0.013980189338326454, + 0.009566323831677437, + 0.009798411279916763, + 0.01305246539413929, + 0.010558908805251122, + 0.013408584520220757, + 0.009465964511036873, + 0.010551776736974716, + 0.011154038831591606, + 0.016089756041765213, + 0.0070502059534192085, + 0.008879026398062706, + 0.00918577890843153, + 0.007755993865430355, + 0.00867012981325388, + 0.019251780584454536, + 0.00557037303224206, + 0.042893506586551666, + 0.006982292514294386, + 0.004851085599511862, + 0.006630787160247564, + 0.02430647425353527, + 0.008643200621008873, + 0.00805238913744688, + 0.002894691424444318, + 0.00224109273403883, + 0.0022337273694574833, + 0.0029239756986498833, + 0.002927827648818493, + 0.0050302790477871895, + 0.0029955273494124413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.026331216096878052 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In Makoto Shinkai's distinctive artistic style, a steady and deliberate push-in slowly pans across a narrow, rain-slicked concrete balcony ledge, perched high above a vast, futuristic cityscape. Below, a dizzying chasm of towering skyscrapers crisscrosses, lined with glowing skytrain tracks, their glass and steel facades drenched in the ethereal glow of giant, flickering holographic advertisements. The scene is dominated by a cool, nocturnal palette of deep blues and purples, punctuated by vibrant neon pinks and cyans that reflect with surreal clarity off every wet surface, casting long, dramatic shadows that dance with the shifting light. On the ledge, two cats are the focus of an intimate courtship ritual. One of the male cats, with sleek black fur that gleams in the multicolored light, cautiously extends its paw and delicately pushes forward a brightly colored fish-shaped kitten as an offering. Its posture is low, its gaze unwavering, sincere and hopeful. The recipient, a calico cat with soft, fluffy white and ginger fur, hesitates, its large, inquisitive green eyes wide, leaning forward slightly, its whiskers twitching in the misty air. The entire scene is imbued with a quiet, romantic melancholy. The gentle, small-scale interaction between the two animals stands in stark contrast to the overwhelming scale and impersonal beauty of this vibrant metropolis, capturing fleeting connections in a vast and indifferent world.", + "seed": 17534, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_8/index_4689_seed_17534_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026500347722321748, + "vlm_logprob": -0.026500347722321748, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.026500347722321748, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023470675572752953 + ], + "opposite_logprob_tokens": [ + -4.023470878601074 + ], + "logprob_reward": -0.023470675572752953, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.023470675572752953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1470160186290741, + 0.5718557834625244, + 0.12481508404016495, + 0.030739180743694305, + 0.026391535997390747, + 0.023282062262296677, + 0.02965591475367546, + 0.015459581278264523, + 0.022667886689305305, + 0.008116941899061203 + ], + "credit_map": [ + 0.017937015742063522, + 0.014538363553583622, + 0.02895178645849228, + 0.018459027633070946, + 0.014160940423607826, + 0.02046283148229122, + 0.015519199892878532, + 0.016986863687634468, + 0.20559541881084442, + 0.023512830957770348, + 0.03595025837421417, + 0.011413795873522758, + 0.026749350130558014, + 0.22926917672157288, + 0.014893563464283943, + 0.02447139285504818, + 0.03923957422375679, + 0.007810395210981369, + 0.00814957357943058, + 0.005290987901389599, + 0.006807644385844469, + 0.042972274124622345, + 0.004122936632484198, + 0.010421701706945896, + 0.0038036296609789133, + 0.0034909930545836687, + 0.0023962424602359533, + 0.002492584753781557, + 0.004926169756799936, + 0.005096151493489742, + 0.0032689357176423073, + 0.00526447594165802, + 0.004457708448171616, + 0.002330712042748928, + 0.003149863798171282, + 0.0023602773435413837, + 0.004126625135540962, + 0.0030566889327019453, + 0.003413833910599351, + 0.003495829412713647, + 0.0037467030342668295, + 0.0022148103453218937, + 0.0030016659293323755, + 0.0022290940396487713, + 0.003330367850139737, + 0.0018950874218717217, + 0.003371858038008213, + 0.003492477349936962, + 0.002982332604005933, + 0.003408965887501836, + 0.003867997555062175, + 0.0026072082109749317, + 0.003918670117855072, + 0.004682198632508516, + 0.003834552364423871, + 0.004353991709649563, + 0.002408072119578719, + 0.0020631595980376005, + 0.0020297097507864237, + 0.001737098558805883, + 0.0019079878693446517, + 0.0011752292048186064, + 0.001877462025731802, + 0.0022608644794672728, + 0.002818079199641943, + 0.0024799080565571785, + 0.0029812329448759556, + 0.0021342337131500244, + 0.0029919908847659826, + 0.003152023535221815, + 0.003063526004552841, + 0.0030468921177089214, + 0.000880061648786068, + 0.0012971949763596058, + 0.0010419365717098117, + 0.0008848379948176444, + 0.0009703375399112701, + 0.0007438866887241602, + 0.0011425487464293838, + 0.0011561382561922073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.023470675572752953 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025012181140482426 + ], + "opposite_logprob_tokens": [ + -6.002501010894775 + ], + "logprob_reward": -0.0025012181140482426, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025012181140482426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.195349782705307, + 0.5705030560493469, + 0.13442575931549072, + 0.02260485477745533, + 0.017948726192116737, + 0.012454904615879059, + 0.022620465606451035, + 0.011340076103806496, + 0.008405483327805996, + 0.004346830304712057 + ], + "credit_map": [ + 0.02979218401014805, + 0.018178479745984077, + 0.03436673432588577, + 0.029276710003614426, + 0.023232022300362587, + 0.01821446418762207, + 0.022400539368391037, + 0.01988864503800869, + 0.20011048018932343, + 0.023366058245301247, + 0.03624774515628815, + 0.014176408760249615, + 0.02754550613462925, + 0.22763806581497192, + 0.013569213449954987, + 0.02784956805408001, + 0.0456978864967823, + 0.007191913202404976, + 0.007171293254941702, + 0.004813182167708874, + 0.006984529085457325, + 0.04745537042617798, + 0.004265978466719389, + 0.010845608077943325, + 0.0031026843935251236, + 0.0027777301147580147, + 0.0017432189779356122, + 0.0020751310512423515, + 0.0044066510163247585, + 0.003016664646565914, + 0.0026240695733577013, + 0.0028587058186531067, + 0.0019182272953912616, + 0.0024326734710484743, + 0.0022595892660319805, + 0.0019569811411201954, + 0.0023708625230938196, + 0.0012828928884118795, + 0.0032001479994505644, + 0.0025273526553064585, + 0.0015344637213274837, + 0.0020983112044632435, + 0.0016648370074108243, + 0.0013666949234902859, + 0.0013507226249203086, + 0.0010069099953398108, + 0.0019656699150800705, + 0.0014672961551696062, + 0.0021817723754793406, + 0.001751340925693512, + 0.0036347866989672184, + 0.00134033162612468, + 0.003572033019736409, + 0.005191355478018522, + 0.0016977194463834167, + 0.003251125803217292, + 0.0010180037934333086, + 0.0006981146871112287, + 0.001938006840646267, + 0.0007192700868472457, + 0.0019013195997104049, + 0.002553474623709917, + 0.0010638253297656775, + 0.0014480610843747854, + 0.0009692424791865051, + 0.000704583537299186, + 0.00137084792368114, + 0.0005095782689750195, + 0.0011597261764109135, + 0.0012534043053165078, + 0.0009056046255864203, + 0.0015324958367273211, + 0.0005273845163173974, + 0.00040697865188121796, + 0.0005602166056632996, + 0.00033680940396152437, + 0.0008050167816691101, + 0.0005971090868115425, + 0.0004704746534116566, + 0.0006428405758924782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0025012181140482426 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305095225572586 + ], + "opposite_logprob_tokens": [ + -3.16805100440979 + ], + "logprob_reward": -0.04305095225572586, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305095225572586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07796134799718857, + 0.6756590008735657, + 0.12172936648130417, + 0.018386175855994225, + 0.034805990755558014, + 0.0168850589543581, + 0.01530479360371828, + 0.020209839567542076, + 0.01379089429974556, + 0.005267513450235128 + ], + "credit_map": [ + 0.007561229635030031, + 0.008061669766902924, + 0.01917148381471634, + 0.010694711469113827, + 0.007926290854811668, + 0.007827906869351864, + 0.007856708951294422, + 0.008861342445015907, + 0.24187041819095612, + 0.02811571955680847, + 0.0421723797917366, + 0.014295695349574089, + 0.032524898648262024, + 0.2738526463508606, + 0.015190827660262585, + 0.02763643115758896, + 0.0375496931374073, + 0.007151093799620867, + 0.00980317872017622, + 0.005587883293628693, + 0.0075586941093206406, + 0.03772091865539551, + 0.006715407129377127, + 0.009642506018280983, + 0.0028986127581447363, + 0.002302082721143961, + 0.0013182430993765593, + 0.0018631520215421915, + 0.002709335647523403, + 0.0026216083206236362, + 0.0019400581950321794, + 0.0027330839075148106, + 0.004452708177268505, + 0.005149850156158209, + 0.0038150562904775143, + 0.004127616994082928, + 0.004506835713982582, + 0.002398153766989708, + 0.005352304317057133, + 0.005003464408218861, + 0.001754725817590952, + 0.0020892261527478695, + 0.0026815568562597036, + 0.0014584636082872748, + 0.0024462277069687843, + 0.0030605143401771784, + 0.0013684958685189486, + 0.0020258482545614243, + 0.0016314440872520208, + 0.0014213750837370753, + 0.002430505584925413, + 0.0010861180489882827, + 0.002195267006754875, + 0.0033889878541231155, + 0.0012886242475360632, + 0.0018624706426635385, + 0.0021545717027038336, + 0.0010820686584338546, + 0.0036196685396134853, + 0.0009548339294269681, + 0.0027809387538582087, + 0.005389241501688957, + 0.0014642846072092652, + 0.0027642319910228252, + 0.0014839529758319259, + 0.001541912672109902, + 0.0022984875831753016, + 0.0009044394828379154, + 0.0018388533499091864, + 0.002697291551157832, + 0.001227634958922863, + 0.0017983210273087025, + 0.0006644725217483938, + 0.0005366491968743503, + 0.0008067962480708957, + 0.00040827016346156597, + 0.0009402325958944857, + 0.00045412787585519254, + 0.0005711282137781382, + 0.0008858363144099712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.04305095225572586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795775189995766 + ], + "opposite_logprob_tokens": [ + -3.5297958850860596 + ], + "logprob_reward": -0.029795775189995766, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795775189995766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24580837786197662, + 0.6722947955131531, + 0.031089216470718384, + 0.009598884731531143, + 0.01159012783318758, + 0.006880247034132481, + 0.008611678145825863, + 0.0049654156900942326, + 0.004996922332793474, + 0.004164252430200577 + ], + "credit_map": [ + 0.03339560702443123, + 0.02283581905066967, + 0.043462082743644714, + 0.03656798228621483, + 0.024264544248580933, + 0.033113911747932434, + 0.027930626645684242, + 0.024237792938947678, + 0.24130336940288544, + 0.027460549026727676, + 0.042276531457901, + 0.014981132932007313, + 0.031591784209012985, + 0.26982414722442627, + 0.01571892574429512, + 0.02913838066160679, + 0.009257186204195023, + 0.0018365606665611267, + 0.0019315365934744477, + 0.0014265241334214807, + 0.0019148867577314377, + 0.010462458245456219, + 0.00162171246483922, + 0.002638352569192648, + 0.0012333500199019909, + 0.0013091835426166654, + 0.0010625682771205902, + 0.0009270281880162656, + 0.0011764586670324206, + 0.0009220931679010391, + 0.0011524753645062447, + 0.001815726631321013, + 0.0017594749806448817, + 0.0014347077812999487, + 0.0017208330100402236, + 0.0011835925979539752, + 0.0013967042323201895, + 0.0008765463717281818, + 0.0013873822754248977, + 0.00183088646735996, + 0.000865855545271188, + 0.0006351475603878498, + 0.001046791672706604, + 0.0005031166365370154, + 0.0010398700833320618, + 0.0012074990663677454, + 0.0005656627472490072, + 0.0010163037804886699, + 0.0008918509702198207, + 0.0008839451475068927, + 0.0013888784451410174, + 0.0006203479715622962, + 0.0012728349538519979, + 0.0019045459339395165, + 0.0006423486047424376, + 0.0010069257114082575, + 0.0007690180791541934, + 0.0005221458850428462, + 0.0006290496094152331, + 0.0005149790667928755, + 0.000672914320603013, + 0.00045332557056099176, + 0.00065256078960374, + 0.0007514223689213395, + 0.0008073116769082844, + 0.0006486347992904484, + 0.0006204331875778735, + 0.0006331239128485322, + 0.0006218876224011183, + 0.00036879690014757216, + 0.0006782172713428736, + 0.0006185164675116539, + 0.0005365582183003426, + 0.0005688570672646165, + 0.0005718178581446409, + 0.0005074283690191805, + 0.0004882175999227911, + 0.00036211282713338733, + 0.0006118813762441278, + 0.0005173790850676596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.029795775189995766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368311747908592 + ], + "opposite_logprob_tokens": [ + -3.4086830615997314 + ], + "logprob_reward": -0.03368311747908592, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368311747908592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08442094177007675, + 0.38204413652420044, + 0.07497502118349075, + 0.15466606616973877, + 0.0756184384226799, + 0.10586665570735931, + 0.04173465073108673, + 0.03817383944988251, + 0.026692252606153488, + 0.01580801047384739 + ], + "credit_map": [ + 0.009361566975712776, + 0.008801349438726902, + 0.009658100083470345, + 0.014258195646107197, + 0.007947309873998165, + 0.008719217032194138, + 0.0141696622595191, + 0.01150553859770298, + 0.10319946706295013, + 0.0347917415201664, + 0.030137553811073303, + 0.015778642147779465, + 0.029753830283880234, + 0.12378765642642975, + 0.01828741282224655, + 0.026307838037610054, + 0.010811623185873032, + 0.009974690154194832, + 0.012326288037002087, + 0.00786618236452341, + 0.007871769368648529, + 0.00799760315567255, + 0.01005085650831461, + 0.008076008409261703, + 0.018980134278535843, + 0.013926807790994644, + 0.01439046673476696, + 0.011787408031523228, + 0.02650955319404602, + 0.021380674093961716, + 0.02423482947051525, + 0.023456189781427383, + 0.010534211993217468, + 0.007161235436797142, + 0.01092239934951067, + 0.005974162369966507, + 0.011540443636476994, + 0.00651048868894577, + 0.009790359064936638, + 0.013185136951506138, + 0.012600150890648365, + 0.01860424317419529, + 0.012123028747737408, + 0.012217085808515549, + 0.012538366951048374, + 0.009070265106856823, + 0.0149811627343297, + 0.013732356950640678, + 0.004533947911113501, + 0.004151947796344757, + 0.005793582182377577, + 0.0035531893372535706, + 0.0041579618118703365, + 0.003929274622350931, + 0.006948687136173248, + 0.008666060864925385, + 0.005236375145614147, + 0.0031683677807450294, + 0.00518004409968853, + 0.0030668105464428663, + 0.005530981812626123, + 0.007812738418579102, + 0.003043886972591281, + 0.005134637001901865, + 0.005095432512462139, + 0.002796859247609973, + 0.0033479882404208183, + 0.0021503837779164314, + 0.004161688964813948, + 0.0030526211485266685, + 0.0027295106556266546, + 0.003357769688591361, + 0.0018525536870583892, + 0.0027448872569948435, + 0.0019184232223778963, + 0.0016511011635884643, + 0.002452897373586893, + 0.001522221602499485, + 0.001742002204991877, + 0.0019239240791648626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.03368311747908592 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In Makoto Shinkai's distinctive artistic style, a steady and deliberate push-in slowly pans across a narrow, rain-slicked concrete balcony ledge, perched high above a vast, futuristic cityscape. Below, a dizzying chasm of towering skyscrapers crisscrosses, lined with glowing skytrain tracks, their glass and steel facades drenched in the ethereal glow of giant, flickering holographic advertisements. The scene is dominated by a cool, nocturnal palette of deep blues and purples, punctuated by vibrant neon pinks and cyans that reflect with surreal clarity off every wet surface, casting long, dramatic shadows that dance with the shifting light. On the ledge, two cats are the focus of an intimate courtship ritual. One of the male cats, with sleek black fur that gleams in the multicolored light, cautiously extends its paw and delicately pushes forward a brightly colored fish-shaped kitten as an offering. Its posture is low, its gaze unwavering, sincere and hopeful. The recipient, a calico cat with soft, fluffy white and ginger fur, hesitates, its large, inquisitive green eyes wide, leaning forward slightly, its whiskers twitching in the misty air. The entire scene is imbued with a quiet, romantic melancholy. The gentle, small-scale interaction between the two animals stands in stark contrast to the overwhelming scale and impersonal beauty of this vibrant metropolis, capturing fleeting connections in a vast and indifferent world.", + "seed": 17535, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_8/index_4689_seed_17535_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0175084849819541, + "vlm_logprob": -0.0175084849819541, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.0175084849819541, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015202251262962818 + ], + "opposite_logprob_tokens": [ + -4.515202045440674 + ], + "logprob_reward": -0.015202251262962818, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.015202251262962818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5091798901557922, + 0.1466149538755417, + 0.053979698568582535, + 0.04827791452407837, + 0.0744616910815239, + 0.06066512688994408, + 0.030146131291985512, + 0.026918886229395866, + 0.017621269449591637, + 0.03213438764214516 + ], + "credit_map": [ + 0.04002518206834793, + 0.07559001445770264, + 0.04598856344819069, + 0.11576443165540695, + 0.03879096359014511, + 0.07585291564464569, + 0.06583109498023987, + 0.051336754113435745, + 0.012743883766233921, + 0.02239643782377243, + 0.018152838572859764, + 0.01839548721909523, + 0.016994984820485115, + 0.02306605689227581, + 0.01343101542443037, + 0.021434254944324493, + 0.0051082768477499485, + 0.010414140298962593, + 0.004512058105319738, + 0.0064646112732589245, + 0.006470576860010624, + 0.006900519132614136, + 0.0063024223782122135, + 0.007807092275470495, + 0.0030787175055593252, + 0.004215238615870476, + 0.0054985880851745605, + 0.010797400958836079, + 0.004652600269764662, + 0.004930559080094099, + 0.01017870381474495, + 0.004926104564219713, + 0.011815422214567661, + 0.00568159157410264, + 0.007851321250200272, + 0.008630681782960892, + 0.005361495539546013, + 0.008713043294847012, + 0.012584357522428036, + 0.013823777437210083, + 0.009507919661700726, + 0.006215325091034174, + 0.009030676446855068, + 0.00989443063735962, + 0.004856540355831385, + 0.007254239171743393, + 0.006018754560500383, + 0.007887246087193489, + 0.004639625549316406, + 0.0024504412431269884, + 0.003933687694370747, + 0.004549774341285229, + 0.0028319486882537603, + 0.0036524974275380373, + 0.00403542909771204, + 0.00405272888019681, + 0.003912711516022682, + 0.0018661362119019032, + 0.0031932396814227104, + 0.004428419284522533, + 0.0032530194148421288, + 0.0022373273968696594, + 0.003909547347575426, + 0.00411848584190011, + 0.0026576058007776737, + 0.0016732559306547046, + 0.002206125296652317, + 0.0032248892821371555, + 0.001357616507448256, + 0.001457446371205151, + 0.0025666861329227686, + 0.002477644244208932, + 0.006461926270276308, + 0.005082126706838608, + 0.004075017292052507, + 0.0046289474703371525, + 0.0031648152507841587, + 0.0026928840670734644, + 0.0025201470125466585, + 0.003508525900542736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.015202251262962818 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747319363057613 + ], + "opposite_logprob_tokens": [ + -5.006747245788574 + ], + "logprob_reward": -0.006747319363057613, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747319363057613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3288987874984741, + 0.13118131458759308, + 0.06606573611497879, + 0.06873802095651627, + 0.07340963184833527, + 0.06750296801328659, + 0.061964768916368484, + 0.04156019911170006, + 0.031787317246198654, + 0.12889128923416138 + ], + "credit_map": [ + 0.02305285446345806, + 0.0525476336479187, + 0.033979590982198715, + 0.06557647138834, + 0.029125677421689034, + 0.050963498651981354, + 0.03720805048942566, + 0.036445021629333496, + 0.01093665137887001, + 0.02069038711488247, + 0.01700310967862606, + 0.01613650470972061, + 0.015701593831181526, + 0.0208115391433239, + 0.012327597476541996, + 0.017573928460478783, + 0.005494585260748863, + 0.008444400504231453, + 0.0071404362097382545, + 0.01153390109539032, + 0.010023270733654499, + 0.010057919658720493, + 0.006652032025158405, + 0.006719185970723629, + 0.004819353111088276, + 0.006124468520283699, + 0.009322398342192173, + 0.01161970291286707, + 0.006140802521258593, + 0.007422701455652714, + 0.012098602950572968, + 0.011189990676939487, + 0.010130098089575768, + 0.004308163654059172, + 0.009143470786511898, + 0.012677228078246117, + 0.00459172623232007, + 0.006405096035450697, + 0.012855613604187965, + 0.01329822838306427, + 0.013169137760996819, + 0.006103274412453175, + 0.008790542371571064, + 0.012086171656847, + 0.007606619503349066, + 0.006057366728782654, + 0.005773446056991816, + 0.007916408590972424, + 0.01054746937006712, + 0.006034610792994499, + 0.007920436561107635, + 0.009914141148328781, + 0.007009692955762148, + 0.007460321765393019, + 0.006637968122959137, + 0.006440130993723869, + 0.006956956814974546, + 0.004060661885887384, + 0.005368040408939123, + 0.006711293943226337, + 0.0033054782543331385, + 0.005114574916660786, + 0.004933371674269438, + 0.005109825171530247, + 0.005377196241170168, + 0.002384328516200185, + 0.004868820309638977, + 0.004295162856578827, + 0.003609591396525502, + 0.003407543757930398, + 0.0032747399527579546, + 0.004569930490106344, + 0.020732494071125984, + 0.013887139037251472, + 0.01660306379199028, + 0.018976738676428795, + 0.010805985890328884, + 0.018418146297335625, + 0.014912065118551254, + 0.014555663801729679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.006747319363057613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793230816721916 + ], + "opposite_logprob_tokens": [ + -3.5297932624816895 + ], + "logprob_reward": -0.029793230816721916, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793230816721916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3598271906375885, + 0.1804356724023819, + 0.08891992270946503, + 0.034527990967035294, + 0.06210958957672119, + 0.06431429833173752, + 0.06667302548885345, + 0.05477314069867134, + 0.04222241789102554, + 0.046196818351745605 + ], + "credit_map": [ + 0.028713850304484367, + 0.04540159925818443, + 0.037801407277584076, + 0.08716122806072235, + 0.025293370708823204, + 0.0468427911400795, + 0.044736601412296295, + 0.04387631639838219, + 0.01766233704984188, + 0.025566142052412033, + 0.027023492380976677, + 0.023089690133929253, + 0.02326963096857071, + 0.027360927313566208, + 0.016319816932082176, + 0.020143622532486916, + 0.008578993380069733, + 0.013909352011978626, + 0.006230804603546858, + 0.01008703000843525, + 0.019513040781021118, + 0.012397602200508118, + 0.008069700561463833, + 0.01013338565826416, + 0.004191949963569641, + 0.003240681253373623, + 0.0036237728781998158, + 0.0038662715815007687, + 0.004653058014810085, + 0.004556860774755478, + 0.005738324020057917, + 0.004657069221138954, + 0.008516652509570122, + 0.005927309859544039, + 0.007197028491646051, + 0.010944747366011143, + 0.006798909977078438, + 0.0073328460566699505, + 0.0068979631178081036, + 0.008494120091199875, + 0.00981292873620987, + 0.007893364876508713, + 0.009290036745369434, + 0.00857163779437542, + 0.005888308864086866, + 0.008406630717217922, + 0.0053779082372784615, + 0.009073476307094097, + 0.013249020092189312, + 0.007525128778070211, + 0.011046496219933033, + 0.009108143858611584, + 0.005896369926631451, + 0.005648770369589329, + 0.008242577314376831, + 0.00595651101320982, + 0.010320831090211868, + 0.005071111489087343, + 0.009273918345570564, + 0.00609499542042613, + 0.006663833279162645, + 0.006834504194557667, + 0.006077793892472982, + 0.004436145070940256, + 0.007376876659691334, + 0.0036718654446303844, + 0.007985708303749561, + 0.005712318234145641, + 0.004329140763729811, + 0.004283818881958723, + 0.004630178678780794, + 0.004232504870742559, + 0.008508894592523575, + 0.003701607696712017, + 0.00865370687097311, + 0.006150201894342899, + 0.0039429073221981525, + 0.004869596567004919, + 0.00495458347722888, + 0.005415312014520168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.029793230816721916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531115673482418 + ], + "opposite_logprob_tokens": [ + -4.387531280517578 + ], + "logprob_reward": -0.012531115673482418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531115673482418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.457120418548584, + 0.12208676338195801, + 0.06596609950065613, + 0.06419793516397476, + 0.09007691591978073, + 0.03867556154727936, + 0.05904431268572807, + 0.04453674331307411, + 0.026503032073378563, + 0.03179217875003815 + ], + "credit_map": [ + 0.0297319944947958, + 0.04708879813551903, + 0.04740777984261513, + 0.09119594097137451, + 0.04655931517481804, + 0.07594681531190872, + 0.06001646816730499, + 0.0591733381152153, + 0.007439717650413513, + 0.010716310702264309, + 0.00944418366998434, + 0.025976134464144707, + 0.014535202644765377, + 0.01348115038126707, + 0.017662666738033295, + 0.022831393405795097, + 0.00685246242210269, + 0.0059377518482506275, + 0.005943055264651775, + 0.010111700743436813, + 0.00912490300834179, + 0.013014172203838825, + 0.007189153227955103, + 0.007792903576046228, + 0.00518388906493783, + 0.009429408237338066, + 0.007441055029630661, + 0.010902666486799717, + 0.007108580786734819, + 0.006579509470611811, + 0.010761599987745285, + 0.00679122656583786, + 0.010193580761551857, + 0.006952945608645678, + 0.013976527377963066, + 0.01467525027692318, + 0.006933559197932482, + 0.010245857760310173, + 0.014107071794569492, + 0.012992117553949356, + 0.006459034513682127, + 0.0030342729296535254, + 0.00443615997210145, + 0.00552669120952487, + 0.005300342570990324, + 0.004345681052654982, + 0.004759975243359804, + 0.004813402891159058, + 0.008870622143149376, + 0.006835614796727896, + 0.008090686053037643, + 0.0075088185258209705, + 0.008255187422037125, + 0.0070508923381567, + 0.006517158355563879, + 0.0059153358452022076, + 0.008402973413467407, + 0.003756658872589469, + 0.005862962454557419, + 0.0063446699641644955, + 0.004872904624789953, + 0.004860883578658104, + 0.005493857432156801, + 0.004941832739859819, + 0.0044020419009029865, + 0.0029477104544639587, + 0.0042999256402254105, + 0.0029722987674176693, + 0.0024662460200488567, + 0.0033983865287154913, + 0.0033943254966288805, + 0.0026220963336527348, + 0.005072910338640213, + 0.0035625959280878305, + 0.004908527713268995, + 0.004166305065155029, + 0.002626312430948019, + 0.003678266890347004, + 0.0035263639874756336, + 0.004250895697623491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.012531115673482418 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023268507793545723 + ], + "opposite_logprob_tokens": [ + -3.773268461227417 + ], + "logprob_reward": -0.023268507793545723, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023268507793545723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31001508235931396, + 0.10027927160263062, + 0.06880250573158264, + 0.09008975327014923, + 0.10106635838747025, + 0.14405691623687744, + 0.12273623049259186, + 0.021012958139181137, + 0.016951654106378555, + 0.024989314377307892 + ], + "credit_map": [ + 0.021052969619631767, + 0.04304283857345581, + 0.036029208451509476, + 0.0692245364189148, + 0.02509351447224617, + 0.04238950461149216, + 0.03842658922076225, + 0.034755878150463104, + 0.007871664129197598, + 0.014299106784164906, + 0.012107162736356258, + 0.018571728840470314, + 0.009595693089067936, + 0.010984471067786217, + 0.014080505818128586, + 0.012768927030265331, + 0.006891225464642048, + 0.007957303896546364, + 0.007402634713798761, + 0.01235195156186819, + 0.007093468215316534, + 0.008327478542923927, + 0.009981674142181873, + 0.00879676453769207, + 0.007042908575385809, + 0.008345495909452438, + 0.011376697570085526, + 0.01693866401910782, + 0.008892272599041462, + 0.009270716458559036, + 0.015722259879112244, + 0.012500726617872715, + 0.014095425605773926, + 0.010894721373915672, + 0.011129464022815228, + 0.014411968179047108, + 0.009532710537314415, + 0.01090913638472557, + 0.01623552292585373, + 0.013857401907444, + 0.025618458166718483, + 0.012813838198781013, + 0.014029479585587978, + 0.018021974712610245, + 0.023953158408403397, + 0.013780406676232815, + 0.01364783477038145, + 0.02219175361096859, + 0.025726936757564545, + 0.00998054537922144, + 0.018638188019394875, + 0.016509488224983215, + 0.009529784321784973, + 0.015141003765165806, + 0.015916381031274796, + 0.011293881572782993, + 0.0034510225523263216, + 0.0018432089127600193, + 0.0019609166774898767, + 0.002976302755996585, + 0.0029855873435735703, + 0.002202741103246808, + 0.003007087390869856, + 0.0025860900059342384, + 0.0026398899499326944, + 0.0016510229324921966, + 0.0026321352925151587, + 0.0019049402326345444, + 0.0016181673854589462, + 0.0019042001804336905, + 0.0024067668709903955, + 0.002194529864937067, + 0.004869833122938871, + 0.0023529776372015476, + 0.0032971154432743788, + 0.003123438684269786, + 0.0023050853051245213, + 0.002764573786407709, + 0.0038896799087524414, + 0.0023866083938628435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.023268507793545723 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In Makoto Shinkai's distinctive artistic style, a steady and deliberate push-in slowly pans across a narrow, rain-slicked concrete balcony ledge, perched high above a vast, futuristic cityscape. Below, a dizzying chasm of towering skyscrapers crisscrosses, lined with glowing skytrain tracks, their glass and steel facades drenched in the ethereal glow of giant, flickering holographic advertisements. The scene is dominated by a cool, nocturnal palette of deep blues and purples, punctuated by vibrant neon pinks and cyans that reflect with surreal clarity off every wet surface, casting long, dramatic shadows that dance with the shifting light. On the ledge, two cats are the focus of an intimate courtship ritual. One of the male cats, with sleek black fur that gleams in the multicolored light, cautiously extends its paw and delicately pushes forward a brightly colored fish-shaped kitten as an offering. Its posture is low, its gaze unwavering, sincere and hopeful. The recipient, a calico cat with soft, fluffy white and ginger fur, hesitates, its large, inquisitive green eyes wide, leaning forward slightly, its whiskers twitching in the misty air. The entire scene is imbued with a quiet, romantic melancholy. The gentle, small-scale interaction between the two animals stands in stark contrast to the overwhelming scale and impersonal beauty of this vibrant metropolis, capturing fleeting connections in a vast and indifferent world.", + "seed": 17536, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_8/index_4689_seed_17536_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020214498043060303, + "vlm_logprob": -0.020214498043060303, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.020214498043060303, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030524134635925293 + ], + "opposite_logprob_tokens": [ + -3.655524253845215 + ], + "logprob_reward": -0.030524134635925293, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.030524134635925293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1432584971189499, + 0.09928341954946518, + 0.0959184318780899, + 0.10275567322969437, + 0.13660944998264313, + 0.23087675869464874, + 0.08578365296125412, + 0.04215236380696297, + 0.044338569045066833, + 0.01902320608496666 + ], + "credit_map": [ + 0.018302910029888153, + 0.00871048029512167, + 0.016248952597379684, + 0.028284184634685516, + 0.010744242928922176, + 0.021793579682707787, + 0.016638299450278282, + 0.02253582887351513, + 0.004098770674318075, + 0.004253406077623367, + 0.006685325875878334, + 0.00661048898473382, + 0.006943585351109505, + 0.03476940095424652, + 0.008113798685371876, + 0.027808627113699913, + 0.00556652108207345, + 0.007046960759907961, + 0.015745319426059723, + 0.011335576884448528, + 0.01151952613145113, + 0.012906229123473167, + 0.016640016809105873, + 0.015158263966441154, + 0.0067657888866961, + 0.010919871740043163, + 0.012403709813952446, + 0.008410687558352947, + 0.01265876367688179, + 0.014899290166795254, + 0.019145958125591278, + 0.017551597207784653, + 0.012097734026610851, + 0.013634949922561646, + 0.014827459119260311, + 0.011621860787272453, + 0.01576106995344162, + 0.019268067553639412, + 0.03717834874987602, + 0.012219950556755066, + 0.030962783843278885, + 0.02529226616024971, + 0.03971702605485916, + 0.021941183134913445, + 0.03262485936284065, + 0.031476426869630814, + 0.024665694683790207, + 0.02419649064540863, + 0.011471454054117203, + 0.008860379457473755, + 0.008987917564809322, + 0.00841746386140585, + 0.012025143951177597, + 0.010730286128818989, + 0.014624183066189289, + 0.010666820220649242, + 0.004689646884799004, + 0.006455047056078911, + 0.005530155263841152, + 0.004400897305458784, + 0.004297946579754353, + 0.00512586859986186, + 0.005471911747008562, + 0.006180889438837767, + 0.006489865947514772, + 0.006641586776822805, + 0.006579251494258642, + 0.004373204428702593, + 0.004646883811801672, + 0.005237812176346779, + 0.0043492792174220085, + 0.0060206810012459755, + 0.0027299108915030956, + 0.0029717145953327417, + 0.0024856706149876118, + 0.0023807906545698643, + 0.002169420011341572, + 0.0028480698820203543, + 0.0016965745016932487, + 0.0017410522559657693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.030524134635925293 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002212459221482277 + ], + "opposite_logprob_tokens": [ + -6.1272125244140625 + ], + "logprob_reward": -0.002212459221482277, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002212459221482277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060882434248924255, + 0.23519644141197205, + 0.04082394018769264, + 0.0876995176076889, + 0.24957287311553955, + 0.08174637705087662, + 0.13116438686847687, + 0.03136659413576126, + 0.022702019661664963, + 0.05884547159075737 + ], + "credit_map": [ + 0.004723267629742622, + 0.005123150534927845, + 0.005061763804405928, + 0.01652780920267105, + 0.004967795684933662, + 0.008975791744887829, + 0.004957027267664671, + 0.01054582092911005, + 0.006947401445358992, + 0.0051987990736961365, + 0.011560972779989243, + 0.010283343493938446, + 0.008287892676889896, + 0.1026151105761528, + 0.008243725635111332, + 0.08205919712781906, + 0.0030014829244464636, + 0.0035337351728230715, + 0.0049954913556575775, + 0.006586454808712006, + 0.004934250842779875, + 0.006512341555207968, + 0.006709984969347715, + 0.004550197161734104, + 0.006776619702577591, + 0.011770740151405334, + 0.013647994957864285, + 0.009436030872166157, + 0.011482059955596924, + 0.01088931318372488, + 0.012231479398906231, + 0.011465269140899181, + 0.017808228731155396, + 0.0208890438079834, + 0.01794547401368618, + 0.017462078481912613, + 0.02626175805926323, + 0.04092056304216385, + 0.08067154884338379, + 0.027614159509539604, + 0.010475413873791695, + 0.00761602446436882, + 0.012084919959306717, + 0.006874740589410067, + 0.012928235344588757, + 0.010671577416360378, + 0.011274440214037895, + 0.009821022860705853, + 0.013488265685737133, + 0.011578280478715897, + 0.014084558002650738, + 0.011001556180417538, + 0.02101389691233635, + 0.017744123935699463, + 0.022772127762436867, + 0.019481578841805458, + 0.004313406068831682, + 0.005119646899402142, + 0.004343630746006966, + 0.003543390892446041, + 0.0035179825499653816, + 0.0038031802978366613, + 0.0032874697353690863, + 0.0034378832206130028, + 0.0025476585142314434, + 0.003997108433395624, + 0.0026425085961818695, + 0.0017536625964567065, + 0.0036948530469089746, + 0.002680085599422455, + 0.0023594978265464306, + 0.0030266435351222754, + 0.007291589863598347, + 0.0058319782838225365, + 0.006629674229770899, + 0.005763842724263668, + 0.008266046643257141, + 0.008118833415210247, + 0.006571474485099316, + 0.010372023098170757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.002212459221482277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065318137407303 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016065318137407303, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065318137407303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09703998267650604, + 0.166822150349617, + 0.12192729860544205, + 0.12285197526216507, + 0.21467135846614838, + 0.11445091664791107, + 0.0432799756526947, + 0.038636595010757446, + 0.0433318130671978, + 0.036987945437431335 + ], + "credit_map": [ + 0.009193461388349533, + 0.00921400636434555, + 0.007408341858536005, + 0.023267928510904312, + 0.009634085930883884, + 0.014968265779316425, + 0.009173291735351086, + 0.01418059878051281, + 0.006170264445245266, + 0.004689136520028114, + 0.009666848927736282, + 0.010169541463255882, + 0.00857180543243885, + 0.06273134797811508, + 0.011437581852078438, + 0.05338561162352562, + 0.008424008265137672, + 0.006755935959517956, + 0.01674273982644081, + 0.021019347012043, + 0.010797594673931599, + 0.017369020730257034, + 0.029009466990828514, + 0.01180918700993061, + 0.008631951175630093, + 0.01506167184561491, + 0.020050466060638428, + 0.010752297937870026, + 0.015767740085721016, + 0.01596038229763508, + 0.01945747621357441, + 0.017169984057545662, + 0.018867263570427895, + 0.03189721703529358, + 0.02663378231227398, + 0.02313271164894104, + 0.024867812171578407, + 0.03221534565091133, + 0.03786338120698929, + 0.01919385977089405, + 0.013862733729183674, + 0.011976503767073154, + 0.01781291700899601, + 0.010482017882168293, + 0.013732231222093105, + 0.01519038062542677, + 0.018384138122200966, + 0.013009995222091675, + 0.00523016694933176, + 0.005901670083403587, + 0.006741664372384548, + 0.006267676595598459, + 0.004373859614133835, + 0.004654698073863983, + 0.004466297570616007, + 0.005643943790346384, + 0.006508759222924709, + 0.005993532482534647, + 0.005365610122680664, + 0.004407878499478102, + 0.002832181053236127, + 0.004904726054519415, + 0.004630208481103182, + 0.003993702121078968, + 0.006376114673912525, + 0.004811793100088835, + 0.004113958682864904, + 0.00402311235666275, + 0.006678453646600246, + 0.005491090007126331, + 0.005529892630875111, + 0.006307393778115511, + 0.004748030565679073, + 0.004073454532772303, + 0.004229042213410139, + 0.00462453393265605, + 0.005166010931134224, + 0.004698572680354118, + 0.004010516218841076, + 0.005437786225229502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.016065318137407303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190951362252235 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014190951362252235, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190951362252235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.322862446308136, + 0.21876966953277588, + 0.10276786983013153, + 0.06816195696592331, + 0.08842013031244278, + 0.08349684625864029, + 0.05575009062886238, + 0.01759408600628376, + 0.018871428444981575, + 0.023305416107177734 + ], + "credit_map": [ + 0.03859107568860054, + 0.019331440329551697, + 0.029362984001636505, + 0.05964616313576698, + 0.029424989596009254, + 0.04009757563471794, + 0.03662501275539398, + 0.06978324055671692, + 0.009207958355545998, + 0.007058853283524513, + 0.017226578667759895, + 0.013134771026670933, + 0.013599387370049953, + 0.08035404235124588, + 0.01555868610739708, + 0.06262941658496857, + 0.006531055551022291, + 0.008510876446962357, + 0.011993848718702793, + 0.012436565943062305, + 0.016854997724294662, + 0.016571329906582832, + 0.016549278050661087, + 0.013319923542439938, + 0.00571421766653657, + 0.00620433920994401, + 0.007023113779723644, + 0.004895798861980438, + 0.01092576328665018, + 0.009345302358269691, + 0.012617554515600204, + 0.011435873806476593, + 0.008898724801838398, + 0.010978711768984795, + 0.009539618156850338, + 0.0076620737090706825, + 0.011756166815757751, + 0.013583535328507423, + 0.01808500476181507, + 0.00791630707681179, + 0.012400602921843529, + 0.008951046504080296, + 0.010211152024567127, + 0.0057050068862736225, + 0.014032218605279922, + 0.01109058316797018, + 0.012861181981861591, + 0.008245060220360756, + 0.0073490324430167675, + 0.0051898653618991375, + 0.005992685444653034, + 0.005180897191166878, + 0.0078521016985178, + 0.0071786572225391865, + 0.009557749144732952, + 0.0074491072446107864, + 0.0023034580517560244, + 0.002876607934013009, + 0.0025823800824582577, + 0.0017405744874849916, + 0.0017283206107094884, + 0.0019835548009723425, + 0.002226166892796755, + 0.0021530252415686846, + 0.0027215457521378994, + 0.002440356882289052, + 0.0027037737891077995, + 0.00178682676050812, + 0.002505476353690028, + 0.002372716087847948, + 0.0019817559514194727, + 0.002358978847041726, + 0.003829666879028082, + 0.003324512392282486, + 0.0032122458796948195, + 0.002059859223663807, + 0.0024667419493198395, + 0.0033839710522443056, + 0.0018246629042550921, + 0.0032037575729191303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014190951362252235 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two cats visible on a narrow concrete balcony ledge in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one cat black and the other a calico with white and ginger fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the black cat extending its paw to push a brightly colored fish-shaped object toward the calico cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a futuristic cityscape with towering skyscrapers, glowing skytrain tracks, and holographic advertisements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a steady and deliberate push-in motion across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038079626858234406 + ], + "opposite_logprob_tokens": [ + -3.2880797386169434 + ], + "logprob_reward": -0.038079626858234406, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038079626858234406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.172232985496521, + 0.425464391708374, + 0.05337821692228317, + 0.07462502270936966, + 0.0736270472407341, + 0.07687027007341385, + 0.08055178821086884, + 0.018385952338576317, + 0.011354941874742508, + 0.013509421609342098 + ], + "credit_map": [ + 0.01937037706375122, + 0.011738928966224194, + 0.016190925613045692, + 0.03156819939613342, + 0.017600692808628082, + 0.028933292254805565, + 0.013848000206053257, + 0.032982565462589264, + 0.012034169398248196, + 0.01588820479810238, + 0.049156446009874344, + 0.0455690398812294, + 0.019876502454280853, + 0.14214973151683807, + 0.027189644053578377, + 0.11360064893960953, + 0.00410475954413414, + 0.004815344233065844, + 0.009306964464485645, + 0.009909541346132755, + 0.005541034042835236, + 0.0070709954015910625, + 0.007532219868153334, + 0.005097357090562582, + 0.005017951596528292, + 0.008975806646049023, + 0.010527645237743855, + 0.005837981589138508, + 0.009938795119524002, + 0.010549113154411316, + 0.011413798667490482, + 0.012363928370177746, + 0.008914383128285408, + 0.008762459270656109, + 0.009298923425376415, + 0.006157252471894026, + 0.008127165026962757, + 0.010212746448814869, + 0.015219546854496002, + 0.006934572476893663, + 0.008139142766594887, + 0.007977322675287724, + 0.007634146139025688, + 0.0053519983775913715, + 0.014671881683170795, + 0.009855715557932854, + 0.013796626590192318, + 0.0094434330239892, + 0.010037887841463089, + 0.007035486865788698, + 0.010173307731747627, + 0.007682801689952612, + 0.010398021899163723, + 0.01044552307575941, + 0.012649686075747013, + 0.012129073962569237, + 0.001968152355402708, + 0.0033768760040402412, + 0.0027885492891073227, + 0.002007081639021635, + 0.0019505739910528064, + 0.0019883541390299797, + 0.0021812841296195984, + 0.002125081140547991, + 0.00193029863294214, + 0.0014253744157031178, + 0.0013124392135068774, + 0.0009434014791622758, + 0.0011622837046161294, + 0.0014754653675481677, + 0.0011882318649441004, + 0.0019174469634890556, + 0.0018179521430283785, + 0.0018569169333204627, + 0.0013569715665653348, + 0.0015135013964027166, + 0.0019589411094784737, + 0.0017775422893464565, + 0.001430156989954412, + 0.0017974392976611853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.038079626858234406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_9/index_tensor([2607, 2607])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_9/index_tensor([2607, 2607])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7d8cc7d4b9397fa6bd7ab176215231a5797072b2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_9/index_tensor([2607, 2607])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide, low-angle shot, the camera slowly panning to the right, sweeps across a poignant scene frozen in time. A thin, morning mist, like the sighs of fallen soldiers, slowly rises from the silent ruins of a battlefield, winding like a ghost across the muddy, charred earth. The skyline is pierced by the faint, almost grayish light of the rising sun. This cold light penetrates the mist, casting long, soft shadows on the damp metal and earth. Broken spears, notched sword blades, and shattered shields riddled with arrows lie scattered across the ground, their cold metal surfaces reflecting a faint light. Several torn battle flags, their faded crimson and gold patterns blurred, hang limply from broken poles, occasionally swaying gently in the barely perceptible breeze. The main colors of the entire scene are composed of earthy brown, charcoal black and cold metallic gray, creating a solemn, sad and tranquil atmosphere, full of a heavy sense of history.", + "seed": 13945, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_9/index_2607_seed_13945_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11017822865396738, + "vlm_logprob": -0.11017822865396738, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.11017822865396738, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3132384717464447 + ], + "opposite_logprob_tokens": [ + -1.563238501548767 + ], + "logprob_reward": -0.3132384717464447, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.3132384717464447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36752036213874817, + 0.08212263882160187, + 0.15921428799629211, + 0.10739786922931671, + 0.08193108439445496, + 0.0389765128493309, + 0.024759912863373756, + 0.051562730222940445, + 0.044514529407024384, + 0.04200005531311035 + ], + "credit_map": [ + 0.036942459642887115, + 0.05786033347249031, + 0.021949708461761475, + 0.028760042041540146, + 0.07880319654941559, + 0.08321389555931091, + 0.033468786627054214, + 0.02652193419635296, + 0.007443114649504423, + 0.008909606374800205, + 0.013774260878562927, + 0.022028395906090736, + 0.0065696099773049355, + 0.007516777142882347, + 0.009389265440404415, + 0.006491613574326038, + 0.02094779722392559, + 0.009116806089878082, + 0.014970619231462479, + 0.034845154732465744, + 0.04503033682703972, + 0.011228149756789207, + 0.009511579759418964, + 0.013563842512667179, + 0.011641530320048332, + 0.011164376512169838, + 0.01666143536567688, + 0.0191029142588377, + 0.015087711624801159, + 0.01553264632821083, + 0.009624120779335499, + 0.008583142422139645, + 0.008371868170797825, + 0.005641640163958073, + 0.0183961670845747, + 0.007350663188844919, + 0.012087973766028881, + 0.007797598373144865, + 0.006518261972814798, + 0.015766914933919907, + 0.003715040162205696, + 0.005305847153067589, + 0.006310970522463322, + 0.004645578563213348, + 0.004733976908028126, + 0.003652060404419899, + 0.005965933669358492, + 0.004647101741284132, + 0.0031212535686790943, + 0.00584456417709589, + 0.0028894031420350075, + 0.0017155767418444157, + 0.003689924255013466, + 0.002971532056108117, + 0.0021342190448194742, + 0.002393440343439579, + 0.006526469718664885, + 0.005072762258350849, + 0.00794378574937582, + 0.007044919766485691, + 0.0031298715621232986, + 0.004048730246722698, + 0.00578561145812273, + 0.012010583654046059, + 0.003993067890405655, + 0.004312550649046898, + 0.006456784438341856, + 0.008836532011628151, + 0.0030592784751206636, + 0.003559342585504055, + 0.006333320401608944, + 0.007963651791214943, + 0.004268506541848183, + 0.006092963740229607, + 0.004439506214112043, + 0.006595042068511248, + 0.0042780241928994656, + 0.0037104040384292603, + 0.006638081278651953, + 0.005977527238428593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.3132384717464447 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799486130475998 + ], + "opposite_logprob_tokens": [ + -1.8049485683441162 + ], + "logprob_reward": -0.1799486130475998, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799486130475998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0636732429265976, + 0.05758468806743622, + 0.5207175016403198, + 0.07520679384469986, + 0.0339302234351635, + 0.07732453942298889, + 0.04443640634417534, + 0.058527301996946335, + 0.04956142604351044, + 0.019037893041968346 + ], + "credit_map": [ + 0.007091234903782606, + 0.007487064693123102, + 0.005471725016832352, + 0.007954372093081474, + 0.010569063015282154, + 0.009793233126401901, + 0.008972754701972008, + 0.006333793979138136, + 0.0068196202628314495, + 0.006957524456083775, + 0.009022585116326809, + 0.009113444946706295, + 0.007408506236970425, + 0.007716769352555275, + 0.006338671315461397, + 0.004207564052194357, + 0.04838881641626358, + 0.02374381199479103, + 0.07316456735134125, + 0.1489376425743103, + 0.12234625220298767, + 0.03143230080604553, + 0.028347160667181015, + 0.04435694217681885, + 0.007333888206630945, + 0.0063259182497859, + 0.013559448532760143, + 0.01906505785882473, + 0.008232912980020046, + 0.006715166848152876, + 0.007314321119338274, + 0.006660078186541796, + 0.0034033788833767176, + 0.004482168238610029, + 0.004264042712748051, + 0.005495695862919092, + 0.004175330977886915, + 0.0024944888427853584, + 0.003834356553852558, + 0.0057807620614767075, + 0.008505738340318203, + 0.012395092286169529, + 0.018456965684890747, + 0.0079115629196167, + 0.006965178065001965, + 0.005895821377635002, + 0.01008609589189291, + 0.007108086720108986, + 0.0058942874893546104, + 0.006599034182727337, + 0.006118545308709145, + 0.0029441218357533216, + 0.007077586837112904, + 0.006331240758299828, + 0.004844702314585447, + 0.004626891575753689, + 0.007527193520218134, + 0.006419938523322344, + 0.008019834756851196, + 0.0067727151326835155, + 0.005438100080937147, + 0.005522510036826134, + 0.006509246304631233, + 0.012317764572799206, + 0.004756846930831671, + 0.00587255647405982, + 0.005543006118386984, + 0.00922471098601818, + 0.004899436607956886, + 0.004067900124937296, + 0.007295054849237204, + 0.007901915349066257, + 0.0022907163947820663, + 0.0017806976102292538, + 0.0014579190174117684, + 0.002198935719206929, + 0.0014738671015948057, + 0.0012792219640687108, + 0.0058370293118059635, + 0.002719508484005928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.1799486130475998 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371930867433548 + ], + "opposite_logprob_tokens": [ + -3.408719301223755 + ], + "logprob_reward": -0.03371930867433548, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371930867433548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3996601700782776, + 0.07273786514997482, + 0.12128525972366333, + 0.12019170075654984, + 0.04673504829406738, + 0.05327038839459419, + 0.022395893931388855, + 0.03459068015217781, + 0.05384232476353645, + 0.07529065757989883 + ], + "credit_map": [ + 0.04066720977425575, + 0.062141019850969315, + 0.02361130528151989, + 0.033989910036325455, + 0.0797719731926918, + 0.09142749011516571, + 0.03560323268175125, + 0.03244803100824356, + 0.007874230854213238, + 0.007161409594118595, + 0.012748167850077152, + 0.011187274008989334, + 0.009775631129741669, + 0.008332554250955582, + 0.009701873175799847, + 0.0059567224234342575, + 0.013918377459049225, + 0.006707367487251759, + 0.029248163104057312, + 0.009209100157022476, + 0.032756198197603226, + 0.00889334175735712, + 0.007405243813991547, + 0.01314746867865324, + 0.010052299126982689, + 0.010066082701086998, + 0.020128784701228142, + 0.025380296632647514, + 0.01480819657444954, + 0.017077555879950523, + 0.010179415345191956, + 0.012499063275754452, + 0.005338652059435844, + 0.004601055756211281, + 0.007694540079683065, + 0.005088064819574356, + 0.006182590965181589, + 0.003723847446963191, + 0.003961675800383091, + 0.010144617408514023, + 0.005045133642852306, + 0.005599027033895254, + 0.015540868043899536, + 0.005602325312793255, + 0.004031721502542496, + 0.002995115704834461, + 0.008835280314087868, + 0.005620914977043867, + 0.0033173942938447, + 0.005375240929424763, + 0.00328494468703866, + 0.00191391259431839, + 0.002348230918869376, + 0.0019000894390046597, + 0.0021241032518446445, + 0.0021319796796888113, + 0.004015497397631407, + 0.00402784300968051, + 0.005169190466403961, + 0.00434742821380496, + 0.002872369484975934, + 0.003016992239281535, + 0.00481309462338686, + 0.006328266579657793, + 0.004207787569612265, + 0.0035023565869778395, + 0.004799137823283672, + 0.012591821141541004, + 0.0074226162396371365, + 0.00410120002925396, + 0.008038808591663837, + 0.00917859748005867, + 0.012023158371448517, + 0.012168046087026596, + 0.008618699386715889, + 0.01071855053305626, + 0.005850402172654867, + 0.005861198529601097, + 0.010827985592186451, + 0.009222617372870445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.03371930867433548 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979274045675993 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.00979274045675993, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979274045675993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14648687839508057, + 0.055865660309791565, + 0.37581199407577515, + 0.13240548968315125, + 0.06052270531654358, + 0.07807638496160507, + 0.035801127552986145, + 0.025101516395807266, + 0.055518098175525665, + 0.034410182386636734 + ], + "credit_map": [ + 0.016126474365592003, + 0.021148307248950005, + 0.011046416126191616, + 0.014011871069669724, + 0.03141368180513382, + 0.023495765402913094, + 0.01799866557121277, + 0.01124570332467556, + 0.005042699631303549, + 0.006675665732473135, + 0.009503242559731007, + 0.011227964423596859, + 0.005788698792457581, + 0.004826693795621395, + 0.0077232494950294495, + 0.005077451001852751, + 0.04142402485013008, + 0.020250163972377777, + 0.03663969784975052, + 0.09095586091279984, + 0.1027604192495346, + 0.026051126420497894, + 0.02183578722178936, + 0.03589491546154022, + 0.011916527524590492, + 0.013574181124567986, + 0.01974172331392765, + 0.022032078355550766, + 0.01763657107949257, + 0.02167104184627533, + 0.012408207170665264, + 0.013425160199403763, + 0.0072834608145058155, + 0.0068465168587863445, + 0.00744754308834672, + 0.007055363617837429, + 0.005390834994614124, + 0.007016963325440884, + 0.005165324546396732, + 0.014316700398921967, + 0.006937527097761631, + 0.00883581954985857, + 0.021323109045624733, + 0.007912077941000462, + 0.007242299150675535, + 0.007206661626696587, + 0.012094696052372456, + 0.0065242028795182705, + 0.00467664236202836, + 0.006834321189671755, + 0.004340951796621084, + 0.002129125874489546, + 0.004129650536924601, + 0.005850581917911768, + 0.003732724580913782, + 0.004107131622731686, + 0.0022057865280658007, + 0.001963652903214097, + 0.0033268609549850225, + 0.003419565036892891, + 0.0036216892767697573, + 0.0024033975787460804, + 0.0030478942207992077, + 0.005112670361995697, + 0.0038597669918090105, + 0.004243051633238792, + 0.004928382579237223, + 0.012765499763190746, + 0.006914190016686916, + 0.0044119274243712425, + 0.007571087218821049, + 0.01082419604063034, + 0.004020026419311762, + 0.0035434612073004246, + 0.0033166171051561832, + 0.0034461377654224634, + 0.005233241710811853, + 0.002584862755611539, + 0.008041835390031338, + 0.004224002826958895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.00979274045675993 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192009344696999 + ], + "opposite_logprob_tokens": [ + -4.2641921043396 + ], + "logprob_reward": -0.014192009344696999, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192009344696999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13253425061702728, + 0.03993605077266693, + 0.395516037940979, + 0.048205699771642685, + 0.0569470040500164, + 0.1691582053899765, + 0.07330817729234695, + 0.021822331473231316, + 0.04429680109024048, + 0.01827549934387207 + ], + "credit_map": [ + 0.012817499227821827, + 0.020351724699139595, + 0.009691190905869007, + 0.01153299119323492, + 0.02655363455414772, + 0.027118930593132973, + 0.01458777580410242, + 0.009880490601062775, + 0.004110794048756361, + 0.004504116252064705, + 0.005870911758393049, + 0.009134305641055107, + 0.004785545170307159, + 0.0031603833194822073, + 0.005570520646870136, + 0.0027994713746011257, + 0.033481791615486145, + 0.01641528494656086, + 0.06537981331348419, + 0.1213764175772667, + 0.08410529792308807, + 0.021333081647753716, + 0.020952243357896805, + 0.03247208520770073, + 0.0038079959340393543, + 0.004304729402065277, + 0.008329528383910656, + 0.010068679228425026, + 0.004928361624479294, + 0.007972621358931065, + 0.004235370084643364, + 0.0045584081672132015, + 0.006502264644950628, + 0.005216773599386215, + 0.009610569104552269, + 0.0063627734780311584, + 0.008256916888058186, + 0.004489007405936718, + 0.005734780803322792, + 0.010773914866149426, + 0.0165583286434412, + 0.019741684198379517, + 0.0503823347389698, + 0.014400339685380459, + 0.01417139358818531, + 0.011170499958097935, + 0.029607193544507027, + 0.013126414269208908, + 0.010027515701949596, + 0.02130618505179882, + 0.009075461886823177, + 0.005726712755858898, + 0.0075982133857905865, + 0.006077513564378023, + 0.006808810867369175, + 0.006687758024781942, + 0.002878469182178378, + 0.002226766664534807, + 0.0030431526247411966, + 0.0028932923451066017, + 0.0018593358108773828, + 0.0025367627386003733, + 0.0028809544164687395, + 0.0035035961773246527, + 0.003140711924061179, + 0.0033119511790573597, + 0.004239625297486782, + 0.010907704941928387, + 0.006184240337461233, + 0.0035250866785645485, + 0.006357086822390556, + 0.006630393210798502, + 0.002056615427136421, + 0.0018071391386911273, + 0.0020936056971549988, + 0.0024753848556429148, + 0.0014623282477259636, + 0.001026199315674603, + 0.005184260196983814, + 0.0021699650678783655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.014192009344696999 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, low-angle shot, the camera slowly panning to the right, sweeps across a poignant scene frozen in time. A thin, morning mist, like the sighs of fallen soldiers, slowly rises from the silent ruins of a battlefield, winding like a ghost across the muddy, charred earth. The skyline is pierced by the faint, almost grayish light of the rising sun. This cold light penetrates the mist, casting long, soft shadows on the damp metal and earth. Broken spears, notched sword blades, and shattered shields riddled with arrows lie scattered across the ground, their cold metal surfaces reflecting a faint light. Several torn battle flags, their faded crimson and gold patterns blurred, hang limply from broken poles, occasionally swaying gently in the barely perceptible breeze. The main colors of the entire scene are composed of earthy brown, charcoal black and cold metallic gray, creating a solemn, sad and tranquil atmosphere, full of a heavy sense of history.", + "seed": 13946, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_9/index_2607_seed_13946_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08939985847100615, + "vlm_logprob": -0.08939985847100615, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.08939985847100615, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1958627700805664 + ], + "opposite_logprob_tokens": [ + -1.9458627700805664 + ], + "logprob_reward": -0.1958627700805664, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1958627700805664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1782960146665573, + 0.05597752332687378, + 0.1619929075241089, + 0.039835136383771896, + 0.03380563482642174, + 0.024219514802098274, + 0.026950746774673462, + 0.2839241921901703, + 0.044745054095983505, + 0.15025320649147034 + ], + "credit_map": [ + 0.012904840521514416, + 0.020225554704666138, + 0.0288451686501503, + 0.04769036918878555, + 0.016835007816553116, + 0.01652030274271965, + 0.02644554153084755, + 0.008829228579998016, + 0.005024440586566925, + 0.011443513445556164, + 0.007569293025881052, + 0.01026122085750103, + 0.005232346244156361, + 0.004307808354496956, + 0.007627536077052355, + 0.0045113638043403625, + 0.012484521605074406, + 0.015093475580215454, + 0.038533296436071396, + 0.019496886059641838, + 0.038744159042835236, + 0.012727541849017143, + 0.015930278226733208, + 0.008982758037745953, + 0.0032619170378893614, + 0.005794927012175322, + 0.005415583960711956, + 0.006052117794752121, + 0.006131256464868784, + 0.005056773778051138, + 0.005010811612010002, + 0.0031117466278374195, + 0.005933041218668222, + 0.005165102891623974, + 0.0050935591571033, + 0.004878675099462271, + 0.0034660061355680227, + 0.0038134315982460976, + 0.0028736134991049767, + 0.002582205692306161, + 0.002687753178179264, + 0.0028189229778945446, + 0.0032069957815110683, + 0.0029919324442744255, + 0.0028558538760989904, + 0.0025350016076117754, + 0.002563097281381488, + 0.0045599564909935, + 0.003148514311760664, + 0.00292822509072721, + 0.0035636264365166426, + 0.003444981761276722, + 0.0041974070481956005, + 0.002318461425602436, + 0.004047955386340618, + 0.0033015760127454996, + 0.029673654586076736, + 0.03980186954140663, + 0.031163018196821213, + 0.05119730532169342, + 0.02898969128727913, + 0.050078392028808594, + 0.016739094629883766, + 0.03628116846084595, + 0.010357269085943699, + 0.006178396288305521, + 0.005746189504861832, + 0.005182701162993908, + 0.00428044144064188, + 0.005644521676003933, + 0.0036383303813636303, + 0.0037172031588852406, + 0.021456481888890266, + 0.04244197532534599, + 0.014984180219471455, + 0.011014247313141823, + 0.03235248476266861, + 0.007839908823370934, + 0.006059272680431604, + 0.014104649424552917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.1958627700805664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560718655586243 + ], + "opposite_logprob_tokens": [ + -1.60060715675354 + ], + "logprob_reward": -0.22560718655586243, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560718655586243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06116323545575142, + 0.024724898859858513, + 0.1537507176399231, + 0.025205107405781746, + 0.03360602632164955, + 0.04157304763793945, + 0.033099640160799026, + 0.19648593664169312, + 0.14416368305683136, + 0.2862277030944824 + ], + "credit_map": [ + 0.004594531841576099, + 0.008568050339818, + 0.009929914027452469, + 0.013292412273585796, + 0.004950082395225763, + 0.006425515748560429, + 0.008541641756892204, + 0.004861088469624519, + 0.002219239016994834, + 0.004087736364454031, + 0.003928068559616804, + 0.004348762799054384, + 0.0020515923388302326, + 0.0024576461873948574, + 0.00320342555642128, + 0.0024284296669065952, + 0.011927417479455471, + 0.02680261805653572, + 0.028530843555927277, + 0.030831707641482353, + 0.02383323572576046, + 0.00977586954832077, + 0.013844193890690804, + 0.008204830810427666, + 0.003452304285019636, + 0.003619366092607379, + 0.003431647550314665, + 0.0037076829466968775, + 0.0026301315519958735, + 0.0026090219616889954, + 0.0023983526043593884, + 0.003356602508574724, + 0.0029234914109110832, + 0.005500458180904388, + 0.005316575523465872, + 0.005225551314651966, + 0.003892656182870269, + 0.004103658255189657, + 0.003176045138388872, + 0.0034675931092351675, + 0.004048565402626991, + 0.005011259112507105, + 0.0049425652250647545, + 0.0059571941383183, + 0.006168201565742493, + 0.007528272923082113, + 0.003585400292649865, + 0.004331593401730061, + 0.0038462646771222353, + 0.004168327897787094, + 0.0039000287652015686, + 0.004527145065367222, + 0.004674280993640423, + 0.005881402175873518, + 0.002657744800671935, + 0.0034444481134414673, + 0.02348390594124794, + 0.02714294195175171, + 0.020828742533922195, + 0.030444148927927017, + 0.02096371538937092, + 0.03821375593543053, + 0.01104560773819685, + 0.024363132193684578, + 0.020570803433656693, + 0.02477729134261608, + 0.021517198532819748, + 0.015814492478966713, + 0.013521961867809296, + 0.02326635830104351, + 0.011893357150256634, + 0.012802226468920708, + 0.02585316076874733, + 0.08758898079395294, + 0.02474989928305149, + 0.022928990423679352, + 0.06678018718957901, + 0.019221991300582886, + 0.01421361044049263, + 0.024890881031751633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.22560718655586243 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794747456908226 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.009794747456908226, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794747456908226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11644130200147629, + 0.03997083380818367, + 0.3099271357059479, + 0.0197195652872324, + 0.024058407172560692, + 0.06306667625904083, + 0.027083583176136017, + 0.12008132040500641, + 0.08485398441553116, + 0.19479718804359436 + ], + "credit_map": [ + 0.012891478836536407, + 0.021241404116153717, + 0.0150452870875597, + 0.017117146402597427, + 0.012421178631484509, + 0.010284708812832832, + 0.015129873529076576, + 0.012310225516557693, + 0.00439430121332407, + 0.007113385479897261, + 0.005745452828705311, + 0.007400772534310818, + 0.0038770553655922413, + 0.00292581575922668, + 0.004545826930552721, + 0.003968226257711649, + 0.027388019487261772, + 0.014012555591762066, + 0.08806174248456955, + 0.025138475000858307, + 0.09226098656654358, + 0.026608772575855255, + 0.021181771531701088, + 0.015274799428880215, + 0.0017172627849504352, + 0.002449226099997759, + 0.002999022603034973, + 0.0029321270994842052, + 0.001946466974914074, + 0.0022328533232212067, + 0.00304280873388052, + 0.0023997968528419733, + 0.0026380319613963366, + 0.003677153028547764, + 0.003680974245071411, + 0.0035364488139748573, + 0.0027351737953722477, + 0.002784223295748234, + 0.002741138217970729, + 0.0022652642801404, + 0.007141243200749159, + 0.006779736373573542, + 0.008217468857765198, + 0.006785654462873936, + 0.012621436268091202, + 0.0057180700823664665, + 0.008424721658229828, + 0.0073783486150205135, + 0.0030826847068965435, + 0.0028563211672008038, + 0.0033130012452602386, + 0.003511544782668352, + 0.004409191198647022, + 0.0028208564035594463, + 0.0040770587511360645, + 0.0030129235237836838, + 0.014765537343919277, + 0.017004892230033875, + 0.013293773867189884, + 0.021382851526141167, + 0.011941504664719105, + 0.019858919084072113, + 0.0071225883439183235, + 0.014711255207657814, + 0.020591288805007935, + 0.009379573166370392, + 0.014237534254789352, + 0.008135388605296612, + 0.01085189450532198, + 0.008138082921504974, + 0.005420590750873089, + 0.008099631406366825, + 0.01755974441766739, + 0.04520069807767868, + 0.017306627705693245, + 0.02094554342329502, + 0.029316529631614685, + 0.02462272346019745, + 0.019123414531350136, + 0.020721904933452606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -0.009794747456908226 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978624727576971 + ], + "opposite_logprob_tokens": [ + -4.634786128997803 + ], + "logprob_reward": -0.00978624727576971, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978624727576971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09203982353210449, + 0.05905616655945778, + 0.3429581820964813, + 0.05598289147019386, + 0.06268060952425003, + 0.018683258444070816, + 0.015200136229395866, + 0.0889628529548645, + 0.1681908816099167, + 0.09624514728784561 + ], + "credit_map": [ + 0.011183698661625385, + 0.01727316528558731, + 0.011894488707184792, + 0.01622866652905941, + 0.008397321216762066, + 0.007477649487555027, + 0.011185536161065102, + 0.00839931145310402, + 0.0050962320528924465, + 0.007013083901256323, + 0.009825658053159714, + 0.012205070815980434, + 0.004664640408009291, + 0.0039892420172691345, + 0.01104450598359108, + 0.005217743571847677, + 0.024509061127901077, + 0.054440468549728394, + 0.07645102590322495, + 0.0580962598323822, + 0.06551045924425125, + 0.02172910049557686, + 0.024799074977636337, + 0.017422737553715706, + 0.0058046323247253895, + 0.011457858607172966, + 0.008852148428559303, + 0.011695992201566696, + 0.004470298998057842, + 0.004274516366422176, + 0.006376796867698431, + 0.003050651866942644, + 0.005313088186085224, + 0.010450414381921291, + 0.00955587811768055, + 0.011082540266215801, + 0.007086738478392363, + 0.007335433270782232, + 0.005980344954878092, + 0.0058761783875525, + 0.0021939543075859547, + 0.00229659560136497, + 0.002301391214132309, + 0.0023689609952270985, + 0.0022716207895427942, + 0.0020547450985759497, + 0.0020744442008435726, + 0.00312154833227396, + 0.001976592931896448, + 0.0016104340320453048, + 0.0017714075511321425, + 0.0017169959610328078, + 0.0023385011591017246, + 0.0013603465631604195, + 0.0020177941769361496, + 0.002408064668998122, + 0.010946724563837051, + 0.012415310367941856, + 0.009951595216989517, + 0.014030816033482552, + 0.009382477961480618, + 0.015739699825644493, + 0.004941274877637625, + 0.01155495923012495, + 0.03722523897886276, + 0.01601644977927208, + 0.03170487657189369, + 0.01807260327041149, + 0.0252592284232378, + 0.015999769791960716, + 0.008677504025399685, + 0.015235230326652527, + 0.013039331883192062, + 0.011891760863363743, + 0.01271485723555088, + 0.011667052283883095, + 0.010284626856446266, + 0.010310027748346329, + 0.0136718163266778, + 0.01266568060964346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.00978624727576971 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059483409859240055 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.0059483409859240055, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059483409859240055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10051023960113525, + 0.09068647027015686, + 0.39086267352104187, + 0.06607764214277267, + 0.028791289776563644, + 0.03186932951211929, + 0.03402525931596756, + 0.045124877244234085, + 0.08547820895910263, + 0.12657403945922852 + ], + "credit_map": [ + 0.009329685010015965, + 0.01998923532664776, + 0.012987443245947361, + 0.020578522235155106, + 0.008867518976330757, + 0.0072901551611721516, + 0.012031210586428642, + 0.0094364695250988, + 0.008951141498982906, + 0.020576277747750282, + 0.012613551691174507, + 0.014425884000957012, + 0.008909059688448906, + 0.00591572280973196, + 0.010728850029408932, + 0.008565993048250675, + 0.02856508642435074, + 0.018639136105775833, + 0.11282116174697876, + 0.03187825530767441, + 0.1242261454463005, + 0.035743456333875656, + 0.024276964366436005, + 0.01471246499568224, + 0.009119575843214989, + 0.009176104329526424, + 0.01012915838509798, + 0.009762268513441086, + 0.00721772201359272, + 0.006156515795737505, + 0.008658588863909245, + 0.005857708863914013, + 0.0030265164095908403, + 0.004770966246724129, + 0.0037024756893515587, + 0.004646399989724159, + 0.0034314109943807125, + 0.0035359107423573732, + 0.003476365702226758, + 0.00220124376937747, + 0.00517281936481595, + 0.003458276391029358, + 0.0045823208056390285, + 0.0037036091089248657, + 0.004536127671599388, + 0.002416814910247922, + 0.0038062918465584517, + 0.004193067084997892, + 0.003913944587111473, + 0.004208061844110489, + 0.004354956094175577, + 0.004752613604068756, + 0.004891213029623032, + 0.0025241118855774403, + 0.004232937470078468, + 0.00514741986989975, + 0.0061461967416107655, + 0.00580778019502759, + 0.005129138473421335, + 0.008347949013113976, + 0.004520602058619261, + 0.006251603830605745, + 0.003788864240050316, + 0.00513274222612381, + 0.0172362569719553, + 0.010715121403336525, + 0.014659850858151913, + 0.009735089726746082, + 0.012352734804153442, + 0.009154196828603745, + 0.004255570005625486, + 0.007369387894868851, + 0.02062583900988102, + 0.027660401538014412, + 0.013130232691764832, + 0.010872839950025082, + 0.023770267143845558, + 0.007983753457665443, + 0.010949654504656792, + 0.011581046506762505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.0059483409859240055 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, low-angle shot, the camera slowly panning to the right, sweeps across a poignant scene frozen in time. A thin, morning mist, like the sighs of fallen soldiers, slowly rises from the silent ruins of a battlefield, winding like a ghost across the muddy, charred earth. The skyline is pierced by the faint, almost grayish light of the rising sun. This cold light penetrates the mist, casting long, soft shadows on the damp metal and earth. Broken spears, notched sword blades, and shattered shields riddled with arrows lie scattered across the ground, their cold metal surfaces reflecting a faint light. Several torn battle flags, their faded crimson and gold patterns blurred, hang limply from broken poles, occasionally swaying gently in the barely perceptible breeze. The main colors of the entire scene are composed of earthy brown, charcoal black and cold metallic gray, creating a solemn, sad and tranquil atmosphere, full of a heavy sense of history.", + "seed": 13947, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_9/index_2607_seed_13947_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06020817076787353, + "vlm_logprob": -0.06020817076787353, + "vlm_yes_no_margin": 3.599999952316284, + "vlm_token_credit": -0.06020817076787353, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16713625192642212 + ], + "opposite_logprob_tokens": [ + -2.1671361923217773 + ], + "logprob_reward": -0.16713625192642212, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.16713625192642212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1576594114303589, + 0.08433200418949127, + 0.08272360265254974, + 0.1747683435678482, + 0.11117904633283615, + 0.04879290238022804, + 0.06413429975509644, + 0.07434095442295074, + 0.09239751845598221, + 0.10967185348272324 + ], + "credit_map": [ + 0.019881822168827057, + 0.01371471956372261, + 0.021640893071889877, + 0.020654866471886635, + 0.01663847267627716, + 0.0209491103887558, + 0.020038535818457603, + 0.024140993133187294, + 0.00728615652769804, + 0.013267873786389828, + 0.008263926021754742, + 0.00874380674213171, + 0.014634844847023487, + 0.009729539044201374, + 0.012700071558356285, + 0.009705786593258381, + 0.007306333631277084, + 0.008477798663079739, + 0.011183545924723148, + 0.0143589461222291, + 0.006583262700587511, + 0.009518811479210854, + 0.015279802493751049, + 0.010015095584094524, + 0.02748589776456356, + 0.0209426898509264, + 0.01724843680858612, + 0.01713884249329567, + 0.0220657866448164, + 0.03299485892057419, + 0.018696580082178116, + 0.0181952565908432, + 0.011840008199214935, + 0.009117907844483852, + 0.014548840932548046, + 0.01295480690896511, + 0.016391785815358162, + 0.018398810178041458, + 0.013499567285180092, + 0.014427312649786472, + 0.0060042329132556915, + 0.007293816190212965, + 0.005529732909053564, + 0.004418870434165001, + 0.00390327419154346, + 0.006983606610447168, + 0.00682596443220973, + 0.007833404466509819, + 0.007752286270260811, + 0.010216849856078625, + 0.004379428457468748, + 0.005115871783345938, + 0.018348727375268936, + 0.0075753796845674515, + 0.0059460224583745, + 0.004799737129360437, + 0.006544749718159437, + 0.01448694709688425, + 0.007199696730822325, + 0.009286848828196526, + 0.010108181275427341, + 0.007128510624170303, + 0.010544076561927795, + 0.009041942656040192, + 0.015180334448814392, + 0.011163858696818352, + 0.007604295387864113, + 0.007413513492792845, + 0.020873526111245155, + 0.011789494194090366, + 0.009205921553075314, + 0.009166574105620384, + 0.020772045478224754, + 0.008559531532227993, + 0.021270954981446266, + 0.010799056850373745, + 0.011562304571270943, + 0.009407158941030502, + 0.011472444981336594, + 0.015828358009457588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.16713625192642212 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032735764980316 + ], + "opposite_logprob_tokens": [ + -2.350327253341675 + ], + "logprob_reward": -0.10032735764980316, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032735764980316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11019415408372879, + 0.057959046214818954, + 0.07893292605876923, + 0.06055312603712082, + 0.11142028123140335, + 0.04798009619116783, + 0.09578148275613785, + 0.166587233543396, + 0.19891144335269928, + 0.07168024033308029 + ], + "credit_map": [ + 0.01286587305366993, + 0.00938117690384388, + 0.013398169539868832, + 0.009780874475836754, + 0.010256703943014145, + 0.01657269150018692, + 0.014234072528779507, + 0.02370459958910942, + 0.009610596112906933, + 0.008533908985555172, + 0.005428867414593697, + 0.0063691711984574795, + 0.008858432061970234, + 0.005930830724537373, + 0.006079301703721285, + 0.007147936150431633, + 0.005602739285677671, + 0.008555734530091286, + 0.008705254644155502, + 0.011001302860677242, + 0.006326820235699415, + 0.012481670826673508, + 0.017304107546806335, + 0.008955293335020542, + 0.012549953535199165, + 0.005536799319088459, + 0.006857826840132475, + 0.0059485482051968575, + 0.00690827053040266, + 0.010389391332864761, + 0.0065228003077209, + 0.005839531775563955, + 0.012873494997620583, + 0.008946851827204227, + 0.015080302022397518, + 0.011763934046030045, + 0.01680304855108261, + 0.024353237822651863, + 0.011934971436858177, + 0.009664435870945454, + 0.006991509348154068, + 0.004597604274749756, + 0.005838220007717609, + 0.004390000831335783, + 0.00646784994751215, + 0.008583971299231052, + 0.007279449142515659, + 0.003831488313153386, + 0.012201141566038132, + 0.0092444634065032, + 0.010635241866111755, + 0.011586910113692284, + 0.01178815495222807, + 0.008999167941510677, + 0.01666245236992836, + 0.01466395054012537, + 0.023639189079403877, + 0.027406373992562294, + 0.01574195921421051, + 0.01211225613951683, + 0.04462166503071785, + 0.01466531865298748, + 0.014085207134485245, + 0.014315265230834484, + 0.032530464231967926, + 0.021212702617049217, + 0.020683487877249718, + 0.026257349178195, + 0.022716937586665154, + 0.0271371491253376, + 0.02367396466434002, + 0.024699386209249496, + 0.008308950811624527, + 0.008401461876928806, + 0.0060138157568871975, + 0.005462135188281536, + 0.020775431767106056, + 0.008602491579949856, + 0.006489490158855915, + 0.00762645760551095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.10032735764980316 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200064077973366 + ], + "opposite_logprob_tokens": [ + -4.018199920654297 + ], + "logprob_reward": -0.018200064077973366, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018200064077973366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11799075454473495, + 0.16872912645339966, + 0.06058868393301964, + 0.10791418701410294, + 0.10855133831501007, + 0.05909980088472366, + 0.06646682322025299, + 0.13316789269447327, + 0.13169503211975098, + 0.04579631984233856 + ], + "credit_map": [ + 0.014523463323712349, + 0.008149824105203152, + 0.015208078548312187, + 0.013359226286411285, + 0.011929438449442387, + 0.01988641358911991, + 0.016561582684516907, + 0.018372731283307076, + 0.01237792894244194, + 0.030640846118330956, + 0.016341259703040123, + 0.019544607028365135, + 0.027603238821029663, + 0.013492274098098278, + 0.01999257132411003, + 0.02873639576137066, + 0.007776238024234772, + 0.006363988388329744, + 0.006870383396744728, + 0.007479039952158928, + 0.004789353348314762, + 0.00932632852345705, + 0.009500741958618164, + 0.008482609875500202, + 0.022668851539492607, + 0.009453603066504002, + 0.014692429453134537, + 0.009503364562988281, + 0.010859251953661442, + 0.01711157150566578, + 0.012586615048348904, + 0.011038499884307384, + 0.012509413994848728, + 0.009022454731166363, + 0.012583802454173565, + 0.015743428841233253, + 0.01403253898024559, + 0.01270991936326027, + 0.019037097692489624, + 0.012912681326270103, + 0.006054764147847891, + 0.006535335909575224, + 0.005704708397388458, + 0.0060371640138328075, + 0.00958542712032795, + 0.0065419492311775684, + 0.009500053711235523, + 0.00914040207862854, + 0.007400174625217915, + 0.007508102338761091, + 0.007163986563682556, + 0.007197704631835222, + 0.015170902013778687, + 0.006113090552389622, + 0.007547306828200817, + 0.008365556597709656, + 0.022987257689237595, + 0.014586149714887142, + 0.013492894358932972, + 0.014652961865067482, + 0.020870033651590347, + 0.010102657601237297, + 0.02220962755382061, + 0.01426631584763527, + 0.021352743729948997, + 0.01622864603996277, + 0.0147332102060318, + 0.016373654827475548, + 0.017355667427182198, + 0.017835741862654686, + 0.013269099406898022, + 0.014546267688274384, + 0.0077050174586474895, + 0.0038553897757083178, + 0.004801946692168713, + 0.004120318219065666, + 0.009450032375752926, + 0.00616878317669034, + 0.005002402234822512, + 0.004692430607974529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.018200064077973366 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643122389912605 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643122389912605, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643122389912605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0854625552892685, + 0.0891997292637825, + 0.22475723922252655, + 0.1779547780752182, + 0.12893062829971313, + 0.030766021460294724, + 0.04827296361327171, + 0.05671810731291771, + 0.11232532560825348, + 0.045612625777721405 + ], + "credit_map": [ + 0.010317311622202396, + 0.010659589432179928, + 0.009511543437838554, + 0.008138332515954971, + 0.008470752276480198, + 0.014568261802196503, + 0.010793542489409447, + 0.013003222644329071, + 0.00781252607703209, + 0.013022087514400482, + 0.009587172418832779, + 0.009867453947663307, + 0.012344387359917164, + 0.010729275643825531, + 0.012203098274767399, + 0.013633727096021175, + 0.018073193728923798, + 0.02760057896375656, + 0.02484235353767872, + 0.016707338392734528, + 0.021882768720388412, + 0.07730700820684433, + 0.017580337822437286, + 0.02076364867389202, + 0.0391312912106514, + 0.019229663535952568, + 0.01833171769976616, + 0.015212570317089558, + 0.022121749818325043, + 0.030003154650330544, + 0.019514672458171844, + 0.014409956522285938, + 0.017624130472540855, + 0.011394907720386982, + 0.015688611194491386, + 0.015211968682706356, + 0.017347777262330055, + 0.020979318767786026, + 0.018168218433856964, + 0.01251569390296936, + 0.005611071828752756, + 0.002672543516382575, + 0.0033108678180724382, + 0.0021780659444630146, + 0.0050775897689163685, + 0.007116545457392931, + 0.002667830092832446, + 0.0021315058693289757, + 0.006137613672763109, + 0.00443153316155076, + 0.0055140163749456406, + 0.006172132678329945, + 0.008484664373099804, + 0.004595795180648565, + 0.005919960793107748, + 0.007017251569777727, + 0.006626889109611511, + 0.009766456671059132, + 0.005828343331813812, + 0.005795623175799847, + 0.009747805073857307, + 0.005234311800450087, + 0.007574578747153282, + 0.00614409614354372, + 0.01956951431930065, + 0.012184943072497845, + 0.011331640183925629, + 0.01328088529407978, + 0.018274426460266113, + 0.012343146838247776, + 0.011521754786372185, + 0.0138190146535635, + 0.005837698467075825, + 0.004460819065570831, + 0.005198935512453318, + 0.0029065702110528946, + 0.012512448243796825, + 0.0066900537349283695, + 0.0034549818374216557, + 0.004551121033728123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.008643122389912605 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734057795256376 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006734057795256376, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734057795256376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15884104371070862, + 0.11858651787042618, + 0.06319592893123627, + 0.06409309804439545, + 0.13169671595096588, + 0.08603622019290924, + 0.06478133797645569, + 0.06583517044782639, + 0.1958058923482895, + 0.05112805962562561 + ], + "credit_map": [ + 0.013755342923104763, + 0.010478866286575794, + 0.01992642879486084, + 0.014873108826577663, + 0.013308924622833729, + 0.021341713145375252, + 0.024173688143491745, + 0.04098297655582428, + 0.0189859289675951, + 0.012814493849873543, + 0.010476584546267986, + 0.015745172277092934, + 0.011832359246909618, + 0.011108643375337124, + 0.02200736291706562, + 0.015615969896316528, + 0.006045665126293898, + 0.00620753038674593, + 0.006314839702099562, + 0.006481457035988569, + 0.005922191310673952, + 0.015241566114127636, + 0.010852247476577759, + 0.006130428984761238, + 0.005706381052732468, + 0.005968947894871235, + 0.007350839674472809, + 0.006810957565903664, + 0.007450149394571781, + 0.014284627512097359, + 0.009903855621814728, + 0.006617339327931404, + 0.015128746628761292, + 0.011542242020368576, + 0.012792486697435379, + 0.008092546835541725, + 0.017997536808252335, + 0.03974773734807968, + 0.012354450300335884, + 0.014040974900126457, + 0.013450449332594872, + 0.010737400501966476, + 0.007888245396316051, + 0.006870257202535868, + 0.016449788585305214, + 0.01643138751387596, + 0.006200151052325964, + 0.008008548058569431, + 0.008015801198780537, + 0.009984712116420269, + 0.004845714662224054, + 0.006286171730607748, + 0.016596760600805283, + 0.0069684116169810295, + 0.006116900593042374, + 0.005966866388916969, + 0.005933406762778759, + 0.008714310824871063, + 0.005720240529626608, + 0.006657568272203207, + 0.017006121575832367, + 0.0073907338082790375, + 0.007520184386521578, + 0.006892603822052479, + 0.025015562772750854, + 0.026526696979999542, + 0.013127924874424934, + 0.015300617553293705, + 0.04804457724094391, + 0.028677554801106453, + 0.01854868419468403, + 0.020564263686537743, + 0.009480427019298077, + 0.005320831201970577, + 0.007046459708362818, + 0.005382560193538666, + 0.0059747155755758286, + 0.005328636150807142, + 0.006399231962859631, + 0.0061951931565999985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.006734057795256376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, low-angle shot, the camera slowly panning to the right, sweeps across a poignant scene frozen in time. A thin, morning mist, like the sighs of fallen soldiers, slowly rises from the silent ruins of a battlefield, winding like a ghost across the muddy, charred earth. The skyline is pierced by the faint, almost grayish light of the rising sun. This cold light penetrates the mist, casting long, soft shadows on the damp metal and earth. Broken spears, notched sword blades, and shattered shields riddled with arrows lie scattered across the ground, their cold metal surfaces reflecting a faint light. Several torn battle flags, their faded crimson and gold patterns blurred, hang limply from broken poles, occasionally swaying gently in the barely perceptible breeze. The main colors of the entire scene are composed of earthy brown, charcoal black and cold metallic gray, creating a solemn, sad and tranquil atmosphere, full of a heavy sense of history.", + "seed": 13948, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000063/group_9/index_2607_seed_13948_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0911154704168439, + "vlm_logprob": -0.0911154704168439, + "vlm_yes_no_margin": 3.274999976158142, + "vlm_token_credit": -0.0911154704168439, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1889304369688034 + ], + "opposite_logprob_tokens": [ + -2.0639305114746094 + ], + "logprob_reward": -0.1889304369688034, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1889304369688034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15909063816070557, + 0.33387529850006104, + 0.0303680170327425, + 0.060703977942466736, + 0.09753590077161789, + 0.06312473863363266, + 0.05292323976755142, + 0.05158945918083191, + 0.10396679490804672, + 0.04682191088795662 + ], + "credit_map": [ + 0.016361959278583527, + 0.025696445256471634, + 0.022392239421606064, + 0.0248873308300972, + 0.022044304758310318, + 0.012277140282094479, + 0.016188599169254303, + 0.01924261264503002, + 0.10954249650239944, + 0.10657186806201935, + 0.02077539451420307, + 0.020894665271043777, + 0.022301282733678818, + 0.02453780360519886, + 0.018300745636224747, + 0.010951049625873566, + 0.0035099012311547995, + 0.004208819940686226, + 0.004033067263662815, + 0.0045124138705432415, + 0.004635774064809084, + 0.003630847204476595, + 0.0029805435333400965, + 0.002856649225577712, + 0.008987144567072392, + 0.013952826149761677, + 0.0059744808822870255, + 0.006662094034254551, + 0.00890719797462225, + 0.004811698105186224, + 0.00769727723672986, + 0.0037112566642463207, + 0.009196951985359192, + 0.014934221282601357, + 0.015936359763145447, + 0.023314863443374634, + 0.011333758011460304, + 0.008073864504694939, + 0.00800012331455946, + 0.0067457580007612705, + 0.010201201774179935, + 0.02281259186565876, + 0.003983624745160341, + 0.003845714731141925, + 0.006834040395915508, + 0.0034208863507956266, + 0.0035699433647096157, + 0.008456732146441936, + 0.00581284100189805, + 0.009738952852785587, + 0.00786621868610382, + 0.005874511785805225, + 0.006270664278417826, + 0.005076925735920668, + 0.007057455368340015, + 0.005225666798651218, + 0.006399448961019516, + 0.005828774068504572, + 0.0041048238053917885, + 0.008302683010697365, + 0.0067789433524012566, + 0.009946920908987522, + 0.006280686240643263, + 0.0039471774362027645, + 0.011729398742318153, + 0.017693260684609413, + 0.010628307238221169, + 0.010935916565358639, + 0.01060409378260374, + 0.011159515008330345, + 0.01677744835615158, + 0.014438853599131107, + 0.00636783242225647, + 0.014512347057461739, + 0.0030665055382996798, + 0.0040375832468271255, + 0.005001181736588478, + 0.0038290205411612988, + 0.004109138622879982, + 0.005898300092667341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.1889304369688034 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22564058005809784 + ], + "opposite_logprob_tokens": [ + -1.6006405353546143 + ], + "logprob_reward": -0.22564058005809784, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22564058005809784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10011795908212662, + 0.11626549065113068, + 0.0375000461935997, + 0.0745779275894165, + 0.061999037861824036, + 0.19541862607002258, + 0.06991412490606308, + 0.054223399609327316, + 0.07771162688732147, + 0.21227172017097473 + ], + "credit_map": [ + 0.015532630495727062, + 0.011047561652958393, + 0.013635051436722279, + 0.011280668899416924, + 0.009935621172189713, + 0.014018272049725056, + 0.014879350550472736, + 0.009788802824914455, + 0.029928740113973618, + 0.032994333654642105, + 0.008319730870425701, + 0.0108293192461133, + 0.008119948208332062, + 0.009664947167038918, + 0.00910126231610775, + 0.007307204883545637, + 0.005158169195055962, + 0.004771631211042404, + 0.005380484741181135, + 0.00380250858142972, + 0.003913844935595989, + 0.007445426192134619, + 0.0035831027198582888, + 0.003444878151640296, + 0.009304028935730457, + 0.014036215841770172, + 0.0065317898988723755, + 0.008943486958742142, + 0.011009267531335354, + 0.010072432458400726, + 0.007983054034411907, + 0.00669765193015337, + 0.006189457140862942, + 0.012770257890224457, + 0.005778809078037739, + 0.008077462203800678, + 0.011883692815899849, + 0.006786487530916929, + 0.00566252414137125, + 0.0048503512516617775, + 0.029942994937300682, + 0.0650782361626625, + 0.015838243067264557, + 0.016033872961997986, + 0.026321420446038246, + 0.01521808747202158, + 0.011631002649664879, + 0.015354759991168976, + 0.008379681967198849, + 0.013778670690953732, + 0.012675064615905285, + 0.008127224631607533, + 0.007172842510044575, + 0.005884886719286442, + 0.00814085453748703, + 0.005754897836595774, + 0.007320689037442207, + 0.005742824170738459, + 0.005632084794342518, + 0.0053316340781748295, + 0.013098474591970444, + 0.006403105333447456, + 0.006042866036295891, + 0.0046517252922058105, + 0.010246764868497849, + 0.014716217294335365, + 0.008151871152222157, + 0.006299328058958054, + 0.0070716748014092445, + 0.007116503082215786, + 0.009703004732728004, + 0.014406261965632439, + 0.02471587061882019, + 0.07053089141845703, + 0.014520540833473206, + 0.015208384953439236, + 0.024499889463186264, + 0.01494269073009491, + 0.015456069260835648, + 0.03239738196134567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.22564058005809784 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608221046626568 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.01608221046626568, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608221046626568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07150042057037354, + 0.36259859800338745, + 0.03879597410559654, + 0.06948943436145782, + 0.04328848049044609, + 0.07635301351547241, + 0.04305778443813324, + 0.06207392364740372, + 0.14631140232086182, + 0.08653093874454498 + ], + "credit_map": [ + 0.013307450339198112, + 0.00931604951620102, + 0.008962654508650303, + 0.007957613095641136, + 0.009455048479139805, + 0.010276987217366695, + 0.006348101422190666, + 0.005876516457647085, + 0.11817732453346252, + 0.1210639625787735, + 0.02217882126569748, + 0.023645637556910515, + 0.02200658619403839, + 0.019344396889209747, + 0.02319503016769886, + 0.012986846268177032, + 0.007112937048077583, + 0.00783886481076479, + 0.005500948987901211, + 0.0050897435285151005, + 0.00291127129457891, + 0.003176873316988349, + 0.00318373111076653, + 0.003981603775173426, + 0.015434285625815392, + 0.02244781330227852, + 0.005480651743710041, + 0.0050826421938836575, + 0.006864537484943867, + 0.004246476572006941, + 0.005010958760976791, + 0.0049220803193748, + 0.005820808000862598, + 0.005780346691608429, + 0.005322971846908331, + 0.006797770503908396, + 0.00430793222039938, + 0.006846627686172724, + 0.0042622447945177555, + 0.00414977828040719, + 0.014468854293227196, + 0.02612333558499813, + 0.006336929742246866, + 0.004921720363199711, + 0.0074394624680280685, + 0.002986750565469265, + 0.004826888907700777, + 0.009249075315892696, + 0.004454385954886675, + 0.006722292862832546, + 0.007106016855686903, + 0.006983357947319746, + 0.003878689371049404, + 0.003703662194311619, + 0.005856179166585207, + 0.004353204742074013, + 0.0054686847142875195, + 0.007482652552425861, + 0.008313687518239021, + 0.007665074896067381, + 0.011752331629395485, + 0.009337290190160275, + 0.006319429259747267, + 0.005734773352742195, + 0.011966037563979626, + 0.016594793647527695, + 0.01869243010878563, + 0.02509138360619545, + 0.011072490364313126, + 0.01881367899477482, + 0.02559089846909046, + 0.018489696085453033, + 0.010294810868799686, + 0.027169113978743553, + 0.006805167067795992, + 0.006168276537209749, + 0.01038939505815506, + 0.007892047986388206, + 0.0064722709357738495, + 0.011339853517711163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.01608221046626568 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067402152344584465 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.0067402152344584465, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067402152344584465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1274491548538208, + 0.16525892913341522, + 0.10530239343643188, + 0.10652591288089752, + 0.10388634353876114, + 0.13846921920776367, + 0.07046500593423843, + 0.03508035093545914, + 0.10580907016992569, + 0.041753582656383514 + ], + "credit_map": [ + 0.018527820706367493, + 0.013045595027506351, + 0.018005570396780968, + 0.014242245815694332, + 0.013917175121605396, + 0.024826601147651672, + 0.012043844908475876, + 0.012840308248996735, + 0.039378900080919266, + 0.051566578447818756, + 0.013305200263857841, + 0.011572563089430332, + 0.012881062924861908, + 0.01677420735359192, + 0.009504469111561775, + 0.01027595903724432, + 0.017044827342033386, + 0.012936555780470371, + 0.0089248763397336, + 0.010050018317997456, + 0.013452969491481781, + 0.022346816956996918, + 0.010001796297729015, + 0.01054452545940876, + 0.01914309710264206, + 0.029473423957824707, + 0.008573165163397789, + 0.009581590071320534, + 0.01127250213176012, + 0.01276257075369358, + 0.008715779520571232, + 0.007003782317042351, + 0.012048558332026005, + 0.010979849845170975, + 0.010264912620186806, + 0.011686808429658413, + 0.011725471355021, + 0.02276010997593403, + 0.013676632195711136, + 0.010743996128439903, + 0.0215495303273201, + 0.0424698181450367, + 0.010492927394807339, + 0.009897850453853607, + 0.01762053556740284, + 0.01030134130269289, + 0.010290044359862804, + 0.01584717258810997, + 0.006349596660584211, + 0.009769397787749767, + 0.009914100170135498, + 0.007315630558878183, + 0.013428174890577793, + 0.00697331503033638, + 0.011643370613455772, + 0.005071415100246668, + 0.0037069288082420826, + 0.004355133976787329, + 0.003983555361628532, + 0.006672646850347519, + 0.004487594589591026, + 0.003958929795771837, + 0.0041369544342160225, + 0.0037786078173667192, + 0.010272136889398098, + 0.021546434611082077, + 0.010712722316384315, + 0.01507183164358139, + 0.008363030850887299, + 0.011522222310304642, + 0.013819397427141666, + 0.014501294121146202, + 0.0054874056950211525, + 0.005871656350791454, + 0.005518820136785507, + 0.0063523645512759686, + 0.004508977755904198, + 0.0043571931309998035, + 0.004923040047287941, + 0.00473412498831749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.0067402152344584465 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are broken spears, notched sword blades, and shattered shields visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are torn battle flags with faded crimson and gold patterns hanging from broken poles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by earthy brown, charcoal black, and cold metallic gray colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183909356594086 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183909356594086, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018183909356594086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14472515881061554, + 0.12464764714241028, + 0.042457565665245056, + 0.021333172917366028, + 0.08371558785438538, + 0.19195528328418732, + 0.028800390660762787, + 0.16161665320396423, + 0.09333235025405884, + 0.10741617530584335 + ], + "credit_map": [ + 0.02023572288453579, + 0.023582449182868004, + 0.015214999206364155, + 0.014652599580585957, + 0.02942216396331787, + 0.016350112855434418, + 0.015553164295852184, + 0.009713949635624886, + 0.01896069198846817, + 0.03153687343001366, + 0.014821144752204418, + 0.018185695633292198, + 0.01176505722105503, + 0.0073212687857449055, + 0.013297418132424355, + 0.008759495802223682, + 0.005861810874193907, + 0.007970791310071945, + 0.006103686057031155, + 0.006019198801368475, + 0.00423203082755208, + 0.0044221412390470505, + 0.0036838145460933447, + 0.00416409270837903, + 0.0024697347544133663, + 0.0028281405102461576, + 0.0027021400164812803, + 0.0033642486669123173, + 0.002354262862354517, + 0.002782519906759262, + 0.002432187320664525, + 0.002399938181042671, + 0.012111681513488293, + 0.013365511782467365, + 0.010703889653086662, + 0.019868921488523483, + 0.007414902560412884, + 0.005369577091187239, + 0.0076878247782588005, + 0.007193286903202534, + 0.035853784531354904, + 0.06660197675228119, + 0.01765197515487671, + 0.015436402522027493, + 0.019332697615027428, + 0.008264443837106228, + 0.014257450588047504, + 0.014556561596691608, + 0.0032447054982185364, + 0.006970118265599012, + 0.0059310682117938995, + 0.00290706567466259, + 0.0027234565932303667, + 0.0018392838537693024, + 0.002999015850946307, + 0.0021856760140508413, + 0.019236942753195763, + 0.013814558275043964, + 0.024458210915327072, + 0.021816661581397057, + 0.02951372042298317, + 0.015720484778285027, + 0.023500436916947365, + 0.013555620796978474, + 0.011302019469439983, + 0.010674651712179184, + 0.012957311235368252, + 0.016795167699456215, + 0.007279846351593733, + 0.00950257945805788, + 0.013214053586125374, + 0.011606722138822079, + 0.013892731629312038, + 0.03685030713677406, + 0.008413702249526978, + 0.006671735551208258, + 0.013214493170380592, + 0.007333557121455669, + 0.006477702874690294, + 0.014561940915882587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.018183909356594086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_0/index_tensor([4438, 4438])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_0/index_tensor([4438, 4438])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f8680ff36f31a56e9aa00f1f402901d2bd499e2c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_0/index_tensor([4438, 4438])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, top-down shot positioned directly overhead captures a mesmerizing scene with lifelike detail. A perfect blob of thick, golden-amber honey rests on a black, perfectly polished, and highly reflective countertop, resembling black obsidian. Soft, directed studio light from outside the camera casts brilliant, specular highlights on the honey's smooth surface, while its translucent bulk glows from within, revealing subtle internal ripples. Under its immense weight, the blob begins to slowly and deliberately deform, its leading edge creeping forward with the authentic, viscous flow. The honey lazily stretches, thins, and becomes more transparent at the advancing front, while the main body sags, forming a thick, syrupy trail. The dark countertop reflects the distorted, elongated reflections of the golden substance, enhancing the sense of depth and texture, creating a tranquil, focused, and almost hypnotic atmosphere that celebrates the simple, beautiful physics of fluid motion.", + "seed": 5525, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_0/index_4438_seed_5525_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06828624196350574, + "vlm_logprob": -0.06828624196350574, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.06828624196350574, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2981453537940979 + ], + "opposite_logprob_tokens": [ + -1.4231452941894531 + ], + "logprob_reward": -0.2981453537940979, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2981453537940979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16773664951324463, + 0.13773401081562042, + 0.04604487866163254, + 0.07305219024419785, + 0.06479812413454056, + 0.1139412373304367, + 0.04005598649382591, + 0.017007391899824142, + 0.09663979709148407, + 0.24298979341983795 + ], + "credit_map": [ + 0.01881515420973301, + 0.01620531640946865, + 0.025035224854946136, + 0.021584777161478996, + 0.01763675920665264, + 0.02450467273592949, + 0.019809626042842865, + 0.024145105853676796, + 0.016899418085813522, + 0.01477137766778469, + 0.01400036085397005, + 0.01864323951303959, + 0.021179985255002975, + 0.011272885836660862, + 0.02322453446686268, + 0.017742207273840904, + 0.007377099245786667, + 0.005339981522411108, + 0.004107512999325991, + 0.006159311626106501, + 0.004670954775065184, + 0.0037248781882226467, + 0.00655377795919776, + 0.008111364208161831, + 0.009802323766052723, + 0.010190736502408981, + 0.008192071691155434, + 0.008371610194444656, + 0.008313585072755814, + 0.00945947878062725, + 0.009065834805369377, + 0.009656544774770737, + 0.009204044006764889, + 0.006762318778783083, + 0.006718417629599571, + 0.007074363064020872, + 0.009417140856385231, + 0.008147108368575573, + 0.008122490718960762, + 0.009352246299386024, + 0.018077004700899124, + 0.008910847827792168, + 0.013587475754320621, + 0.017156559973955154, + 0.011005266569554806, + 0.013904299587011337, + 0.0142058115452528, + 0.017093973234295845, + 0.00574917858466506, + 0.006379306316375732, + 0.0034997067414224148, + 0.007234342861920595, + 0.003992382436990738, + 0.0035429259296506643, + 0.004209217615425587, + 0.005448923446238041, + 0.002495638793334365, + 0.001685021328739822, + 0.0017140737036243081, + 0.0028453164268285036, + 0.0015216409228742123, + 0.00264765415340662, + 0.001796285854652524, + 0.0023017614148557186, + 0.011493206024169922, + 0.009580923244357109, + 0.013643361628055573, + 0.014746759086847305, + 0.008918327279388905, + 0.012288441881537437, + 0.013347071595489979, + 0.012621707282960415, + 0.020828083157539368, + 0.019631192088127136, + 0.031076226383447647, + 0.029656896367669106, + 0.026601120829582214, + 0.0425223782658577, + 0.03399123623967171, + 0.038682661950588226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.2981453537940979 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016096990555524826 + ], + "opposite_logprob_tokens": [ + -4.141097068786621 + ], + "logprob_reward": -0.016096990555524826, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016096990555524826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35079124569892883, + 0.11822906881570816, + 0.034798260778188705, + 0.03989977017045021, + 0.026338940486311913, + 0.2400745153427124, + 0.022445090115070343, + 0.0126603739336133, + 0.06072717905044556, + 0.0940355733036995 + ], + "credit_map": [ + 0.033687230199575424, + 0.03651946038007736, + 0.04896216094493866, + 0.05904937535524368, + 0.03512878715991974, + 0.03858054429292679, + 0.04871135577559471, + 0.05015233904123306, + 0.01966739259660244, + 0.011860514059662819, + 0.010449083521962166, + 0.014436563476920128, + 0.01868954673409462, + 0.007218627724796534, + 0.019455572590231895, + 0.0164517629891634, + 0.005199659150093794, + 0.004860264249145985, + 0.0024300438817590475, + 0.00534187164157629, + 0.0029211777728050947, + 0.004018821287900209, + 0.0043247840367257595, + 0.005701638292521238, + 0.005799076985567808, + 0.004857675172388554, + 0.0052289157174527645, + 0.00470363674685359, + 0.0040032584220170975, + 0.005501518025994301, + 0.004947667941451073, + 0.004858023952692747, + 0.0033548176288604736, + 0.002544914837926626, + 0.0020618736743927, + 0.003149584634229541, + 0.0034774774685502052, + 0.0043014418333768845, + 0.002906650071963668, + 0.004542181268334389, + 0.03545784205198288, + 0.022151069715619087, + 0.02386733517050743, + 0.028558632358908653, + 0.028687922284007072, + 0.032883159816265106, + 0.03222951292991638, + 0.03623904287815094, + 0.003003063378855586, + 0.002412312664091587, + 0.0028128637932240963, + 0.003013658570125699, + 0.002733847824856639, + 0.0027697389014065266, + 0.002674589166417718, + 0.0030250167474150658, + 0.0018199619371443987, + 0.0011925019789487123, + 0.0013819059822708368, + 0.0019721195567399263, + 0.0008789181592874229, + 0.0018394713988527656, + 0.0020189243368804455, + 0.001556570059619844, + 0.006966030690819025, + 0.006857541389763355, + 0.0069879270158708096, + 0.009631272405385971, + 0.005957498215138912, + 0.006742351688444614, + 0.009206578135490417, + 0.008377978578209877, + 0.009423439390957355, + 0.008812658488750458, + 0.011714179068803787, + 0.011764656752347946, + 0.010982146486639977, + 0.014615895226597786, + 0.01233486458659172, + 0.014387736096978188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.016096990555524826 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792149998247623 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.009792149998247623, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792149998247623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4317030608654022, + 0.0730816200375557, + 0.034961946308612823, + 0.07901762425899506, + 0.10974659025669098, + 0.0488605722784996, + 0.0253624077886343, + 0.017963068559765816, + 0.06266247481107712, + 0.11664056777954102 + ], + "credit_map": [ + 0.0416666641831398, + 0.03449108079075813, + 0.06646376103162766, + 0.049880724400281906, + 0.05058228224515915, + 0.05745550990104675, + 0.06912129372358322, + 0.06204177439212799, + 0.007765492424368858, + 0.00591424060985446, + 0.010174400173127651, + 0.008681570179760456, + 0.009047301486134529, + 0.004736283794045448, + 0.015438860282301903, + 0.011323479935526848, + 0.0053947498090565205, + 0.0035024655517190695, + 0.0032284187618643045, + 0.003623559605330229, + 0.004096690099686384, + 0.006332231219857931, + 0.003429383272305131, + 0.005354452412575483, + 0.008791974745690823, + 0.011205427348613739, + 0.008862467482686043, + 0.009353641420602798, + 0.008793915621936321, + 0.012613986618816853, + 0.009097978472709656, + 0.010298238135874271, + 0.013845628127455711, + 0.011086140759289265, + 0.009820292703807354, + 0.01186094619333744, + 0.01803594082593918, + 0.017656736075878143, + 0.011166809126734734, + 0.0162741020321846, + 0.006456951145082712, + 0.006280756555497646, + 0.005502166226506233, + 0.008364876732230186, + 0.004541848786175251, + 0.005734228994697332, + 0.004978557117283344, + 0.007001187652349472, + 0.0033355557825416327, + 0.0028210768941789865, + 0.00234358012676239, + 0.003919769078493118, + 0.0037802245933562517, + 0.003127280157059431, + 0.002838159678503871, + 0.003196762641891837, + 0.0021900178398936987, + 0.0015311703318729997, + 0.0018402448622509837, + 0.002438188064843416, + 0.002056871773675084, + 0.0034808816853910685, + 0.001907173078507185, + 0.0025185232516378164, + 0.007096433080732822, + 0.0054253097623586655, + 0.005454640835523605, + 0.006633815821260214, + 0.005147140938788652, + 0.015516464598476887, + 0.0067488583736121655, + 0.010639815591275692, + 0.011124846525490284, + 0.011346651241183281, + 0.015145513229072094, + 0.013528939336538315, + 0.012860216200351715, + 0.018684007227420807, + 0.015682073310017586, + 0.018268326297402382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.009792149998247623 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201176352798939 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014201176352798939, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201176352798939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22973178327083588, + 0.10743270069360733, + 0.09536884725093842, + 0.04616944119334221, + 0.04577562212944031, + 0.0829293429851532, + 0.04744362831115723, + 0.03682851418852806, + 0.11796043068170547, + 0.19035959243774414 + ], + "credit_map": [ + 0.03096819296479225, + 0.023999178782105446, + 0.032712336629629135, + 0.03338390588760376, + 0.020587844774127007, + 0.019381318241357803, + 0.03371516242623329, + 0.03498387336730957, + 0.012971987016499043, + 0.012271642684936523, + 0.0109851760789752, + 0.015896160155534744, + 0.012759526260197163, + 0.009042126126587391, + 0.017917441204190254, + 0.015588649548590183, + 0.018072407692670822, + 0.010943155735731125, + 0.008695848286151886, + 0.012329412624239922, + 0.00817765761166811, + 0.009847462177276611, + 0.009673921391367912, + 0.0176289863884449, + 0.005062086042016745, + 0.005484489258378744, + 0.0050149839371442795, + 0.005344146862626076, + 0.00567352632060647, + 0.007604818791151047, + 0.004936420824378729, + 0.00704897241666913, + 0.005965844262391329, + 0.004952582065016031, + 0.004892490804195404, + 0.005403613205999136, + 0.005710430908948183, + 0.006803698837757111, + 0.005412353668361902, + 0.00663461210206151, + 0.011080491356551647, + 0.0066599189303815365, + 0.01156198512762785, + 0.011475599370896816, + 0.009212660603225231, + 0.010360132902860641, + 0.010100801475346088, + 0.012477762065827847, + 0.0061223674565553665, + 0.00440334714949131, + 0.0043254815973341465, + 0.009520214051008224, + 0.004951547831296921, + 0.00617833249270916, + 0.005998862907290459, + 0.005943481810390949, + 0.005520192440599203, + 0.005337610375136137, + 0.0035876098554581404, + 0.005813260562717915, + 0.002639530226588249, + 0.005428105592727661, + 0.004276735242456198, + 0.004225472453981638, + 0.010346186347305775, + 0.013962720520794392, + 0.012372537516057491, + 0.017922576516866684, + 0.011434551328420639, + 0.01935071498155594, + 0.016692453995347023, + 0.015878701582551003, + 0.019282950088381767, + 0.022281464189291, + 0.020484810695052147, + 0.025461414828896523, + 0.020847782492637634, + 0.028775442391633987, + 0.024236049503087997, + 0.02898968756198883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.014201176352798939 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195539116859436 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.003195539116859436, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003195539116859436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19167156517505646, + 0.07967141270637512, + 0.04930247366428375, + 0.03818810358643532, + 0.03818007558584213, + 0.2280881255865097, + 0.02216513082385063, + 0.02058657631278038, + 0.05069645866751671, + 0.281450092792511 + ], + "credit_map": [ + 0.021828925237059593, + 0.02137933485209942, + 0.025477169081568718, + 0.024582257494330406, + 0.022846383973956108, + 0.021858718246221542, + 0.02714860625565052, + 0.026550181210041046, + 0.010966078378260136, + 0.008512484841048717, + 0.0074829598888754845, + 0.010780530981719494, + 0.013246925547719002, + 0.0053980122320353985, + 0.012628279626369476, + 0.010656134225428104, + 0.007422326598316431, + 0.003963099326938391, + 0.004234240390360355, + 0.0047345287166535854, + 0.00485125370323658, + 0.009621231816709042, + 0.005357223097234964, + 0.009118568152189255, + 0.006505304481834173, + 0.003580706426873803, + 0.003693304257467389, + 0.004014177713543177, + 0.004234102088958025, + 0.0071251047775149345, + 0.0036203220952302217, + 0.005415080580860376, + 0.004830514080822468, + 0.004243853036314249, + 0.0030765177216380835, + 0.003698460292071104, + 0.005233626812696457, + 0.0064704883843660355, + 0.004319815430790186, + 0.006306799594312906, + 0.034231483936309814, + 0.019742583855986595, + 0.025527289137244225, + 0.02944689430296421, + 0.02708505280315876, + 0.028316209092736244, + 0.030513955280184746, + 0.0332246758043766, + 0.002695538802072406, + 0.0033296383917331696, + 0.0020773897413164377, + 0.003923011478036642, + 0.0020951239857822657, + 0.0026570498012006283, + 0.002128457184880972, + 0.0032589216716587543, + 0.003001393284648657, + 0.0015458929119631648, + 0.002521038521081209, + 0.002810630016028881, + 0.001558744115754962, + 0.003908855374902487, + 0.002667608205229044, + 0.002572414930909872, + 0.005815950222313404, + 0.005447729956358671, + 0.007379638496786356, + 0.007693129125982523, + 0.005034657660871744, + 0.005697434302419424, + 0.00724446726962924, + 0.006383453495800495, + 0.02578580379486084, + 0.026128225028514862, + 0.03628738597035408, + 0.034886207431554794, + 0.029662586748600006, + 0.04615484178066254, + 0.03841618448495865, + 0.04412884637713432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.003195539116859436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, top-down shot positioned directly overhead captures a mesmerizing scene with lifelike detail. A perfect blob of thick, golden-amber honey rests on a black, perfectly polished, and highly reflective countertop, resembling black obsidian. Soft, directed studio light from outside the camera casts brilliant, specular highlights on the honey's smooth surface, while its translucent bulk glows from within, revealing subtle internal ripples. Under its immense weight, the blob begins to slowly and deliberately deform, its leading edge creeping forward with the authentic, viscous flow. The honey lazily stretches, thins, and becomes more transparent at the advancing front, while the main body sags, forming a thick, syrupy trail. The dark countertop reflects the distorted, elongated reflections of the golden substance, enhancing the sense of depth and texture, creating a tranquil, focused, and almost hypnotic atmosphere that celebrates the simple, beautiful physics of fluid motion.", + "seed": 5526, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_0/index_4438_seed_5526_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019411072274670003, + "vlm_logprob": -0.019411072274670003, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.019411072274670003, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06349442899227142 + ], + "opposite_logprob_tokens": [ + -3.1884944438934326 + ], + "logprob_reward": -0.06349442899227142, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06349442899227142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08360522985458374, + 0.06820034980773926, + 0.26874637603759766, + 0.055319104343652725, + 0.04708376154303551, + 0.15093101561069489, + 0.274685800075531, + 0.01460043340921402, + 0.017367854714393616, + 0.019460106268525124 + ], + "credit_map": [ + 0.014066837728023529, + 0.008247928693890572, + 0.012966185808181763, + 0.007730924990028143, + 0.015849262475967407, + 0.0065727331675589085, + 0.006678085308521986, + 0.011493266560137272, + 0.013317395932972431, + 0.0053594475612044334, + 0.014670956879854202, + 0.005756942555308342, + 0.005419477354735136, + 0.00927943829447031, + 0.0061681438237428665, + 0.008228546008467674, + 0.020057063549757004, + 0.009383910335600376, + 0.10315025597810745, + 0.006013700272887945, + 0.006231693550944328, + 0.10366155952215195, + 0.0047992803156375885, + 0.0154489167034626, + 0.008869320154190063, + 0.008933855220675468, + 0.007615977898240089, + 0.006677727214992046, + 0.004078049678355455, + 0.00771494023501873, + 0.005845101550221443, + 0.005584133788943291, + 0.012301993556320667, + 0.0044201435521245, + 0.007550852373242378, + 0.003287743777036667, + 0.0029222688172012568, + 0.007497662212699652, + 0.003060808638110757, + 0.006042285356670618, + 0.04228755459189415, + 0.012917568907141685, + 0.017128167673945427, + 0.013616899959743023, + 0.009216411970555782, + 0.01838569901883602, + 0.008298731409013271, + 0.029079971835017204, + 0.04778766632080078, + 0.00873027928173542, + 0.048380494117736816, + 0.0059168837033212185, + 0.008046912029385567, + 0.12076617032289505, + 0.00837645772844553, + 0.02668091282248497, + 0.0023094192147254944, + 0.001127265626564622, + 0.002646154025569558, + 0.0025046500377357006, + 0.000995239824987948, + 0.001315173925831914, + 0.0015237489715218544, + 0.0021787816658616066, + 0.002353247255086899, + 0.0014352081343531609, + 0.0031452691182494164, + 0.0017995026428252459, + 0.0013041102793067694, + 0.0021248902194201946, + 0.0014625885523855686, + 0.0037430382799357176, + 0.0019930992275476456, + 0.0022055478766560555, + 0.0027365456335246563, + 0.002925034612417221, + 0.0022166066337376833, + 0.0022132762242108583, + 0.0019249801989644766, + 0.0032450174912810326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.06349442899227142 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011097377166152 + ], + "opposite_logprob_tokens": [ + -4.511097431182861 + ], + "logprob_reward": -0.011097377166152, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011097377166152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07558953016996384, + 0.04825643450021744, + 0.09031592309474945, + 0.022783296182751656, + 0.04419619217514992, + 0.0719546377658844, + 0.5244323015213013, + 0.03157894313335419, + 0.04669518023729324, + 0.04419759660959244 + ], + "credit_map": [ + 0.013843382708728313, + 0.006536028813570738, + 0.011156182736158371, + 0.005047122482210398, + 0.013403009623289108, + 0.00749894930049777, + 0.0057375566102564335, + 0.012367300689220428, + 0.009003514423966408, + 0.005572207737714052, + 0.008025824092328548, + 0.0040480210445821285, + 0.006635033525526524, + 0.004765934776514769, + 0.00407456373795867, + 0.006131331901997328, + 0.008080016821622849, + 0.004939651116728783, + 0.03044719062745571, + 0.002914669457823038, + 0.005498334765434265, + 0.029213257133960724, + 0.0036876783706247807, + 0.00553512666374445, + 0.004297198727726936, + 0.0025953424628823996, + 0.004032652359455824, + 0.002197183668613434, + 0.0013558960054069757, + 0.003551722038537264, + 0.001960206311196089, + 0.002793095773085952, + 0.010687602683901787, + 0.004312505945563316, + 0.006435886491090059, + 0.004494244698435068, + 0.002485501579940319, + 0.007174247410148382, + 0.0036047345492988825, + 0.005001470446586609, + 0.012148316949605942, + 0.007451266050338745, + 0.008187969215214252, + 0.010556432418525219, + 0.005139281041920185, + 0.011809022165834904, + 0.005944252945482731, + 0.010718092322349548, + 0.10169654339551926, + 0.01827320270240307, + 0.09024589508771896, + 0.011709026992321014, + 0.01804148219525814, + 0.22368207573890686, + 0.017357943579554558, + 0.04342612996697426, + 0.007060045376420021, + 0.003251878544688225, + 0.004570174030959606, + 0.0034386201296001673, + 0.0019758983980864286, + 0.004632242489606142, + 0.002589845797047019, + 0.004060239065438509, + 0.006518173962831497, + 0.0022653716150671244, + 0.012587777338922024, + 0.0036828338634222746, + 0.0015933278482407331, + 0.004426447208970785, + 0.0027347782161086798, + 0.0128864711150527, + 0.005861035548150539, + 0.0056927986443042755, + 0.007077659014612436, + 0.00785259623080492, + 0.0037529761902987957, + 0.004508485551923513, + 0.0042055025696754456, + 0.005246540531516075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.011097377166152 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646931774914265 + ], + "opposite_logprob_tokens": [ + -5.3796467781066895 + ], + "logprob_reward": -0.004646931774914265, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646931774914265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1063523069024086, + 0.06363678723573685, + 0.6043318510055542, + 0.0167153999209404, + 0.028640715405344963, + 0.03017769567668438, + 0.10194507241249084, + 0.005290994420647621, + 0.019970882683992386, + 0.02293824777007103 + ], + "credit_map": [ + 0.021512730047106743, + 0.010304533876478672, + 0.017102155834436417, + 0.008743036538362503, + 0.013061167672276497, + 0.009689977392554283, + 0.0070466031320393085, + 0.018892105668783188, + 0.012348427437245846, + 0.005320257507264614, + 0.014986272901296616, + 0.004409549292176962, + 0.0055345031432807446, + 0.008668764494359493, + 0.004627618473023176, + 0.007741391193121672, + 0.04381297901272774, + 0.021983906626701355, + 0.22927114367485046, + 0.014498674310743809, + 0.014637266285717487, + 0.23242443799972534, + 0.011240709573030472, + 0.03646279498934746, + 0.0029842748772352934, + 0.0023973339702934027, + 0.0028690816834568977, + 0.0018907986814156175, + 0.0014540639240294695, + 0.0021204689983278513, + 0.0014551919884979725, + 0.001544185564853251, + 0.006395319476723671, + 0.0024572238326072693, + 0.004543308634310961, + 0.0030616912990808487, + 0.0015604088548570871, + 0.004951074253767729, + 0.0023222484160214663, + 0.003349443431943655, + 0.009669375605881214, + 0.0015319706872105598, + 0.0038381642661988735, + 0.001491180737502873, + 0.001467996509745717, + 0.0038519378285855055, + 0.0018441206775605679, + 0.006482947152107954, + 0.019776832312345505, + 0.0031018282752484083, + 0.017961211502552032, + 0.0019914780277758837, + 0.0030575969722121954, + 0.04435853287577629, + 0.0031936673913151026, + 0.008503926917910576, + 0.0009167062235064805, + 0.0004459407937247306, + 0.0008889743476174772, + 0.0007655284134671092, + 0.0003265741397626698, + 0.0006219453061930835, + 0.00046364031732082367, + 0.0008616847917437553, + 0.003089943900704384, + 0.0015572477132081985, + 0.003484036773443222, + 0.0017602622974663973, + 0.0016030902042984962, + 0.003192015690729022, + 0.0016131020383909345, + 0.0036711846478283405, + 0.003650916740298271, + 0.0026205109898000956, + 0.003538625780493021, + 0.0035040813963860273, + 0.0019394726259633899, + 0.0023118904791772366, + 0.002433529356494546, + 0.002939219120889902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.004646931774914265 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196004718542099 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014196004718542099, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196004718542099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1591491848230362, + 0.06623899191617966, + 0.20478200912475586, + 0.10572375357151031, + 0.07267869263887405, + 0.08439687639474869, + 0.12431054562330246, + 0.05205168202519417, + 0.03155868873000145, + 0.09910950809717178 + ], + "credit_map": [ + 0.02672937698662281, + 0.013457067310810089, + 0.025661161169409752, + 0.012660031206905842, + 0.02932795137166977, + 0.013917258009314537, + 0.010915957391262054, + 0.026480378583073616, + 0.014982515946030617, + 0.006002161186188459, + 0.013045215979218483, + 0.005583054386079311, + 0.005653801839798689, + 0.0072938683442771435, + 0.006373469717800617, + 0.0073049128986895084, + 0.015567559748888016, + 0.007657054346054792, + 0.07735457271337509, + 0.004825340583920479, + 0.005708666052669287, + 0.07654920220375061, + 0.004006783943623304, + 0.013112860731780529, + 0.02712111361324787, + 0.0058555579744279385, + 0.030276373028755188, + 0.006056907586753368, + 0.0070841652341187, + 0.013836299069225788, + 0.0059978230856359005, + 0.009495520032942295, + 0.016576377674937248, + 0.0062391790561378, + 0.011759286746382713, + 0.006659971550107002, + 0.004319354426115751, + 0.01134968176484108, + 0.005889447405934334, + 0.00988539308309555, + 0.016948556527495384, + 0.008922702632844448, + 0.008461983874440193, + 0.007760506588965654, + 0.007787470705807209, + 0.011708669364452362, + 0.006791532970964909, + 0.016015464439988136, + 0.011757370084524155, + 0.006483639124780893, + 0.015882425010204315, + 0.005815497133880854, + 0.007077978458255529, + 0.042561404407024384, + 0.005804309155791998, + 0.028927939012646675, + 0.005969239864498377, + 0.0032879437785595655, + 0.006464418955147266, + 0.004451468586921692, + 0.002159244380891323, + 0.01386571116745472, + 0.0028506768867373466, + 0.013002978637814522, + 0.0041068075224757195, + 0.0029828681144863367, + 0.0054817283526062965, + 0.004593901801854372, + 0.0027652347926050425, + 0.003913652617484331, + 0.0027503222227096558, + 0.00496417423710227, + 0.011119053699076176, + 0.009298969991505146, + 0.014194673858582973, + 0.011031495407223701, + 0.011298304423689842, + 0.014888100326061249, + 0.010324693284928799, + 0.01695423014461994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.014196004718542099 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036206187214702368 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036206187214702368, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036206187214702368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058936115354299545, + 0.16164326667785645, + 0.16244423389434814, + 0.07165814191102982, + 0.08017434179782867, + 0.09543095529079437, + 0.16546930372714996, + 0.06047515943646431, + 0.04309011623263359, + 0.10067836195230484 + ], + "credit_map": [ + 0.010889209806919098, + 0.005569943226873875, + 0.008764268830418587, + 0.00450439378619194, + 0.010925756767392159, + 0.004862704314291477, + 0.005047110840678215, + 0.008372731506824493, + 0.03001776896417141, + 0.013640256598591805, + 0.033281099051237106, + 0.013084657490253448, + 0.016510315239429474, + 0.02092461846768856, + 0.01682722009718418, + 0.01735733635723591, + 0.012832175008952618, + 0.00766752427443862, + 0.05821898952126503, + 0.004748588893562555, + 0.005399596411734819, + 0.05930623039603233, + 0.004549684468656778, + 0.009721439331769943, + 0.015732906758785248, + 0.006521477829664946, + 0.015649715438485146, + 0.0056998515501618385, + 0.0040579005144536495, + 0.010354764759540558, + 0.0060352180153131485, + 0.00760630751028657, + 0.01694629155099392, + 0.006945491768419743, + 0.012670221738517284, + 0.007890990935266018, + 0.007322689518332481, + 0.01307825930416584, + 0.006790263578295708, + 0.008530129678547382, + 0.022052738815546036, + 0.00713437981903553, + 0.011686191894114017, + 0.007960837334394455, + 0.006780553143471479, + 0.017216436564922333, + 0.007155787665396929, + 0.01544403750449419, + 0.01564478687942028, + 0.01069699414074421, + 0.02563558705151081, + 0.00668897433206439, + 0.0060487729497253895, + 0.06138061732053757, + 0.005746642593294382, + 0.033626917749643326, + 0.009918497875332832, + 0.0055137150920927525, + 0.010291735641658306, + 0.007298166863620281, + 0.003579579759389162, + 0.006495280656963587, + 0.003636979265138507, + 0.013741206377744675, + 0.007372352294623852, + 0.003845543134957552, + 0.006965422537177801, + 0.005839736666530371, + 0.002742521930485964, + 0.005902472417801619, + 0.003789637703448534, + 0.0066324262879788876, + 0.015145131386816502, + 0.012671156786382198, + 0.018333904445171356, + 0.0187720637768507, + 0.007970532402396202, + 0.009047293104231358, + 0.009725592099130154, + 0.009012687020003796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.0036206187214702368 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, top-down shot positioned directly overhead captures a mesmerizing scene with lifelike detail. A perfect blob of thick, golden-amber honey rests on a black, perfectly polished, and highly reflective countertop, resembling black obsidian. Soft, directed studio light from outside the camera casts brilliant, specular highlights on the honey's smooth surface, while its translucent bulk glows from within, revealing subtle internal ripples. Under its immense weight, the blob begins to slowly and deliberately deform, its leading edge creeping forward with the authentic, viscous flow. The honey lazily stretches, thins, and becomes more transparent at the advancing front, while the main body sags, forming a thick, syrupy trail. The dark countertop reflects the distorted, elongated reflections of the golden substance, enhancing the sense of depth and texture, creating a tranquil, focused, and almost hypnotic atmosphere that celebrates the simple, beautiful physics of fluid motion.", + "seed": 5527, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_0/index_4438_seed_5527_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.022336539812386036, + "vlm_logprob": -0.022336539812386036, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.022336539812386036, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05612235888838768 + ], + "opposite_logprob_tokens": [ + -3.306122303009033 + ], + "logprob_reward": -0.05612235888838768, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05612235888838768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23177281022071838, + 0.05397672578692436, + 0.05550365522503853, + 0.06680544465780258, + 0.17064698040485382, + 0.05207620561122894, + 0.12267925590276718, + 0.018077440559864044, + 0.03260263428092003, + 0.19585886597633362 + ], + "credit_map": [ + 0.04366562142968178, + 0.026188194751739502, + 0.04636127129197121, + 0.01953883282840252, + 0.03467579931020737, + 0.01981162466108799, + 0.026094194501638412, + 0.015437254682183266, + 0.011220287531614304, + 0.005959851201623678, + 0.008875933475792408, + 0.006244645453989506, + 0.007834043353796005, + 0.005251314956694841, + 0.0063461121171712875, + 0.002244536532089114, + 0.010842357762157917, + 0.004784764256328344, + 0.008689356967806816, + 0.004918931983411312, + 0.009893360547721386, + 0.005671591032296419, + 0.007379482034593821, + 0.003323809476569295, + 0.012664517387747765, + 0.007884052582085133, + 0.011473001912236214, + 0.0054524741135537624, + 0.010753457434475422, + 0.006354474928230047, + 0.008117827586829662, + 0.004105635453015566, + 0.03441172093153, + 0.014544785022735596, + 0.02646995149552822, + 0.016384491696953773, + 0.02836894430220127, + 0.021380290389060974, + 0.01632455177605152, + 0.012762225233018398, + 0.011770771816372871, + 0.006992043927311897, + 0.008658542297780514, + 0.0041253576055169106, + 0.007709318771958351, + 0.004423671867698431, + 0.004802338778972626, + 0.003594152396544814, + 0.022658074274659157, + 0.02106434293091297, + 0.020578835159540176, + 0.010178826749324799, + 0.015112683176994324, + 0.011957095935940742, + 0.011645086109638214, + 0.009484312497079372, + 0.002865143120288849, + 0.0032170589547604322, + 0.0026034745387732983, + 0.0017323867650702596, + 0.0034886968787759542, + 0.0018013332737609744, + 0.0016005313955247402, + 0.0007688137120567262, + 0.007014373317360878, + 0.004073896910995245, + 0.0064187669195234776, + 0.002431499771773815, + 0.0029230036307126284, + 0.003385083517059684, + 0.004359393380582333, + 0.0019966154359281063, + 0.08673567324876785, + 0.008392293937504292, + 0.05845613032579422, + 0.005866943392902613, + 0.006389036774635315, + 0.012514135800302029, + 0.00672485027462244, + 0.010779788717627525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.05612235888838768 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370810344815254 + ], + "opposite_logprob_tokens": [ + -3.6513707637786865 + ], + "logprob_reward": -0.026370810344815254, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370810344815254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12918174266815186, + 0.1130545437335968, + 0.11167792975902557, + 0.08501047641038895, + 0.12197741866111755, + 0.14118164777755737, + 0.025295689702033997, + 0.02576497569680214, + 0.056222520768642426, + 0.19063305854797363 + ], + "credit_map": [ + 0.021218135952949524, + 0.013048301450908184, + 0.018086647614836693, + 0.013034320436418056, + 0.020179035142064095, + 0.01453301589936018, + 0.02130838856101036, + 0.007773892488330603, + 0.018964989110827446, + 0.013073241338133812, + 0.019288888201117516, + 0.011107334867119789, + 0.016671154648065567, + 0.013910703361034393, + 0.013606052845716476, + 0.006432176101952791, + 0.019974038004875183, + 0.013742035254836082, + 0.017345387488603592, + 0.010773209854960442, + 0.015311654657125473, + 0.011714374646544456, + 0.012914863415062428, + 0.009902359917759895, + 0.01426311768591404, + 0.01096232607960701, + 0.014629394747316837, + 0.00991868507117033, + 0.010795588605105877, + 0.01129048690199852, + 0.007958510890603065, + 0.005192368291318417, + 0.021836020052433014, + 0.009627901948988438, + 0.020119640976190567, + 0.01156344823539257, + 0.021919602528214455, + 0.01568594202399254, + 0.012572574429214, + 0.008652287535369396, + 0.024323824793100357, + 0.023305824026465416, + 0.021464040502905846, + 0.014968999661505222, + 0.022642672061920166, + 0.013478623703122139, + 0.01159321516752243, + 0.009404458105564117, + 0.006069738417863846, + 0.0025447981897741556, + 0.004447903484106064, + 0.0021806268487125635, + 0.002627141308039427, + 0.0037631008308380842, + 0.0017965262522920966, + 0.0018658561166375875, + 0.00668608583509922, + 0.00283325114287436, + 0.00466523040086031, + 0.0017910476308315992, + 0.003739080624654889, + 0.0025891545228660107, + 0.001627556630410254, + 0.0018335686763748527, + 0.014550724998116493, + 0.004349603317677975, + 0.017677875235676765, + 0.0042176321148872375, + 0.003350592451170087, + 0.0061765918508172035, + 0.0029993068892508745, + 0.0029001960065215826, + 0.06459047645330429, + 0.015923654660582542, + 0.04647834226489067, + 0.011301703751087189, + 0.014462856575846672, + 0.013147660531103611, + 0.011358614079654217, + 0.013369759544730186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.026370810344815254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551246210932732 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.012551246210932732, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551246210932732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3995719254016876, + 0.06297323852777481, + 0.10147672891616821, + 0.08108481764793396, + 0.05717771500349045, + 0.12515661120414734, + 0.062093768268823624, + 0.02657686546444893, + 0.019319437444210052, + 0.06456885486841202 + ], + "credit_map": [ + 0.07966116815805435, + 0.02889525704085827, + 0.06991812586784363, + 0.02872590534389019, + 0.04902280122041702, + 0.03507436439394951, + 0.0783161148428917, + 0.029958199709653854, + 0.010767179541289806, + 0.009047729894518852, + 0.008453291840851307, + 0.005776511505246162, + 0.011879161931574345, + 0.005278715398162603, + 0.006592741701751947, + 0.005177905783057213, + 0.016586605459451675, + 0.01138990931212902, + 0.015532254241406918, + 0.009839843958616257, + 0.018295804038643837, + 0.011853226460516453, + 0.011608037166297436, + 0.006371049676090479, + 0.013487143442034721, + 0.011434147134423256, + 0.014412353746592999, + 0.008808820508420467, + 0.008380324579775333, + 0.011720028705894947, + 0.0069839367642998695, + 0.0058580548502504826, + 0.010489325039088726, + 0.006457126699388027, + 0.01009003259241581, + 0.006997823715209961, + 0.009589357301592827, + 0.004547677002847195, + 0.005921024363487959, + 0.0030853464268147945, + 0.02478291280567646, + 0.016069088131189346, + 0.020275458693504333, + 0.014969179406762123, + 0.01725625991821289, + 0.013881918974220753, + 0.009051026776432991, + 0.008870773017406464, + 0.010605166666209698, + 0.012353447265923023, + 0.010060679167509079, + 0.005147835705429316, + 0.007478960324078798, + 0.005582417361438274, + 0.006113158538937569, + 0.004752103704959154, + 0.0050292424857616425, + 0.0026370834093540907, + 0.004430110566318035, + 0.002397662028670311, + 0.00537756085395813, + 0.0028010390233248472, + 0.0022193382028490305, + 0.0016848290106281638, + 0.0045363144017755985, + 0.0013169513549655676, + 0.006355503108352423, + 0.0010461027268320322, + 0.0012523983605206013, + 0.0024348825681954622, + 0.0013609224697574973, + 0.0010163614060729742, + 0.021009869873523712, + 0.004755460657179356, + 0.014220592565834522, + 0.0037557114847004414, + 0.0047582900151610374, + 0.007502728141844273, + 0.004816994536668062, + 0.003749205730855465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.012551246210932732 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541239149868488 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541239149868488, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541239149868488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10931342095136642, + 0.08473239094018936, + 0.06887117028236389, + 0.04299641028046608, + 0.19816069304943085, + 0.07353446632623672, + 0.02697835862636566, + 0.019270645454525948, + 0.02867273986339569, + 0.3474697172641754 + ], + "credit_map": [ + 0.018498627468943596, + 0.009003674611449242, + 0.019197627902030945, + 0.007901718840003014, + 0.01485986728221178, + 0.010883653536438942, + 0.021342547610402107, + 0.007625696249306202, + 0.013184483163058758, + 0.00966079905629158, + 0.010989350266754627, + 0.00757494755089283, + 0.017049899324774742, + 0.009214358404278755, + 0.0101312305778265, + 0.00692732585594058, + 0.011451374739408493, + 0.008663715794682503, + 0.010164839215576649, + 0.007633318193256855, + 0.012332417070865631, + 0.0078815259039402, + 0.007497356738895178, + 0.003246618900448084, + 0.007074487861245871, + 0.0049545359797775745, + 0.007255061995238066, + 0.00525016849860549, + 0.006196504924446344, + 0.005230674985796213, + 0.004022639710456133, + 0.003012332832440734, + 0.04169687256217003, + 0.020505892112851143, + 0.0359005331993103, + 0.014989608898758888, + 0.03265569731593132, + 0.018770158290863037, + 0.01791621744632721, + 0.015725696459412575, + 0.013587791472673416, + 0.008727977983653545, + 0.011063464917242527, + 0.007216352038085461, + 0.013230646960437298, + 0.0069832210429012775, + 0.007074889726936817, + 0.005650114733725786, + 0.004082954488694668, + 0.004976126365363598, + 0.004220686387270689, + 0.003270629560574889, + 0.003996318206191063, + 0.0025148356799036264, + 0.0023665328044444323, + 0.0015502728056162596, + 0.0042035626247525215, + 0.00149526447057724, + 0.0044173491187393665, + 0.0015887641347944736, + 0.0025819079019129276, + 0.001974061131477356, + 0.0016314118402078748, + 0.0013783230679109693, + 0.008522321470081806, + 0.0016105732647702098, + 0.00805646926164627, + 0.0013631335459649563, + 0.0029624830931425095, + 0.0020876543130725622, + 0.002043002052232623, + 0.0020270992536097765, + 0.1591378152370453, + 0.010332770645618439, + 0.10636294633150101, + 0.006650895345956087, + 0.007156210485845804, + 0.02661227062344551, + 0.010041872970759869, + 0.02117490954697132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.012541239149868488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097044467926025 + ], + "opposite_logprob_tokens": [ + -5.504096984863281 + ], + "logprob_reward": -0.004097044467926025, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097044467926025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23188361525535583, + 0.07585466653108597, + 0.16450923681259155, + 0.05011170357465744, + 0.08136069029569626, + 0.030515853315591812, + 0.05761202424764633, + 0.08921930938959122, + 0.034476347267627716, + 0.18445658683776855 + ], + "credit_map": [ + 0.04893883317708969, + 0.023839576169848442, + 0.03641451150178909, + 0.02078302577137947, + 0.033341724425554276, + 0.022306954488158226, + 0.030913975089788437, + 0.015345006249845028, + 0.015864327549934387, + 0.0062485141679644585, + 0.013003570027649403, + 0.004932656418532133, + 0.010705815628170967, + 0.006600181572139263, + 0.008056910708546638, + 0.010442689061164856, + 0.02815847285091877, + 0.019292796030640602, + 0.026634778827428818, + 0.018573110923171043, + 0.022761277854442596, + 0.021145684644579887, + 0.016420383006334305, + 0.011522726155817509, + 0.0077674658969044685, + 0.006563832517713308, + 0.008417366072535515, + 0.005750145297497511, + 0.006742163561284542, + 0.005996938329190016, + 0.00555249722674489, + 0.0033212897833436728, + 0.014042247086763382, + 0.0073773907497525215, + 0.0115429712459445, + 0.005984498653560877, + 0.01965676061809063, + 0.007908279076218605, + 0.010353846475481987, + 0.00449469406157732, + 0.005096727050840855, + 0.003700221423059702, + 0.004735512193292379, + 0.004119386896491051, + 0.004433221183717251, + 0.0032950222957879305, + 0.0030146059580147266, + 0.00212115328758955, + 0.010344157926738262, + 0.008732154034078121, + 0.011102087795734406, + 0.005088194739073515, + 0.006719613913446665, + 0.006375819444656372, + 0.005084082949906588, + 0.004165909718722105, + 0.015232248231768608, + 0.013922951184213161, + 0.016662726178765297, + 0.009074053727090359, + 0.00802548322826624, + 0.010583384893834591, + 0.0077920109033584595, + 0.007926448248326778, + 0.0073215109296143055, + 0.004226771183311939, + 0.0074310991913080215, + 0.002364144427701831, + 0.003227602457627654, + 0.0038385987281799316, + 0.003126533003523946, + 0.0029400838539004326, + 0.0702395811676979, + 0.009074164554476738, + 0.04836311563849449, + 0.00813521258533001, + 0.012100839987397194, + 0.016520818695425987, + 0.010180043987929821, + 0.009842793457210064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.004097044467926025 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, top-down shot positioned directly overhead captures a mesmerizing scene with lifelike detail. A perfect blob of thick, golden-amber honey rests on a black, perfectly polished, and highly reflective countertop, resembling black obsidian. Soft, directed studio light from outside the camera casts brilliant, specular highlights on the honey's smooth surface, while its translucent bulk glows from within, revealing subtle internal ripples. Under its immense weight, the blob begins to slowly and deliberately deform, its leading edge creeping forward with the authentic, viscous flow. The honey lazily stretches, thins, and becomes more transparent at the advancing front, while the main body sags, forming a thick, syrupy trail. The dark countertop reflects the distorted, elongated reflections of the golden substance, enhancing the sense of depth and texture, creating a tranquil, focused, and almost hypnotic atmosphere that celebrates the simple, beautiful physics of fluid motion.", + "seed": 5528, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_0/index_4438_seed_5528_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04252835912629962, + "vlm_logprob": -0.04252835912629962, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.04252835912629962, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15924285352230072 + ], + "opposite_logprob_tokens": [ + -2.159242868423462 + ], + "logprob_reward": -0.15924285352230072, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.15924285352230072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12350331991910934, + 0.019452111795544624, + 0.021000590175390244, + 0.07006007432937622, + 0.02727488800883293, + 0.2898179292678833, + 0.15094318985939026, + 0.08612669259309769, + 0.06619735807180405, + 0.14562389254570007 + ], + "credit_map": [ + 0.016050167381763458, + 0.013006162829697132, + 0.029953425750136375, + 0.015312207862734795, + 0.01585087925195694, + 0.009802534244954586, + 0.012267031706869602, + 0.011260905303061008, + 0.00210447795689106, + 0.0020692022517323494, + 0.003846373874694109, + 0.0029127809684723616, + 0.002359001198783517, + 0.0020203387830406427, + 0.0020716325379908085, + 0.0020683028269559145, + 0.002920306520536542, + 0.0017005741829052567, + 0.0038942357059568167, + 0.0039850217290222645, + 0.0017546523595228791, + 0.0020884452387690544, + 0.0021323305554687977, + 0.002525022253394127, + 0.00908232294023037, + 0.006652952171862125, + 0.009796624071896076, + 0.013071389868855476, + 0.00819318275898695, + 0.006006429437547922, + 0.009313439019024372, + 0.007943728007376194, + 0.003875578287988901, + 0.0023612959776073694, + 0.005261964630335569, + 0.004691174253821373, + 0.002235960215330124, + 0.0029685760382562876, + 0.0030178215820342302, + 0.0028625151608139277, + 0.032485976815223694, + 0.029235754162073135, + 0.05052502825856209, + 0.046490829437971115, + 0.023244308307766914, + 0.03418383747339249, + 0.045564789324998856, + 0.028087357059121132, + 0.02985265664756298, + 0.014894533902406693, + 0.02463018335402012, + 0.01720118708908558, + 0.01409319881349802, + 0.020173335447907448, + 0.011660102754831314, + 0.018437961116433144, + 0.0084318732842803, + 0.00668557221069932, + 0.016276562586426735, + 0.01477332878857851, + 0.006433545146137476, + 0.013076831586658955, + 0.00870652124285698, + 0.011742456816136837, + 0.00751921720802784, + 0.00631080474704504, + 0.012154807336628437, + 0.012205364182591438, + 0.00514227757230401, + 0.008767549879848957, + 0.007167210336774588, + 0.006930124014616013, + 0.011618279851973057, + 0.009462324902415276, + 0.03557170554995537, + 0.015384121797978878, + 0.009257986210286617, + 0.013627078384160995, + 0.008313987404108047, + 0.04238839074969292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.15924285352230072 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020609166473150253 + ], + "opposite_logprob_tokens": [ + -3.895609140396118 + ], + "logprob_reward": -0.020609166473150253, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020609166473150253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08801094442605972, + 0.06708087772130966, + 0.06643083691596985, + 0.025672143325209618, + 0.05027855932712555, + 0.08130030333995819, + 0.03393889218568802, + 0.09608792513608932, + 0.3579977750778198, + 0.1332017034292221 + ], + "credit_map": [ + 0.013015905395150185, + 0.011079106479883194, + 0.013796060346066952, + 0.009937417693436146, + 0.013144787400960922, + 0.007816207595169544, + 0.009828019887208939, + 0.009393434971570969, + 0.008475187234580517, + 0.007569029461592436, + 0.012152323499321938, + 0.009255423210561275, + 0.009169015102088451, + 0.005972946994006634, + 0.008097254671156406, + 0.006389691960066557, + 0.006058664061129093, + 0.0072470312006771564, + 0.009425872005522251, + 0.008926601149141788, + 0.01263767946511507, + 0.00724238995462656, + 0.00827126856893301, + 0.006621335633099079, + 0.00345276715233922, + 0.0018580472096800804, + 0.004492874722927809, + 0.003414392238482833, + 0.0016903714276850224, + 0.003442601300776005, + 0.002334014978259802, + 0.004987074062228203, + 0.006255354266613722, + 0.0052642361260950565, + 0.008537409827113152, + 0.008006048388779163, + 0.004454921931028366, + 0.005797787569463253, + 0.006704594008624554, + 0.005258208140730858, + 0.00844579003751278, + 0.00908577349036932, + 0.01298503391444683, + 0.012097958475351334, + 0.0071937767788767815, + 0.009353846311569214, + 0.014644281938672066, + 0.007493846584111452, + 0.004308897536247969, + 0.0033444236032664776, + 0.005113472696393728, + 0.004856302868574858, + 0.003514569718390703, + 0.005491461604833603, + 0.0037260407116264105, + 0.0035837236791849136, + 0.009159252978861332, + 0.009859152138233185, + 0.017549606040120125, + 0.016463419422507286, + 0.006641204934567213, + 0.021452926099300385, + 0.006372204516083002, + 0.008590159006416798, + 0.030520958825945854, + 0.024690119549632072, + 0.07545166462659836, + 0.047263383865356445, + 0.02696550451219082, + 0.07007399946451187, + 0.037644162774086, + 0.045388005673885345, + 0.014600016176700592, + 0.00914475042372942, + 0.028774430975317955, + 0.019801434129476547, + 0.011745069175958633, + 0.013654879294335842, + 0.008783970959484577, + 0.02669716067612171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.020609166473150253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958769004791975 + ], + "opposite_logprob_tokens": [ + -5.130958557128906 + ], + "logprob_reward": -0.005958769004791975, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958769004791975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0915217325091362, + 0.02426202967762947, + 0.0400582030415535, + 0.04924396798014641, + 0.05200766772031784, + 0.05008407682180405, + 0.02834656834602356, + 0.1270626187324524, + 0.43584391474723816, + 0.10156919807195663 + ], + "credit_map": [ + 0.012089000083506107, + 0.010738076642155647, + 0.017239006236195564, + 0.012031519785523415, + 0.01298260223120451, + 0.007258664350956678, + 0.010243259370326996, + 0.008939600549638271, + 0.0030797382351011038, + 0.002937041688710451, + 0.00538244005292654, + 0.0030475512612611055, + 0.0034669972956180573, + 0.0018140408210456371, + 0.0022059879265725613, + 0.0023282335605472326, + 0.00432389322668314, + 0.0034981961362063885, + 0.006772433407604694, + 0.007185301277786493, + 0.00474722171202302, + 0.00469721807166934, + 0.00381521531380713, + 0.005018720403313637, + 0.006702843122184277, + 0.004448622930794954, + 0.0074500045739114285, + 0.008550623431801796, + 0.0046103522181510925, + 0.004872289951890707, + 0.0056645506992936134, + 0.006944679655134678, + 0.006600400898605585, + 0.00578540051355958, + 0.008770216256380081, + 0.008713535964488983, + 0.004861387889832258, + 0.0051496075466275215, + 0.006870439276099205, + 0.005256676580756903, + 0.007083254400640726, + 0.004523939918726683, + 0.0075769489631056786, + 0.008441464975476265, + 0.0037299073301255703, + 0.0063671208918094635, + 0.007698416709899902, + 0.004663021303713322, + 0.004498730413615704, + 0.0019101820653304458, + 0.003997286316007376, + 0.004354665521532297, + 0.002897014142945409, + 0.0037809310015290976, + 0.0037438615690916777, + 0.0031638965010643005, + 0.013146381825208664, + 0.012184272520244122, + 0.035131193697452545, + 0.016503000631928444, + 0.006777155213057995, + 0.02378571778535843, + 0.007307916879653931, + 0.012226973660290241, + 0.031439173966646194, + 0.03198632597923279, + 0.09984714537858963, + 0.04775555431842804, + 0.02264402247965336, + 0.10188352316617966, + 0.036828864365816116, + 0.06345932185649872, + 0.007186753675341606, + 0.005102012772113085, + 0.028396498411893845, + 0.007093222811818123, + 0.0060512470081448555, + 0.009980790317058563, + 0.004805013537406921, + 0.032953664660453796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.005958769004791975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818847469985485 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.01818847469985485, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818847469985485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14826078712940216, + 0.02845025807619095, + 0.025383900851011276, + 0.04892939701676369, + 0.03866037353873253, + 0.4481350779533386, + 0.0874551311135292, + 0.0625833198428154, + 0.07911454886198044, + 0.03302721306681633 + ], + "credit_map": [ + 0.020952939987182617, + 0.014346007257699966, + 0.03756789118051529, + 0.016446050256490707, + 0.019212080165743828, + 0.011016751639544964, + 0.01518168207257986, + 0.013537377119064331, + 0.004998031537979841, + 0.0030885408632457256, + 0.006112728267908096, + 0.0032790405675768852, + 0.0034634871408343315, + 0.0017620655708014965, + 0.002382536418735981, + 0.003363825147971511, + 0.0026983816642314196, + 0.002744670957326889, + 0.0034562854561954737, + 0.003954174462705851, + 0.004433784633874893, + 0.002185496501624584, + 0.003149523166939616, + 0.0027615863364189863, + 0.006901873275637627, + 0.004604045767337084, + 0.007273769937455654, + 0.009070814587175846, + 0.006007944233715534, + 0.004488223232328892, + 0.005209671799093485, + 0.005373054184019566, + 0.0050392975099384785, + 0.0035022550728172064, + 0.007088589947670698, + 0.0068560377694666386, + 0.0031367670744657516, + 0.004280759021639824, + 0.004582189489156008, + 0.004174478817731142, + 0.05335451290011406, + 0.04423600807785988, + 0.07927557826042175, + 0.07144775241613388, + 0.03519880399107933, + 0.054017018526792526, + 0.06586384028196335, + 0.044741563498973846, + 0.016175640746951103, + 0.008880987763404846, + 0.013517056591808796, + 0.011545930989086628, + 0.008104809559881687, + 0.011790655553340912, + 0.006653187796473503, + 0.01078686211258173, + 0.005498481914401054, + 0.004375393036752939, + 0.012724598869681358, + 0.011650179512798786, + 0.004400300793349743, + 0.011991851031780243, + 0.005824380088597536, + 0.0061181336641311646, + 0.007001268211752176, + 0.005087745375931263, + 0.01841599866747856, + 0.009376694448292255, + 0.0054114386439323425, + 0.016253022477030754, + 0.007669202517718077, + 0.009899181313812733, + 0.003527363296598196, + 0.0025604283437132835, + 0.004955901298671961, + 0.004844030365347862, + 0.004259249661117792, + 0.004508928395807743, + 0.0037547647953033447, + 0.004616547375917435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.01818847469985485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and positioned directly overhead (top-down)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfect blob of thick, golden-amber honey?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the honey resting on a black, highly reflective countertop resembling obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the honey slowly deform, with its leading edge creeping forward and stretching?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are brilliant specular highlights visible on the honey's surface due to soft, directed studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642531931400299 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642531931400299, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642531931400299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0878848060965538, + 0.036689892411231995, + 0.044424887746572495, + 0.06074530631303787, + 0.02903028205037117, + 0.33189845085144043, + 0.037216246128082275, + 0.04924248158931732, + 0.2381623387336731, + 0.08470532298088074 + ], + "credit_map": [ + 0.009440124966204166, + 0.011207913979887962, + 0.01242323312908411, + 0.011833718046545982, + 0.01396033726632595, + 0.008087282069027424, + 0.011600617319345474, + 0.009331576526165009, + 0.003925255499780178, + 0.00427861837670207, + 0.005515141878277063, + 0.005023470148444176, + 0.005209637340158224, + 0.0036450133193284273, + 0.0051788147538900375, + 0.003913941327482462, + 0.004536341410130262, + 0.004646077752113342, + 0.007454747334122658, + 0.007212781347334385, + 0.005750473588705063, + 0.005423591937869787, + 0.00466808769851923, + 0.004732787609100342, + 0.009902624413371086, + 0.005719656590372324, + 0.009930907748639584, + 0.011400077491998672, + 0.0058984821662306786, + 0.005634540691971779, + 0.004901170264929533, + 0.007357841823250055, + 0.004054488614201546, + 0.0030943467281758785, + 0.005253474228084087, + 0.004805571865290403, + 0.002566492184996605, + 0.002994891256093979, + 0.0036520552821457386, + 0.0026089611928910017, + 0.03852405026555061, + 0.03482082486152649, + 0.05718863010406494, + 0.05350755155086517, + 0.02680775336921215, + 0.03617829084396362, + 0.054673295468091965, + 0.030198032036423683, + 0.004908339586108923, + 0.004292035475373268, + 0.005390217062085867, + 0.00577775901183486, + 0.003915635868906975, + 0.005214462988078594, + 0.003971238620579243, + 0.003746560774743557, + 0.00528906611725688, + 0.0056823198683559895, + 0.010230469517409801, + 0.0068489485420286655, + 0.003854099428281188, + 0.009621985256671906, + 0.003339486662298441, + 0.0043761092238128185, + 0.017642496153712273, + 0.016207987442612648, + 0.057134345173835754, + 0.02628352679312229, + 0.014478570781648159, + 0.050490669906139374, + 0.023470226675271988, + 0.03245450183749199, + 0.011274571530520916, + 0.006790280807763338, + 0.016455337405204773, + 0.013751639984548092, + 0.007160290610045195, + 0.009891407564282417, + 0.007494642864912748, + 0.011887152679264545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.008642531931400299 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_1/index_tensor([1184, 1184])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_1/index_tensor([1184, 1184])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8b524d5f15ba5861b6614a2cf6aa3e84a478fbcd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_1/index_tensor([1184, 1184])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a scene steeped in dramatic, high-contrast lighting, a close-up shot from an extremely low angle, captured with the shaky feel of a handheld camera, captures a Muay Thai fighter's muscular, sweat-beaded calf, like a whip tautly released, fiercely lashing out at a suspended sepak takraw. Shot in exquisite slow-motion detail, the camera follows closely, each violent impact trembling the frame, clearly revealing the momentary, subtle deformation of the sepak takraw's rough, intricately woven rattan surface. Simultaneously, countless glistening beads of sweat are flung from the skin of the leg, transforming into countless splatters of light under the harsh overhead lighting. The background is a profound darkness, highlighting the outline of the leg and the splattering liquid. The entire scene is imbued with a sense of primal power and breathtaking dynamic tension. The cold tones are punctuated only by the warm tones of the skin and the luminous white of the sweat, creating a visual experience of focus, violence, and profound beauty.", + "seed": 20626, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_1/index_1184_seed_20626_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03234412893652916, + "vlm_logprob": -0.03234412893652916, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.03234412893652916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02800796739757061 + ], + "opposite_logprob_tokens": [ + -4.028007984161377 + ], + "logprob_reward": -0.02800796739757061, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02800796739757061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1487312763929367, + 0.10620227456092834, + 0.12667997181415558, + 0.052069198340177536, + 0.03366681933403015, + 0.3301682472229004, + 0.04728648439049721, + 0.0328415185213089, + 0.06183495745062828, + 0.06051921099424362 + ], + "credit_map": [ + 0.006929975468665361, + 0.009555651806294918, + 0.014870671555399895, + 0.004945278633385897, + 0.003916630987077951, + 0.00544927641749382, + 0.047113705426454544, + 0.0559501014649868, + 0.008211334235966206, + 0.021770359948277473, + 0.01689017005264759, + 0.009889083914458752, + 0.006411202251911163, + 0.00597195653244853, + 0.01971261575818062, + 0.017345553264021873, + 0.01833246275782585, + 0.01450480055063963, + 0.015305785462260246, + 0.02243536151945591, + 0.014820437878370285, + 0.0112306484952569, + 0.010834449902176857, + 0.019216030836105347, + 0.007258592639118433, + 0.007233161944895983, + 0.008424224331974983, + 0.007444718386977911, + 0.006660700310021639, + 0.003740772372111678, + 0.004904302768409252, + 0.006402723025530577, + 0.004032393451780081, + 0.003552905051037669, + 0.0051015145145356655, + 0.0042489031329751015, + 0.004658594261854887, + 0.0024295065086334944, + 0.004144120961427689, + 0.005498881451785564, + 0.013188590295612812, + 0.013877345249056816, + 0.06559428572654724, + 0.14057657122612, + 0.018276087939739227, + 0.04268117994070053, + 0.018840447068214417, + 0.017133738845586777, + 0.0038159540854394436, + 0.005703379400074482, + 0.003317002672702074, + 0.003984353039413691, + 0.0035297907888889313, + 0.010501040145754814, + 0.003055739914998412, + 0.013379225507378578, + 0.004084025975316763, + 0.0048935445956885815, + 0.002881897846236825, + 0.0038102257531136274, + 0.004397427663207054, + 0.004177435301244259, + 0.00377786997705698, + 0.004819088615477085, + 0.005391753278672695, + 0.007940622977912426, + 0.0067018745467066765, + 0.0123301325365901, + 0.007684085518121719, + 0.008935677818953991, + 0.005705949850380421, + 0.007144860923290253, + 0.004637165926396847, + 0.011195608414709568, + 0.005849369801580906, + 0.011349999345839024, + 0.004892148543149233, + 0.0058419411070644855, + 0.007794864475727081, + 0.008958112448453903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.02800796739757061 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038144346326589584 + ], + "opposite_logprob_tokens": [ + -3.28814435005188 + ], + "logprob_reward": -0.038144346326589584, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038144346326589584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13994529843330383, + 0.052279919385910034, + 0.20800454914569855, + 0.04579991102218628, + 0.044649917632341385, + 0.15451622009277344, + 0.09795763343572617, + 0.17749738693237305, + 0.028230631723999977, + 0.05111853778362274 + ], + "credit_map": [ + 0.00824039801955223, + 0.022955600172281265, + 0.016317036002874374, + 0.00901268795132637, + 0.006040685344487429, + 0.005865768063813448, + 0.039363086223602295, + 0.0321500301361084, + 0.005124807823449373, + 0.008294866420328617, + 0.006655143573880196, + 0.005628685466945171, + 0.0028007354121655226, + 0.004453711677342653, + 0.009544129483401775, + 0.00977784302085638, + 0.02815130166709423, + 0.025203226134181023, + 0.01960274763405323, + 0.039824649691581726, + 0.028500789776444435, + 0.022816035896539688, + 0.015949472784996033, + 0.027956334874033928, + 0.005238188896328211, + 0.00812119897454977, + 0.006002339534461498, + 0.006259908899664879, + 0.0060631935484707355, + 0.002958941040560603, + 0.004659728147089481, + 0.006496409885585308, + 0.008081869222223759, + 0.005501686129719019, + 0.0069446200504899025, + 0.005358651280403137, + 0.0032034178730100393, + 0.004088921006768942, + 0.004168237559497356, + 0.0073025161400437355, + 0.008730023168027401, + 0.009634895250201225, + 0.025609929114580154, + 0.06388859450817108, + 0.010924586094915867, + 0.017378292977809906, + 0.009582064114511013, + 0.008767833933234215, + 0.012115482240915298, + 0.011487431824207306, + 0.009863410145044327, + 0.013607347384095192, + 0.006406060885637999, + 0.012981266714632511, + 0.007825318723917007, + 0.023671312257647514, + 0.01409487146884203, + 0.01856161653995514, + 0.0088918162509799, + 0.020931681618094444, + 0.023071179166436195, + 0.028189385309815407, + 0.019844915717840195, + 0.04391191527247429, + 0.002958418568596244, + 0.0038491387385874987, + 0.003981932997703552, + 0.005378414411097765, + 0.0032090535387396812, + 0.003302201395854354, + 0.002459346316754818, + 0.003092127153649926, + 0.004667428322136402, + 0.010992789641022682, + 0.00517173670232296, + 0.005191766656935215, + 0.005028580315411091, + 0.006145895924419165, + 0.005210350267589092, + 0.008709991350769997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.038144346326589584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055007439106702805 + ], + "opposite_logprob_tokens": [ + -2.9300074577331543 + ], + "logprob_reward": -0.055007439106702805, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055007439106702805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08550349622964859, + 0.02965824119746685, + 0.036400094628334045, + 0.04758879542350769, + 0.09262514114379883, + 0.5182409882545471, + 0.0648631602525711, + 0.07054566591978073, + 0.03544867783784866, + 0.019125809893012047 + ], + "credit_map": [ + 0.005233002360910177, + 0.014971991069614887, + 0.00933961383998394, + 0.0060080732218921185, + 0.003619050607085228, + 0.00441743666306138, + 0.020992623642086983, + 0.02092169225215912, + 0.0024844715371727943, + 0.005294497590512037, + 0.004574503283947706, + 0.001691402867436409, + 0.0012671686708927155, + 0.0012524406192824244, + 0.006772987544536591, + 0.006320766173303127, + 0.004559315741062164, + 0.004137782845646143, + 0.003136544954031706, + 0.0061632851138710976, + 0.0027733349706977606, + 0.005725609604269266, + 0.004178754519671202, + 0.005725460592657328, + 0.005587183870375156, + 0.005239947233349085, + 0.007554139010608196, + 0.008502538315951824, + 0.006677047815173864, + 0.0035926462151110172, + 0.003952479921281338, + 0.006482809316366911, + 0.009765317663550377, + 0.011600961908698082, + 0.013029179535806179, + 0.01371539942920208, + 0.012552584521472454, + 0.008098975755274296, + 0.011673239059746265, + 0.012189465574920177, + 0.020147480070590973, + 0.022487659007310867, + 0.09126605838537216, + 0.22516797482967377, + 0.02237546071410179, + 0.08279141038656235, + 0.02854607254266739, + 0.02545878104865551, + 0.00800030305981636, + 0.006599711254239082, + 0.006811336148530245, + 0.008029383607208729, + 0.004566124174743891, + 0.00899964477866888, + 0.0046923356130719185, + 0.017164310440421104, + 0.0059038265608251095, + 0.008442860096693039, + 0.0042745331302285194, + 0.009329689666628838, + 0.009432987309992313, + 0.010412346571683884, + 0.007543935440480709, + 0.01520547829568386, + 0.0034621262457221746, + 0.005001791752874851, + 0.004004099406301975, + 0.006900989916175604, + 0.004445627797394991, + 0.005340320989489555, + 0.002939273603260517, + 0.003354440676048398, + 0.0017997261602431536, + 0.0024214754812419415, + 0.002476713852956891, + 0.002867472358047962, + 0.0016324231401085854, + 0.0018869842169806361, + 0.003148977179080248, + 0.00289203436113894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.055007439106702805 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205407351255417 + ], + "opposite_logprob_tokens": [ + -4.264205455780029 + ], + "logprob_reward": -0.014205407351255417, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205407351255417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1758875697851181, + 0.03388463705778122, + 0.3281170725822449, + 0.034971222281455994, + 0.087143175303936, + 0.189186692237854, + 0.052229732275009155, + 0.01642514020204544, + 0.02587682567536831, + 0.05627791956067085 + ], + "credit_map": [ + 0.005468078888952732, + 0.016671236604452133, + 0.021502112969756126, + 0.005512673407793045, + 0.004277501255273819, + 0.006424938328564167, + 0.052114684134721756, + 0.06391634047031403, + 0.0028623719699680805, + 0.004111183807253838, + 0.004581304732710123, + 0.003717259271070361, + 0.0017654466209933162, + 0.002636430785059929, + 0.006575098726898432, + 0.007635542657226324, + 0.04073804244399071, + 0.03575393557548523, + 0.029894234612584114, + 0.057912666350603104, + 0.044155318289995193, + 0.04178189858794212, + 0.028711235150694847, + 0.049169741570949554, + 0.004060889128595591, + 0.003756132209673524, + 0.006581383291631937, + 0.006936540361493826, + 0.003990171942859888, + 0.0024101673625409603, + 0.002459176816046238, + 0.004776760935783386, + 0.012120491825044155, + 0.009576791897416115, + 0.014063048176467419, + 0.0112658916041255, + 0.009284991770982742, + 0.007639447692781687, + 0.008770464919507504, + 0.014422050677239895, + 0.008575793355703354, + 0.009191201068460941, + 0.026136379688978195, + 0.07821206748485565, + 0.007218773011118174, + 0.040713198482990265, + 0.00929353479295969, + 0.009845729917287827, + 0.004389691632241011, + 0.006004679016768932, + 0.004717694595456123, + 0.005505449138581753, + 0.004446303937584162, + 0.008375437930226326, + 0.004961384925991297, + 0.013829094357788563, + 0.0018120978493243456, + 0.0019678391981869936, + 0.001241560559719801, + 0.001657084678299725, + 0.002330239862203598, + 0.0017795246094465256, + 0.003371593775227666, + 0.002265198854729533, + 0.0034669507294893265, + 0.002441822085529566, + 0.003256849944591522, + 0.0031809594947844744, + 0.0044317059218883514, + 0.0037438087165355682, + 0.0019116868497803807, + 0.0034430427476763725, + 0.00492406589910388, + 0.009830975905060768, + 0.006126458756625652, + 0.008091764524579048, + 0.004202600568532944, + 0.0062986197881400585, + 0.008229480125010014, + 0.008573955856263638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.014205407351255417 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355484500527382 + ], + "opposite_logprob_tokens": [ + -3.651355504989624 + ], + "logprob_reward": -0.026355484500527382, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355484500527382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06520332396030426, + 0.029824640601873398, + 0.30456995964050293, + 0.01372919324785471, + 0.03727707639336586, + 0.4642414450645447, + 0.010476568713784218, + 0.029736017808318138, + 0.010680834762752056, + 0.034260932356119156 + ], + "credit_map": [ + 0.005757538136094809, + 0.009417473338544369, + 0.008458437398076057, + 0.004499304108321667, + 0.0034087253734469414, + 0.004553552716970444, + 0.012909159064292908, + 0.016199132427573204, + 0.0025979785714298487, + 0.005026162136346102, + 0.003639442380517721, + 0.002629669615998864, + 0.0015917749842628837, + 0.001800872036255896, + 0.006600554101169109, + 0.005938189569860697, + 0.03766052424907684, + 0.03352789953351021, + 0.02794550731778145, + 0.05540981516242027, + 0.04284161329269409, + 0.0395599864423275, + 0.02496623620390892, + 0.04265839606523514, + 0.0019133079331368208, + 0.0020490572787821293, + 0.0017312783747911453, + 0.0019595527555793524, + 0.001630618586204946, + 0.0012519859010353684, + 0.0012777216034010053, + 0.0019156697671860456, + 0.0034360794816166162, + 0.003952818922698498, + 0.00526801822707057, + 0.005949496291577816, + 0.004618464037775993, + 0.004658789839595556, + 0.004985415376722813, + 0.004407994914799929, + 0.018190568313002586, + 0.020092127844691277, + 0.07163067162036896, + 0.19557368755340576, + 0.018585748970508575, + 0.0927966982126236, + 0.02523849718272686, + 0.02213345654308796, + 0.00107859680429101, + 0.0010878561297431588, + 0.0010673394426703453, + 0.0009242225205525756, + 0.0015140275936573744, + 0.0016052405117079616, + 0.0014253434492275119, + 0.0017739433096721768, + 0.001859742682427168, + 0.0026544907595962286, + 0.001016091788187623, + 0.0032942029647529125, + 0.003998427651822567, + 0.005095126107335091, + 0.0035261542070657015, + 0.00829178374260664, + 0.0007942395750433207, + 0.0013291798532009125, + 0.0009526687790639699, + 0.0017071346519514918, + 0.001536156632937491, + 0.001848725718446076, + 0.001010274630971253, + 0.0015024541644379497, + 0.0028976898174732924, + 0.005847747437655926, + 0.004178371746093035, + 0.004440550692379475, + 0.002651879098266363, + 0.003814798779785633, + 0.005091232713311911, + 0.005338661838322878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.026355484500527382 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene steeped in dramatic, high-contrast lighting, a close-up shot from an extremely low angle, captured with the shaky feel of a handheld camera, captures a Muay Thai fighter's muscular, sweat-beaded calf, like a whip tautly released, fiercely lashing out at a suspended sepak takraw. Shot in exquisite slow-motion detail, the camera follows closely, each violent impact trembling the frame, clearly revealing the momentary, subtle deformation of the sepak takraw's rough, intricately woven rattan surface. Simultaneously, countless glistening beads of sweat are flung from the skin of the leg, transforming into countless splatters of light under the harsh overhead lighting. The background is a profound darkness, highlighting the outline of the leg and the splattering liquid. The entire scene is imbued with a sense of primal power and breathtaking dynamic tension. The cold tones are punctuated only by the warm tones of the skin and the luminous white of the sweat, creating a visual experience of focus, violence, and profound beauty.", + "seed": 20627, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_1/index_1184_seed_20627_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.048818362317979334, + "vlm_logprob": -0.048818362317979334, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.048818362317979334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034756194800138474 + ], + "opposite_logprob_tokens": [ + -4.284756183624268 + ], + "logprob_reward": -0.034756194800138474, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.034756194800138474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09363794326782227, + 0.06034078076481819, + 0.07851014286279678, + 0.20806710422039032, + 0.06949594616889954, + 0.031256675720214844, + 0.1446128487586975, + 0.06717447191476822, + 0.04285309836268425, + 0.2040509581565857 + ], + "credit_map": [ + 0.01728600636124611, + 0.010547407902777195, + 0.013536885380744934, + 0.0057118842378258705, + 0.00900441873818636, + 0.015687236562371254, + 0.013048841618001461, + 0.008815264329314232, + 0.006909057032316923, + 0.007588853128254414, + 0.00998985767364502, + 0.0064386832527816296, + 0.003509798552840948, + 0.009021450765430927, + 0.01195976696908474, + 0.004923312924802303, + 0.006413668394088745, + 0.009608776308596134, + 0.01346791535615921, + 0.008736730553209782, + 0.0030050245113670826, + 0.009916715323925018, + 0.0037113751750439405, + 0.023649930953979492, + 0.017531413584947586, + 0.015917159616947174, + 0.037752341479063034, + 0.01280123833566904, + 0.011675046756863594, + 0.06990527361631393, + 0.011143723502755165, + 0.031340908259153366, + 0.006186227314174175, + 0.007945667020976543, + 0.009977479465305805, + 0.009683800861239433, + 0.005969434976577759, + 0.009215088561177254, + 0.010277930647134781, + 0.010240321047604084, + 0.003714081598445773, + 0.0024498512502759695, + 0.0034012410324066877, + 0.003006372367963195, + 0.003445689333602786, + 0.004893459379673004, + 0.005341476760804653, + 0.005004505161195993, + 0.018602322787046432, + 0.008100463077425957, + 0.01388217881321907, + 0.01472401898354292, + 0.006281808950006962, + 0.008925152942538261, + 0.03873500972986221, + 0.035361889749765396, + 0.005852835718542337, + 0.0028379959985613823, + 0.014593008905649185, + 0.005216591060161591, + 0.006084993481636047, + 0.0066549465991556644, + 0.01602814346551895, + 0.00990595668554306, + 0.005353901535272598, + 0.0034440308809280396, + 0.009591146372258663, + 0.002388885011896491, + 0.0031493071001023054, + 0.0038091177120804787, + 0.004290039651095867, + 0.010826666839420795, + 0.02462628483772278, + 0.020982732996344566, + 0.03110126592218876, + 0.04176789149641991, + 0.03087846376001835, + 0.013822786509990692, + 0.0187828429043293, + 0.022088684141635895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.034756194800138474 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870803281664848 + ], + "opposite_logprob_tokens": [ + -3.048707962036133 + ], + "logprob_reward": -0.04870803281664848, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870803281664848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15597228705883026, + 0.05718124285340309, + 0.11874270439147949, + 0.10098466277122498, + 0.05393168702721596, + 0.019542226567864418, + 0.07273271679878235, + 0.044678039848804474, + 0.05070978403091431, + 0.3255246579647064 + ], + "credit_map": [ + 0.02627858892083168, + 0.021989166736602783, + 0.01941242441534996, + 0.010494422167539597, + 0.02151646465063095, + 0.023570168763399124, + 0.020329244434833527, + 0.012381796725094318, + 0.005798787344247103, + 0.007644519675523043, + 0.009083036333322525, + 0.007538233883678913, + 0.004384668543934822, + 0.008305386640131474, + 0.009256844408810139, + 0.005169765092432499, + 0.010850398801267147, + 0.011794018559157848, + 0.021328963339328766, + 0.018225232139229774, + 0.005714924540370703, + 0.02011321857571602, + 0.007661113049834967, + 0.023054827004671097, + 0.010672308504581451, + 0.01919223926961422, + 0.01006328035145998, + 0.014504378661513329, + 0.009222937747836113, + 0.01601993851363659, + 0.006302703637629747, + 0.01500687189400196, + 0.004433987196534872, + 0.005587291903793812, + 0.006435200106352568, + 0.007259650155901909, + 0.00608815485611558, + 0.006318831816315651, + 0.009900853969156742, + 0.007907716557383537, + 0.0028210426680743694, + 0.0018965478520840406, + 0.0015870837960392237, + 0.0020611314103007317, + 0.0017814193852245808, + 0.0038666033651679754, + 0.0032045196276158094, + 0.0023238796275109053, + 0.006844364106655121, + 0.0031928441021591425, + 0.004212055820971727, + 0.006383990403264761, + 0.0031790416687726974, + 0.0034961372148245573, + 0.02396022528409958, + 0.02146405354142189, + 0.007734277285635471, + 0.0033200483303517103, + 0.007796462159603834, + 0.004889810457825661, + 0.002409368520602584, + 0.0027065735775977373, + 0.006687150802463293, + 0.009134349413216114, + 0.0063109733164310455, + 0.0074478317983448505, + 0.007312223315238953, + 0.0053255301900208, + 0.006752349436283112, + 0.005260988138616085, + 0.006601364817470312, + 0.005698519758880138, + 0.05137602984905243, + 0.03469053655862808, + 0.056485023349523544, + 0.05230805650353432, + 0.04488227143883705, + 0.02178572490811348, + 0.03026442416012287, + 0.033732593059539795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.04870803281664848 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10040510445833206 + ], + "opposite_logprob_tokens": [ + -2.350405216217041 + ], + "logprob_reward": -0.10040510445833206, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10040510445833206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10273575037717819, + 0.0806947872042656, + 0.15181422233581543, + 0.13552725315093994, + 0.0709533840417862, + 0.027786478400230408, + 0.03138623386621475, + 0.09393063932657242, + 0.13824346661567688, + 0.1669277846813202 + ], + "credit_map": [ + 0.013768852688372135, + 0.014274363406002522, + 0.01606549322605133, + 0.00907871127128601, + 0.01152305118739605, + 0.02113056741654873, + 0.007427266798913479, + 0.009467448107898235, + 0.008698086254298687, + 0.008930718526244164, + 0.013512705452740192, + 0.00847772229462862, + 0.0060910494066774845, + 0.013576238416135311, + 0.01485645491629839, + 0.006551808677613735, + 0.015146477147936821, + 0.013774966821074486, + 0.025898680090904236, + 0.012647404335439205, + 0.00573233375325799, + 0.018238894641399384, + 0.010065258480608463, + 0.050310201942920685, + 0.012946550734341145, + 0.019133659079670906, + 0.01618611440062523, + 0.016886992380023003, + 0.01144605316221714, + 0.027026863768696785, + 0.010738763026893139, + 0.021162256598472595, + 0.004522215574979782, + 0.006768624298274517, + 0.007615689653903246, + 0.008840327151119709, + 0.008162017911672592, + 0.009317968040704727, + 0.01484165620058775, + 0.010884881019592285, + 0.006119461264461279, + 0.002963503822684288, + 0.0033254544250667095, + 0.0038336601573973894, + 0.0027746581472456455, + 0.0033506511244922876, + 0.002631731564179063, + 0.0027873560320585966, + 0.006336374673992395, + 0.0026677618734538555, + 0.004514411091804504, + 0.003417080966755748, + 0.0023351290728896856, + 0.0026600935962051153, + 0.00536695122718811, + 0.004088428802788258, + 0.010596263222396374, + 0.007807715330272913, + 0.01012470293790102, + 0.010643922723829746, + 0.008703315630555153, + 0.01193277072161436, + 0.014956552535295486, + 0.019165394827723503, + 0.013424663804471493, + 0.01726371981203556, + 0.016745025292038918, + 0.015449588187038898, + 0.017618263140320778, + 0.019066669046878815, + 0.02013740874826908, + 0.018538126721978188, + 0.027732914313673973, + 0.018109794706106186, + 0.029857944697141647, + 0.021260259672999382, + 0.023312631994485855, + 0.012125607579946518, + 0.0177626833319664, + 0.016765931621193886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.10040510445833206 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257352255284786 + ], + "opposite_logprob_tokens": [ + -5.2552571296691895 + ], + "logprob_reward": -0.005257352255284786, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257352255284786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05291881784796715, + 0.04667629674077034, + 0.2650657296180725, + 0.05860673636198044, + 0.03786202520132065, + 0.04133376479148865, + 0.11959633231163025, + 0.04059692472219467, + 0.12204981595277786, + 0.21529361605644226 + ], + "credit_map": [ + 0.005995886400341988, + 0.005083269439637661, + 0.007286986336112022, + 0.004499311093240976, + 0.006965029984712601, + 0.010596238076686859, + 0.003879896132275462, + 0.008612200617790222, + 0.004333248361945152, + 0.0044227903708815575, + 0.00817626342177391, + 0.0045912303030490875, + 0.00321845430880785, + 0.012799558229744434, + 0.004080580081790686, + 0.005054172594100237, + 0.023060401901602745, + 0.015325753018260002, + 0.047693900763988495, + 0.025928055867552757, + 0.010169293731451035, + 0.04515965282917023, + 0.014778324402868748, + 0.08295033872127533, + 0.0053591132164001465, + 0.009985950775444508, + 0.006962069775909185, + 0.00832553394138813, + 0.005440037697553635, + 0.012529563158750534, + 0.004316714126616716, + 0.0056877536699175835, + 0.0031950476113706827, + 0.004550939425826073, + 0.005464855115860701, + 0.005663423798978329, + 0.0031967065297067165, + 0.004210332874208689, + 0.006378776393830776, + 0.005201943684369326, + 0.005137882195413113, + 0.005310774315148592, + 0.0032866850960999727, + 0.004095173440873623, + 0.005049946717917919, + 0.006172454450279474, + 0.005423345603048801, + 0.006857503671199083, + 0.007833369076251984, + 0.0035106190480291843, + 0.005285081919282675, + 0.005601631477475166, + 0.007950520142912865, + 0.008476611226797104, + 0.042742323130369186, + 0.03819616138935089, + 0.004051111172884703, + 0.002966658677905798, + 0.004835653118789196, + 0.0031768744811415672, + 0.005388875026255846, + 0.006487044971436262, + 0.006215250585228205, + 0.007475457154214382, + 0.0157065037637949, + 0.008352046832442284, + 0.02755468152463436, + 0.0057478900998830795, + 0.00959985051304102, + 0.01415233127772808, + 0.012772147543728352, + 0.02816435880959034, + 0.014574339613318443, + 0.011875065974891186, + 0.016918400302529335, + 0.015017805621027946, + 0.04430245980620384, + 0.0346086286008358, + 0.041862014681100845, + 0.036134909838438034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.005257352255284786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054965127259492874 + ], + "opposite_logprob_tokens": [ + -2.929965019226074 + ], + "logprob_reward": -0.054965127259492874, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054965127259492874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15315556526184082, + 0.06854435801506042, + 0.13053198158740997, + 0.19760406017303467, + 0.10555795580148697, + 0.05473332852125168, + 0.13064002990722656, + 0.05322503298521042, + 0.028813809156417847, + 0.07719387114048004 + ], + "credit_map": [ + 0.012439300306141376, + 0.019574493169784546, + 0.02871294878423214, + 0.013657025061547756, + 0.0172246266156435, + 0.03441859781742096, + 0.013865816406905651, + 0.01326275896281004, + 0.006140531040728092, + 0.008013944141566753, + 0.014404388144612312, + 0.006701059173792601, + 0.006004032678902149, + 0.01266397163271904, + 0.006047840695828199, + 0.008568592369556427, + 0.013399483636021614, + 0.012545593082904816, + 0.022712165489792824, + 0.016224900260567665, + 0.006670835893601179, + 0.02651037462055683, + 0.007075679954141378, + 0.02539294771850109, + 0.01770308054983616, + 0.027932552620768547, + 0.04104452580213547, + 0.024132484570145607, + 0.01329176314175129, + 0.04260912910103798, + 0.01151476800441742, + 0.01937575824558735, + 0.0067972224205732346, + 0.012081468477845192, + 0.016236456111073494, + 0.014999299310147762, + 0.011314361356198788, + 0.01076662540435791, + 0.01816588081419468, + 0.015196643769741058, + 0.006946722511202097, + 0.006694580428302288, + 0.004201657138764858, + 0.004122896585613489, + 0.005608567968010902, + 0.010711580514907837, + 0.008106647059321404, + 0.008340675383806229, + 0.00752321258187294, + 0.005024970043450594, + 0.0050771706737577915, + 0.007105915807187557, + 0.007968946360051632, + 0.009175773710012436, + 0.04576469957828522, + 0.04299933835864067, + 0.005731465294957161, + 0.004196134861558676, + 0.006041349843144417, + 0.0056957704946398735, + 0.005831629037857056, + 0.007416173350065947, + 0.010084563866257668, + 0.008227945305407047, + 0.002541579306125641, + 0.0041634500958025455, + 0.003923843614757061, + 0.0029788659885525703, + 0.004647630266845226, + 0.004129174165427685, + 0.002776070963591337, + 0.0036531933583319187, + 0.007512794341892004, + 0.006663173902779818, + 0.007957058027386665, + 0.009034405462443829, + 0.014290684834122658, + 0.009801550768315792, + 0.011340866796672344, + 0.010593338869512081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.054965127259492874 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene steeped in dramatic, high-contrast lighting, a close-up shot from an extremely low angle, captured with the shaky feel of a handheld camera, captures a Muay Thai fighter's muscular, sweat-beaded calf, like a whip tautly released, fiercely lashing out at a suspended sepak takraw. Shot in exquisite slow-motion detail, the camera follows closely, each violent impact trembling the frame, clearly revealing the momentary, subtle deformation of the sepak takraw's rough, intricately woven rattan surface. Simultaneously, countless glistening beads of sweat are flung from the skin of the leg, transforming into countless splatters of light under the harsh overhead lighting. The background is a profound darkness, highlighting the outline of the leg and the splattering liquid. The entire scene is imbued with a sense of primal power and breathtaking dynamic tension. The cold tones are punctuated only by the warm tones of the skin and the luminous white of the sweat, creating a visual experience of focus, violence, and profound beauty.", + "seed": 20628, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_1/index_1184_seed_20628_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.030055393092334272, + "vlm_logprob": -0.030055393092334272, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.030055393092334272, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023715177550911903 + ], + "opposite_logprob_tokens": [ + -4.898715019226074 + ], + "logprob_reward": -0.023715177550911903, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.023715177550911903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08436285704374313, + 0.10161728411912918, + 0.02064666338264942, + 0.054525624960660934, + 0.24021895229816437, + 0.060560088604688644, + 0.053718362003564835, + 0.11657433956861496, + 0.14655449986457825, + 0.12122133374214172 + ], + "credit_map": [ + 0.010347466915845871, + 0.008943775668740273, + 0.007908911444246769, + 0.009147871285676956, + 0.012064745649695396, + 0.009904921986162663, + 0.013929033651947975, + 0.012116131372749805, + 0.007776905782520771, + 0.009368585422635078, + 0.017821289598941803, + 0.012280390597879887, + 0.01337013952434063, + 0.009189534932374954, + 0.019314788281917572, + 0.012495648115873337, + 0.0020061524119228125, + 0.0029005217365920544, + 0.002808833494782448, + 0.0020956238731741905, + 0.003286950523033738, + 0.0024944422766566277, + 0.002729525323957205, + 0.0023246139753609896, + 0.00970430113375187, + 0.006038704887032509, + 0.007862016558647156, + 0.005488775670528412, + 0.004569866694509983, + 0.004060819745063782, + 0.007347343489527702, + 0.009453793056309223, + 0.026678701862692833, + 0.009709547273814678, + 0.062059272080659866, + 0.017430203035473824, + 0.05593764781951904, + 0.011726764030754566, + 0.047793347388505936, + 0.008883452042937279, + 0.006673711352050304, + 0.004676040727645159, + 0.012269217520952225, + 0.00594493979588151, + 0.010579253546893597, + 0.004627589136362076, + 0.012061087414622307, + 0.003728247247636318, + 0.005975113715976477, + 0.007057162467390299, + 0.003945217002183199, + 0.00625580083578825, + 0.008186928927898407, + 0.0056224362924695015, + 0.005808316636830568, + 0.010867384262382984, + 0.010744559578597546, + 0.011927284300327301, + 0.022437172010540962, + 0.029967404901981354, + 0.012162870727479458, + 0.009550202637910843, + 0.012445867992937565, + 0.0073389820754528046, + 0.012057471089065075, + 0.02267167903482914, + 0.023014681413769722, + 0.013530944474041462, + 0.022219665348529816, + 0.021558159962296486, + 0.015963995829224586, + 0.015537903644144535, + 0.010987450368702412, + 0.017458299174904823, + 0.018420275300741196, + 0.013929666951298714, + 0.014928193762898445, + 0.02271428517997265, + 0.01366074662655592, + 0.009122412651777267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.023715177550911903 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038135282695293427 + ], + "opposite_logprob_tokens": [ + -3.288135290145874 + ], + "logprob_reward": -0.038135282695293427, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038135282695293427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10740836709737778, + 0.12157902121543884, + 0.04397296532988548, + 0.10640020668506622, + 0.23267720639705658, + 0.05194483697414398, + 0.06114008277654648, + 0.10912052541971207, + 0.09660255163908005, + 0.06915421038866043 + ], + "credit_map": [ + 0.011151164770126343, + 0.01086503453552723, + 0.009903795085847378, + 0.015033932402729988, + 0.012338975444436073, + 0.010872317478060722, + 0.020914016291499138, + 0.016329128295183182, + 0.00967563409358263, + 0.013690675608813763, + 0.014766362495720387, + 0.019919343292713165, + 0.01364186406135559, + 0.013975526206195354, + 0.02311249077320099, + 0.012797120027244091, + 0.004222055897116661, + 0.006534116342663765, + 0.004878047853708267, + 0.0066489968448877335, + 0.005823791027069092, + 0.003646467812359333, + 0.006729016080498695, + 0.005490471143275499, + 0.017121409997344017, + 0.014953306876122952, + 0.015331446193158627, + 0.01626581884920597, + 0.005852620117366314, + 0.00696538295596838, + 0.010927965864539146, + 0.018982257694005966, + 0.028524648398160934, + 0.009637501090765, + 0.0785381868481636, + 0.010927017778158188, + 0.05741699039936066, + 0.010582990013062954, + 0.02766508422791958, + 0.009384780190885067, + 0.007807128597050905, + 0.006006630137562752, + 0.006450504995882511, + 0.004865546710789204, + 0.006659986916929483, + 0.005456858314573765, + 0.008825931698083878, + 0.00587225379422307, + 0.007132750004529953, + 0.008202854543924332, + 0.007096025627106428, + 0.00967862457036972, + 0.01027317438274622, + 0.006257084663957357, + 0.007796678692102432, + 0.004702890291810036, + 0.008815486915409565, + 0.008707118220627308, + 0.017916331067681313, + 0.01936374232172966, + 0.018196752294898033, + 0.010823452845215797, + 0.016123268753290176, + 0.009174374863505363, + 0.00688942102715373, + 0.015714241191744804, + 0.015620280988514423, + 0.011626520194113255, + 0.011963168159127235, + 0.014494270086288452, + 0.009773657657206059, + 0.010520994663238525, + 0.007366882171481848, + 0.00775910634547472, + 0.014390138909220695, + 0.0066975438967347145, + 0.006317836698144674, + 0.010574317537248135, + 0.008121286518871784, + 0.007927094586193562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.038135282695293427 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054997287690639496 + ], + "opposite_logprob_tokens": [ + -2.929997205734253 + ], + "logprob_reward": -0.054997287690639496, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054997287690639496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09343551844358444, + 0.13040846586227417, + 0.09776651859283447, + 0.07608703523874283, + 0.2229509800672531, + 0.08045686781406403, + 0.057698849588632584, + 0.07296963781118393, + 0.07101272791624069, + 0.09721342474222183 + ], + "credit_map": [ + 0.007834939286112785, + 0.00863554235547781, + 0.009404413402080536, + 0.009789970703423023, + 0.011184202507138252, + 0.014269320294260979, + 0.01764243096113205, + 0.014674700796604156, + 0.00812176987528801, + 0.010836973786354065, + 0.012683928944170475, + 0.019827937707304955, + 0.01806199736893177, + 0.012369289062917233, + 0.0341968834400177, + 0.01430968102067709, + 0.007732927333563566, + 0.018310029059648514, + 0.008983180858194828, + 0.01535482332110405, + 0.012948082759976387, + 0.011481584049761295, + 0.012029223144054413, + 0.01092666294425726, + 0.007083171047270298, + 0.009871489368379116, + 0.009499032981693745, + 0.009715724736452103, + 0.006976637057960033, + 0.0058404263108968735, + 0.014137010090053082, + 0.01296354178339243, + 0.026822097599506378, + 0.01007916871458292, + 0.07771336287260056, + 0.01166295725852251, + 0.05862100422382355, + 0.009174524806439877, + 0.018965166062116623, + 0.009912701323628426, + 0.008153099566698074, + 0.004807396791875362, + 0.017809003591537476, + 0.006966427434235811, + 0.014858175069093704, + 0.005831252317875624, + 0.01740375906229019, + 0.004627751652151346, + 0.004137792624533176, + 0.007536751683801413, + 0.0043189143761992455, + 0.0066015771590173244, + 0.011643418110907078, + 0.006574090104550123, + 0.009757626801729202, + 0.007128680124878883, + 0.00603229645639658, + 0.0064208246767520905, + 0.010577237233519554, + 0.010552866384387016, + 0.012862988747656345, + 0.0082770474255085, + 0.012113763019442558, + 0.006132608279585838, + 0.005785644520074129, + 0.008433562703430653, + 0.008078174665570259, + 0.009888904169201851, + 0.013153758831322193, + 0.010331516154110432, + 0.009816487319767475, + 0.005524673033505678, + 0.009247484616935253, + 0.008668097667396069, + 0.023076588287949562, + 0.0057473937049508095, + 0.01013708021491766, + 0.01802115887403488, + 0.01136708538979292, + 0.010948531329631805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.054997287690639496 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003619787283241749 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.003619787283241749, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003619787283241749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04267864301800728, + 0.1787901669740677, + 0.031488869339227676, + 0.3227173984050751, + 0.19351324439048767, + 0.03653575852513313, + 0.03172031417489052, + 0.040705375373363495, + 0.06725309044122696, + 0.054597143083810806 + ], + "credit_map": [ + 0.005682341754436493, + 0.006277226842939854, + 0.006919077597558498, + 0.004366858396679163, + 0.004692251328378916, + 0.004214247222989798, + 0.005587074439972639, + 0.004939567297697067, + 0.012907734140753746, + 0.015233955346047878, + 0.043056733906269073, + 0.01816132664680481, + 0.02323288284242153, + 0.01724272593855858, + 0.02308713085949421, + 0.025867683812975883, + 0.002886422909796238, + 0.003736660582944751, + 0.00422532157972455, + 0.0038861215580254793, + 0.004815727472305298, + 0.0038581627886742353, + 0.004445027094334364, + 0.0036354251205921173, + 0.05618513375520706, + 0.02853899449110031, + 0.04975579306483269, + 0.028798533603549004, + 0.01921907253563404, + 0.02223941870033741, + 0.0324026457965374, + 0.0855778157711029, + 0.02201959490776062, + 0.01134368684142828, + 0.044218435883522034, + 0.015308341011404991, + 0.039019543677568436, + 0.00869999174028635, + 0.04494320601224899, + 0.00796043872833252, + 0.003851741086691618, + 0.002041206695139408, + 0.009335544891655445, + 0.002521718619391322, + 0.006860538385808468, + 0.0020168956834822893, + 0.00794225838035345, + 0.0019658543169498444, + 0.0036258590407669544, + 0.00394454738125205, + 0.0027157028671354055, + 0.00404710927978158, + 0.0065829637460410595, + 0.0030793892219662666, + 0.0039508468471467495, + 0.003773893928155303, + 0.0033642854541540146, + 0.004585481248795986, + 0.008230561390519142, + 0.008614487014710903, + 0.005872896406799555, + 0.0032429902348667383, + 0.004318234045058489, + 0.002476440742611885, + 0.007598792668431997, + 0.01782587729394436, + 0.009522801265120506, + 0.006608135066926479, + 0.004413119982928038, + 0.010263822972774506, + 0.0035866370890289545, + 0.007433904800564051, + 0.004799420014023781, + 0.00654312688857317, + 0.009674490429461002, + 0.004042667802423239, + 0.0063712215051054955, + 0.010196976363658905, + 0.00487334793433547, + 0.008095888420939445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.003619787283241749 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029809430241584778 + ], + "opposite_logprob_tokens": [ + -3.5298094749450684 + ], + "logprob_reward": -0.029809430241584778, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029809430241584778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08452820777893066, + 0.05081807076931, + 0.06500548124313354, + 0.05265288054943085, + 0.17923592031002045, + 0.08686036616563797, + 0.1441764086484909, + 0.038459524512290955, + 0.15777967870235443, + 0.14048343896865845 + ], + "credit_map": [ + 0.013101832009851933, + 0.009233827702701092, + 0.0143088698387146, + 0.008101746439933777, + 0.008661492727696896, + 0.00809010025113821, + 0.011890229769051075, + 0.011140107177197933, + 0.0030261583160609007, + 0.004022080916911364, + 0.009883147664368153, + 0.005637906491756439, + 0.004927659407258034, + 0.006464190315455198, + 0.00450865039601922, + 0.012348275631666183, + 0.005959768779575825, + 0.013781899586319923, + 0.009393292479217052, + 0.008868684060871601, + 0.007150439079850912, + 0.004662327468395233, + 0.00704220263287425, + 0.008146868087351322, + 0.007455542217940092, + 0.005189794581383467, + 0.0068596890196204185, + 0.007546716835349798, + 0.005350579507648945, + 0.0041230106726288795, + 0.009595783427357674, + 0.006531762424856424, + 0.01895882561802864, + 0.01088644377887249, + 0.05674266070127487, + 0.013632992282509804, + 0.04357936605811119, + 0.010504882782697678, + 0.01606837287545204, + 0.008862362243235111, + 0.00798119604587555, + 0.005428791046142578, + 0.019220426678657532, + 0.00716814212501049, + 0.018170807510614395, + 0.005676156375557184, + 0.017900994047522545, + 0.00531384814530611, + 0.01745687611401081, + 0.017731819301843643, + 0.011835788376629353, + 0.017659660428762436, + 0.022355927154421806, + 0.013436302542686462, + 0.018362747505307198, + 0.02533729001879692, + 0.003343705553561449, + 0.004497716203331947, + 0.007547338493168354, + 0.009459669701755047, + 0.00498701399192214, + 0.0031243946868926287, + 0.0029958761297166348, + 0.0025038092862814665, + 0.012939474545419216, + 0.0266825370490551, + 0.027241045609116554, + 0.019139964133501053, + 0.018185855820775032, + 0.023814046755433083, + 0.011978170834481716, + 0.01779857836663723, + 0.012170907109975815, + 0.016538327559828758, + 0.024269791319966316, + 0.014782593585550785, + 0.019105345010757446, + 0.02059173956513405, + 0.017393164336681366, + 0.01563158631324768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.029809430241584778 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene steeped in dramatic, high-contrast lighting, a close-up shot from an extremely low angle, captured with the shaky feel of a handheld camera, captures a Muay Thai fighter's muscular, sweat-beaded calf, like a whip tautly released, fiercely lashing out at a suspended sepak takraw. Shot in exquisite slow-motion detail, the camera follows closely, each violent impact trembling the frame, clearly revealing the momentary, subtle deformation of the sepak takraw's rough, intricately woven rattan surface. Simultaneously, countless glistening beads of sweat are flung from the skin of the leg, transforming into countless splatters of light under the harsh overhead lighting. The background is a profound darkness, highlighting the outline of the leg and the splattering liquid. The entire scene is imbued with a sense of primal power and breathtaking dynamic tension. The cold tones are punctuated only by the warm tones of the skin and the luminous white of the sweat, creating a visual experience of focus, violence, and profound beauty.", + "seed": 20629, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_1/index_1184_seed_20629_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027831710502505304, + "vlm_logprob": -0.027831710502505304, + "vlm_yes_no_margin": 3.699999952316284, + "vlm_token_credit": -0.027831710502505304, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035849057137966156 + ], + "opposite_logprob_tokens": [ + -3.785849094390869 + ], + "logprob_reward": -0.035849057137966156, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.035849057137966156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08910448104143143, + 0.0616321861743927, + 0.05116266757249832, + 0.06850022077560425, + 0.05589041858911514, + 0.03716101124882698, + 0.055813685059547424, + 0.4093158543109894, + 0.1139591857790947, + 0.05746028572320938 + ], + "credit_map": [ + 0.005753607023507357, + 0.011212541721761227, + 0.01313974428921938, + 0.012234744615852833, + 0.011219445616006851, + 0.00802734587341547, + 0.01642700284719467, + 0.011090043000876904, + 0.006652302574366331, + 0.008431995287537575, + 0.009197895415127277, + 0.004678642377257347, + 0.008234105072915554, + 0.005990355275571346, + 0.01038782112300396, + 0.008059062995016575, + 0.0067101954482495785, + 0.006632914301007986, + 0.006459161173552275, + 0.009548845700919628, + 0.004883953835815191, + 0.004030268173664808, + 0.006998659111559391, + 0.005898667965084314, + 0.00694667175412178, + 0.013481918722391129, + 0.011123309843242168, + 0.01176226045936346, + 0.006090508308261633, + 0.004959189798682928, + 0.007302331272512674, + 0.006834026891738176, + 0.010574720799922943, + 0.011326530948281288, + 0.008857663720846176, + 0.0068343584425747395, + 0.004938781727105379, + 0.0030206525698304176, + 0.006526317447423935, + 0.0038113873451948166, + 0.005374575033783913, + 0.0043230848386883736, + 0.005291530396789312, + 0.0045976596884429455, + 0.006857241503894329, + 0.0038388418033719063, + 0.004778899718075991, + 0.0020991794299334288, + 0.0069899167865514755, + 0.006458869203925133, + 0.006367796566337347, + 0.0065288362093269825, + 0.009527841582894325, + 0.0045813885517418385, + 0.010992144234478474, + 0.004366889595985413, + 0.09677668660879135, + 0.02817852981388569, + 0.06883646547794342, + 0.02500205487012863, + 0.030416127294301987, + 0.047148145735263824, + 0.01953856274485588, + 0.09341927617788315, + 0.004555184859782457, + 0.004369362723082304, + 0.035669125616550446, + 0.002926895162090659, + 0.0037717558443546295, + 0.04145929962396622, + 0.002762445015832782, + 0.018445106223225594, + 0.005795736331492662, + 0.004991945810616016, + 0.005539409816265106, + 0.005273556336760521, + 0.01015693973749876, + 0.006379880476742983, + 0.01383020170032978, + 0.005492619704455137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.035849057137966156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026381026953458786 + ], + "opposite_logprob_tokens": [ + -3.651381015777588 + ], + "logprob_reward": -0.026381026953458786, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026381026953458786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10178454965353012, + 0.02673223242163658, + 0.07082649320363998, + 0.061746541410684586, + 0.04192739352583885, + 0.11044713854789734, + 0.09908255189657211, + 0.19979338347911835, + 0.26806309819221497, + 0.019596653059124947 + ], + "credit_map": [ + 0.00655423803254962, + 0.015286379493772984, + 0.018694672733545303, + 0.012203855440020561, + 0.01111776102334261, + 0.007951933890581131, + 0.01298784464597702, + 0.016987863928079605, + 0.003254730487242341, + 0.003855752293020487, + 0.004316932987421751, + 0.002560214838013053, + 0.0034439866431057453, + 0.002772828796878457, + 0.0036505169700831175, + 0.002877268008887768, + 0.007844564504921436, + 0.007796524092555046, + 0.009240581654012203, + 0.009615839459002018, + 0.007938691414892673, + 0.007852395996451378, + 0.011511200107634068, + 0.00902668945491314, + 0.0053016794845461845, + 0.00902746245265007, + 0.017307456582784653, + 0.010539999231696129, + 0.004793998785316944, + 0.004627360496670008, + 0.0044674016535282135, + 0.005681173875927925, + 0.005428735166788101, + 0.0063450997695326805, + 0.005909097846597433, + 0.005685010459274054, + 0.00555384298786521, + 0.003576402785256505, + 0.005444263108074665, + 0.003984937909990549, + 0.01005647424608469, + 0.012154881842434406, + 0.010974753648042679, + 0.011986901052296162, + 0.022430911660194397, + 0.01196670439094305, + 0.013774988241493702, + 0.01710151694715023, + 0.014694638550281525, + 0.011952958069741726, + 0.012721051461994648, + 0.010351229459047318, + 0.015520019456744194, + 0.008622467517852783, + 0.016254721209406853, + 0.00896545872092247, + 0.04479262977838516, + 0.014526036567986012, + 0.03241981938481331, + 0.015428842045366764, + 0.01661060005426407, + 0.022565562278032303, + 0.011254001408815384, + 0.0421958826482296, + 0.011860677972435951, + 0.007034990470856428, + 0.09887150675058365, + 0.006546475924551487, + 0.009618406184017658, + 0.10996914654970169, + 0.004810607060790062, + 0.019351253286004066, + 0.0024024692829698324, + 0.0024457736872136593, + 0.0016108053969219327, + 0.0015174811705946922, + 0.0037133502773940563, + 0.002137188334017992, + 0.003356943605467677, + 0.0024126407224684954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.026381026953458786 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038131266832351685 + ], + "opposite_logprob_tokens": [ + -3.2881312370300293 + ], + "logprob_reward": -0.038131266832351685, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038131266832351685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32417014241218567, + 0.03388648107647896, + 0.0827215388417244, + 0.07730457931756973, + 0.057742808014154434, + 0.04692805930972099, + 0.17725926637649536, + 0.13453683257102966, + 0.04010312259197235, + 0.025347180664539337 + ], + "credit_map": [ + 0.017451444640755653, + 0.04899408668279648, + 0.05784399434924126, + 0.045397598296403885, + 0.04432723671197891, + 0.02839992567896843, + 0.042053356766700745, + 0.039702486246824265, + 0.0036212708801031113, + 0.0036835770588368177, + 0.0035978308878839016, + 0.002417599782347679, + 0.0040049925446510315, + 0.005351428408175707, + 0.006100801285356283, + 0.005108978599309921, + 0.01118634082376957, + 0.008890196681022644, + 0.018324678763747215, + 0.009831049479544163, + 0.007231372408568859, + 0.009180045686662197, + 0.009949621744453907, + 0.008128227666020393, + 0.007404079660773277, + 0.016811467707157135, + 0.012740184552967548, + 0.012542616575956345, + 0.007139771711081266, + 0.005909342784434557, + 0.007240571081638336, + 0.0075165461748838425, + 0.007158261723816395, + 0.008539061062037945, + 0.007991557940840721, + 0.00931855384260416, + 0.006281909067183733, + 0.005203397013247013, + 0.008456220850348473, + 0.0047938465140759945, + 0.0070810397155582905, + 0.003570618573576212, + 0.006607433315366507, + 0.0042136237025260925, + 0.005236412398517132, + 0.008701488375663757, + 0.004700294230133295, + 0.006817151792347431, + 0.01838402822613716, + 0.01791451871395111, + 0.016799578443169594, + 0.018836429342627525, + 0.036031194031238556, + 0.01724688522517681, + 0.038116954267024994, + 0.013929675333201885, + 0.03334593027830124, + 0.007344632409512997, + 0.023152418434619904, + 0.005905239377170801, + 0.008289407938718796, + 0.016729583963751793, + 0.005837470293045044, + 0.03393213450908661, + 0.0027284864336252213, + 0.003026757389307022, + 0.01102230604737997, + 0.0016366993077099323, + 0.0026322107296437025, + 0.012617688626050949, + 0.0015581295592710376, + 0.004880842287093401, + 0.002607761649414897, + 0.0030099039431661367, + 0.0020255055278539658, + 0.0020720085594803095, + 0.004672873765230179, + 0.002710500033572316, + 0.005165423266589642, + 0.0030832020565867424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.038131266832351685 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200064077973366 + ], + "opposite_logprob_tokens": [ + -4.018199920654297 + ], + "logprob_reward": -0.018200064077973366, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018200064077973366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07097821682691574, + 0.04846961423754692, + 0.11866002529859543, + 0.05581771954894066, + 0.04245981201529503, + 0.028234394267201424, + 0.072491355240345, + 0.40820977091789246, + 0.11980196833610535, + 0.03487715497612953 + ], + "credit_map": [ + 0.004394119139760733, + 0.012565468437969685, + 0.009034812450408936, + 0.01073009055107832, + 0.012728353030979633, + 0.005501193925738335, + 0.008252350613474846, + 0.0077718328684568405, + 0.005939811002463102, + 0.005604711826890707, + 0.007854255847632885, + 0.003804583102464676, + 0.006598719395697117, + 0.005703013390302658, + 0.006697369273751974, + 0.006267151795327663, + 0.01633983664214611, + 0.017932100221514702, + 0.018875012174248695, + 0.031616128981113434, + 0.006725164130330086, + 0.007465833332389593, + 0.010722680948674679, + 0.008983268402516842, + 0.0048133982345461845, + 0.007112649269402027, + 0.013889491558074951, + 0.00878885667771101, + 0.005022656172513962, + 0.004720534663647413, + 0.005406764801591635, + 0.006063371431082487, + 0.005210255738347769, + 0.006619072053581476, + 0.006276249419897795, + 0.006222872994840145, + 0.004917084239423275, + 0.0036105767358094454, + 0.005807322915643454, + 0.0037963753566145897, + 0.0036659506149590015, + 0.0026564777363091707, + 0.00324353389441967, + 0.0024437885731458664, + 0.005928632337599993, + 0.003271003719419241, + 0.003716391045600176, + 0.0033086175099015236, + 0.011204127222299576, + 0.0076105413027107716, + 0.009346039965748787, + 0.008817055262625217, + 0.011524233967065811, + 0.0064249481074512005, + 0.011665116995573044, + 0.005899291951209307, + 0.09808683395385742, + 0.026400215923786163, + 0.0668352022767067, + 0.023171955719590187, + 0.027999650686979294, + 0.04725494980812073, + 0.020832160487771034, + 0.09762880951166153, + 0.007851235568523407, + 0.0057310136035084724, + 0.026795823127031326, + 0.004977429751306772, + 0.004921252839267254, + 0.03149271756410599, + 0.004309571348130703, + 0.033722929656505585, + 0.004604535177350044, + 0.00481245806440711, + 0.0036529977805912495, + 0.003141655120998621, + 0.00537382485345006, + 0.0044127400033175945, + 0.004822309594601393, + 0.004056632984429598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.018200064077973366 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Muay Thai fighter's muscular calf covered in sweat beads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the action a calf lashing out to strike a suspended sepak takraw ball?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sepak takraw ball visibly deforming upon impact with the leg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sweat beads flung from the leg and shown splattering into light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up from an extremely low angle with a shaky handheld camera feel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059713751077652 + ], + "opposite_logprob_tokens": [ + -3.895597219467163 + ], + "logprob_reward": -0.02059713751077652, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059713751077652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1354776918888092, + 0.03770679980516434, + 0.030973711982369423, + 0.031085046008229256, + 0.0594467893242836, + 0.03897963836789131, + 0.054217636585235596, + 0.49398231506347656, + 0.10281585901975632, + 0.01531448494642973 + ], + "credit_map": [ + 0.009089941158890724, + 0.028635947033762932, + 0.02660052850842476, + 0.01869037002325058, + 0.021446244791150093, + 0.008293516002595425, + 0.0114872045814991, + 0.011233935132622719, + 0.004069515969604254, + 0.004548702854663134, + 0.00498279370367527, + 0.0023229329381138086, + 0.003984251990914345, + 0.006121842656284571, + 0.0062668644823133945, + 0.005409894045442343, + 0.004604771733283997, + 0.004988628905266523, + 0.005364470183849335, + 0.007673637475818396, + 0.001670928904786706, + 0.0017448909347876906, + 0.002791985170915723, + 0.002134398091584444, + 0.0027492123190313578, + 0.005189931020140648, + 0.0061961752362549305, + 0.007374197710305452, + 0.002353131305426359, + 0.002293936675414443, + 0.002650714013725519, + 0.0022777460981160402, + 0.008788194507360458, + 0.010980705730617046, + 0.010470579378306866, + 0.010239041410386562, + 0.006562608759850264, + 0.0031396171543747187, + 0.006101394537836313, + 0.0031646527349948883, + 0.0057635679841041565, + 0.0058891004882752895, + 0.006010298617184162, + 0.005375024396926165, + 0.006093932315707207, + 0.0037397723644971848, + 0.0041996389627456665, + 0.001908304519020021, + 0.009112670086324215, + 0.006018369924277067, + 0.007112368009984493, + 0.005155303981155157, + 0.008961879648268223, + 0.004572802688926458, + 0.008557161316275597, + 0.004727078601717949, + 0.11679428070783615, + 0.03325842693448067, + 0.07978757470846176, + 0.03075443021953106, + 0.03702159598469734, + 0.05493282154202461, + 0.027128420770168304, + 0.11430475860834122, + 0.008614011108875275, + 0.00894372258335352, + 0.016644813120365143, + 0.007743990980088711, + 0.007612315006554127, + 0.021679842844605446, + 0.007506910711526871, + 0.02407025545835495, + 0.0012858284171670675, + 0.001465739798732102, + 0.00150326790753752, + 0.0013854533899575472, + 0.0027181203477084637, + 0.0020810659043490887, + 0.0029468913562595844, + 0.0019281173590570688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.02059713751077652 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_10/index_tensor([5869, 5869])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_10/index_tensor([5869, 5869])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7698cacdc8ad50e4d12c9d190af7bd6bf3541085 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_10/index_tensor([5869, 5869])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tightly static, horizontal close-up focuses on the round headlight of a vintage car, captured in a stylized, high-contrast black and white image. Its raised glass lens is surrounded by a gleaming chrome bezel, which reflects sharp, dazzling highlights, creating a dramatic contrast with the deep, velvety shadows of the dark, polished metal surrounding it. The headlight, motionless and resolute, projects a powerful, focused beam forward, piercing the endless blackness of the frame. Within the cone of light, tiny dust particles suspended in the air dance slowly, lending the light itself a tangible quality. The scene is hushed and imbued with the suspense of film noir and a sense of eternity frozen in time, as if the car is quietly awaiting something in a forgotten corner.", + "seed": 12388, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_10/index_5869_seed_12388_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01608839150285348, + "vlm_logprob": -0.01608839150285348, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.01608839150285348, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035832952708005905 + ], + "opposite_logprob_tokens": [ + -3.7858328819274902 + ], + "logprob_reward": -0.035832952708005905, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.035832952708005905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3218490183353424, + 0.1337173581123352, + 0.03314194828271866, + 0.027676569297909737, + 0.07764387130737305, + 0.08819331228733063, + 0.15045149624347687, + 0.09212491661310196, + 0.05224689096212387, + 0.022954601794481277 + ], + "credit_map": [ + 0.049877554178237915, + 0.022603310644626617, + 0.04283647611737251, + 0.023509640246629715, + 0.01986185647547245, + 0.08080866187810898, + 0.018984468653798103, + 0.0633670836687088, + 0.022902000695466995, + 0.007656667847186327, + 0.01865886151790619, + 0.00656411936506629, + 0.00753659475594759, + 0.03279748186469078, + 0.007494171150028706, + 0.030107473954558372, + 0.005714787170290947, + 0.0028243365231901407, + 0.004456935916095972, + 0.0029258194845169783, + 0.0025767870247364044, + 0.006782813463360071, + 0.002266567200422287, + 0.0055939024314284325, + 0.004125749692320824, + 0.0020266184583306313, + 0.004035537131130695, + 0.002393157919868827, + 0.0019089799607172608, + 0.006457946263253689, + 0.00185208220500499, + 0.004876500461250544, + 0.010608134791254997, + 0.005626590922474861, + 0.012815448455512524, + 0.0076097771525382996, + 0.004433073103427887, + 0.016477711498737335, + 0.005030849017202854, + 0.01504229474812746, + 0.01173631101846695, + 0.007907282561063766, + 0.011683419346809387, + 0.008161209523677826, + 0.005890092812478542, + 0.020796386525034904, + 0.005532731767743826, + 0.01648588851094246, + 0.019272740930318832, + 0.01245182380080223, + 0.02225508540868759, + 0.014822022058069706, + 0.010441312566399574, + 0.03304717689752579, + 0.01110343262553215, + 0.027057908475399017, + 0.010785428807139397, + 0.00872910674661398, + 0.01131256204098463, + 0.009136958979070187, + 0.006744866259396076, + 0.020312808454036713, + 0.007728148717433214, + 0.0173750389367342, + 0.0074938456527888775, + 0.003509851871058345, + 0.007460481952875853, + 0.004121682606637478, + 0.002948162844404578, + 0.012870599515736103, + 0.003218567930161953, + 0.010623705573379993, + 0.0031432188116014004, + 0.0023901225067675114, + 0.0021429513581097126, + 0.0026520013343542814, + 0.0013359561562538147, + 0.0056526330299675465, + 0.0013310372596606612, + 0.004306682851165533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.035832952708005905 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017469158628955483 + ], + "opposite_logprob_tokens": [ + -6.376747131347656 + ], + "logprob_reward": -0.0017469158628955483, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017469158628955483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28537416458129883, + 0.14527088403701782, + 0.06860005855560303, + 0.05668624863028526, + 0.05338522791862488, + 0.07328677922487259, + 0.09577474743127823, + 0.0833757072687149, + 0.07400427758693695, + 0.06424194574356079 + ], + "credit_map": [ + 0.034831929951906204, + 0.024281147867441177, + 0.049303922802209854, + 0.02766108140349388, + 0.019156239926815033, + 0.06816334277391434, + 0.018231458961963654, + 0.04374503716826439, + 0.019689472392201424, + 0.017278769984841347, + 0.015109626576304436, + 0.019697027280926704, + 0.011535395868122578, + 0.027839727699756622, + 0.011741158552467823, + 0.02237970009446144, + 0.010699694976210594, + 0.005872352514415979, + 0.010200394317507744, + 0.0062102689407765865, + 0.0053566452115774155, + 0.014164301566779613, + 0.005101443268358707, + 0.010994956828653812, + 0.010226097889244556, + 0.005556048359721899, + 0.009469503536820412, + 0.004973440431058407, + 0.004272854421287775, + 0.009352769702672958, + 0.004397741053253412, + 0.008437790907919407, + 0.005998789332807064, + 0.004104346502572298, + 0.009114124812185764, + 0.004214563872665167, + 0.0031302468851208687, + 0.012599914334714413, + 0.004120904020965099, + 0.010102333500981331, + 0.006679926998913288, + 0.008507697843015194, + 0.007172571495175362, + 0.010130313225090504, + 0.004885772243142128, + 0.017079878598451614, + 0.005060954485088587, + 0.013769660145044327, + 0.011622984893620014, + 0.010860856622457504, + 0.009386549703776836, + 0.011429933831095695, + 0.005986378528177738, + 0.020722784101963043, + 0.007421244867146015, + 0.018344009295105934, + 0.009020524099469185, + 0.008725373074412346, + 0.01309565082192421, + 0.008974333293735981, + 0.006529814563691616, + 0.01737171970307827, + 0.007199314422905445, + 0.012458966113626957, + 0.011146953329443932, + 0.006302570458501577, + 0.010369761846959591, + 0.005752923432737589, + 0.004795567598193884, + 0.018473660573363304, + 0.00460185157135129, + 0.012560982257127762, + 0.007978707551956177, + 0.007525013759732246, + 0.007887190207839012, + 0.009519710205495358, + 0.0051542045548558235, + 0.010963977314531803, + 0.005185853224247694, + 0.010027281939983368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0017469158628955483 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029814058914780617 + ], + "opposite_logprob_tokens": [ + -3.5298140048980713 + ], + "logprob_reward": -0.029814058914780617, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029814058914780617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46240392327308655, + 0.1205267459154129, + 0.05068681761622429, + 0.0732293426990509, + 0.06648112088441849, + 0.04963436722755432, + 0.052657365798950195, + 0.04198780655860901, + 0.039312995970249176, + 0.04307945817708969 + ], + "credit_map": [ + 0.08172865211963654, + 0.02915208227932453, + 0.07106661796569824, + 0.03462279960513115, + 0.02863701619207859, + 0.10168084502220154, + 0.0265573188662529, + 0.08895860612392426, + 0.015524971298873425, + 0.01219527330249548, + 0.013988513499498367, + 0.016890665516257286, + 0.007616815622895956, + 0.023898644372820854, + 0.011180896311998367, + 0.019230972975492477, + 0.00777340866625309, + 0.003328692866489291, + 0.008730095811188221, + 0.0033363623078912497, + 0.002755095949396491, + 0.012314382940530777, + 0.0038726208731532097, + 0.008576161228120327, + 0.009561720304191113, + 0.007076735142618418, + 0.008835827000439167, + 0.009023604914546013, + 0.005057823844254017, + 0.015414397232234478, + 0.005550957750529051, + 0.012708280235528946, + 0.007925753481686115, + 0.005266029387712479, + 0.011346067301928997, + 0.005033980123698711, + 0.004866162780672312, + 0.015424801036715508, + 0.005072322208434343, + 0.011546006426215172, + 0.006048440933227539, + 0.003530127927660942, + 0.007150685880333185, + 0.004814367275685072, + 0.0030899858102202415, + 0.012307717464864254, + 0.003306241473183036, + 0.009386797435581684, + 0.005365725141018629, + 0.004958615638315678, + 0.007165734656155109, + 0.00431527616456151, + 0.00389618962071836, + 0.013263524509966373, + 0.004160255193710327, + 0.00953204557299614, + 0.003586534643545747, + 0.004636805504560471, + 0.006844951305538416, + 0.004622063599526882, + 0.003386734751984477, + 0.009210902266204357, + 0.003980991896241903, + 0.0057188221253454685, + 0.005551706068217754, + 0.0038737496361136436, + 0.0063383448868989944, + 0.004266819916665554, + 0.00297165266238153, + 0.007534147705882788, + 0.0030733002349734306, + 0.005703274626284838, + 0.007682782597839832, + 0.002757496200501919, + 0.004804032854735851, + 0.0028890788089483976, + 0.0026215112302452326, + 0.010378207080066204, + 0.002235827501863241, + 0.009710521437227726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.029814058914780617 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254771463572979 + ], + "opposite_logprob_tokens": [ + -4.387547492980957 + ], + "logprob_reward": -0.01254771463572979, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254771463572979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32060757279396057, + 0.10915940999984741, + 0.08723936975002289, + 0.05985134840011597, + 0.08048103004693985, + 0.0988389402627945, + 0.11444747447967529, + 0.06276103854179382, + 0.033790938556194305, + 0.03282284364104271 + ], + "credit_map": [ + 0.03306441754102707, + 0.03756903484463692, + 0.046060960739851, + 0.04368295520544052, + 0.02146814949810505, + 0.06594209372997284, + 0.021380741149187088, + 0.05143921822309494, + 0.01442971546202898, + 0.010328170843422413, + 0.0190038550645113, + 0.014953714795410633, + 0.005961020011454821, + 0.019381649792194366, + 0.008245548233389854, + 0.01685573346912861, + 0.014326067641377449, + 0.005420931149274111, + 0.012362184934318066, + 0.006083069834858179, + 0.004141501151025295, + 0.021166397258639336, + 0.005460898391902447, + 0.01827831007540226, + 0.006277813110500574, + 0.00570544321089983, + 0.007311862427741289, + 0.005299717653542757, + 0.0031175301410257816, + 0.0161196980625391, + 0.0036866427399218082, + 0.012332641519606113, + 0.010206959210336208, + 0.00828971154987812, + 0.011382581666111946, + 0.009343467652797699, + 0.005106810014694929, + 0.017339587211608887, + 0.005989826750010252, + 0.012822084128856659, + 0.008993430063128471, + 0.010407284833490849, + 0.009354567155241966, + 0.0119401216506958, + 0.005996947642415762, + 0.02537691220641136, + 0.006058963015675545, + 0.020710714161396027, + 0.013755694963037968, + 0.011769386008381844, + 0.015924783423542976, + 0.010563031770288944, + 0.009202486835420132, + 0.028223110362887383, + 0.009680760093033314, + 0.015328221023082733, + 0.007517318241298199, + 0.007328161504119635, + 0.008051942102611065, + 0.006834161002188921, + 0.003652225947007537, + 0.013791846111416817, + 0.0042055100202560425, + 0.011379876174032688, + 0.0052515300922095776, + 0.002985818311572075, + 0.003956449218094349, + 0.0037385441828519106, + 0.002643296029418707, + 0.007582615129649639, + 0.0021371571347117424, + 0.005495526362210512, + 0.004806263372302055, + 0.004375155549496412, + 0.00266003655269742, + 0.004595579579472542, + 0.0022461514454334974, + 0.006859907414764166, + 0.0020383172668516636, + 0.005241429433226585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.01254771463572979 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005003153928555548 + ], + "opposite_logprob_tokens": [ + -7.625500202178955 + ], + "logprob_reward": -0.0005003153928555548, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0005003153928555548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33552801609039307, + 0.08583146333694458, + 0.037068258970975876, + 0.04323119670152664, + 0.08507432043552399, + 0.11828142404556274, + 0.12032407522201538, + 0.07641977071762085, + 0.04800226911902428, + 0.050239112228155136 + ], + "credit_map": [ + 0.04423484951257706, + 0.029994267970323563, + 0.04407058283686638, + 0.032310619950294495, + 0.01863388530910015, + 0.07948458194732666, + 0.027588529512286186, + 0.059210751205682755, + 0.011478639207780361, + 0.007367048412561417, + 0.01156679168343544, + 0.008754830807447433, + 0.005276317708194256, + 0.01972345821559429, + 0.006676481105387211, + 0.014987904578447342, + 0.005086694844067097, + 0.0034581839572638273, + 0.004186803475022316, + 0.003950407262891531, + 0.002640040125697851, + 0.008241225965321064, + 0.0029189945198595524, + 0.006585913710296154, + 0.008392744697630405, + 0.0034286300651729107, + 0.005266403779387474, + 0.003402003087103367, + 0.0028439394664019346, + 0.009821546263992786, + 0.0026247960049659014, + 0.0074511365965008736, + 0.010221113450825214, + 0.007079554256051779, + 0.012379297986626625, + 0.009439623914659023, + 0.005623010918498039, + 0.01810739003121853, + 0.006513151340186596, + 0.015711188316345215, + 0.00924753025174141, + 0.012464223429560661, + 0.010299472138285637, + 0.014499149285256863, + 0.00808792281895876, + 0.030459482222795486, + 0.00744802039116621, + 0.02577563002705574, + 0.017248094081878662, + 0.010035740211606026, + 0.01666250452399254, + 0.011696460656821728, + 0.007890810258686543, + 0.025745298713445663, + 0.009192818775773048, + 0.02185235545039177, + 0.007152134552598, + 0.007635785266757011, + 0.011085709556937218, + 0.007178041618317366, + 0.0060501256957650185, + 0.017565341666340828, + 0.006576118990778923, + 0.013176524080336094, + 0.006470164749771357, + 0.00511972326785326, + 0.007475558202713728, + 0.004664183594286442, + 0.002742979908362031, + 0.010070149786770344, + 0.0030783878173679113, + 0.008381125517189503, + 0.005612189881503582, + 0.004053896758705378, + 0.0061014872044324875, + 0.003545809304341674, + 0.0034162981901317835, + 0.014179868623614311, + 0.0032102407421916723, + 0.010119320824742317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0005003153928555548 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tightly static, horizontal close-up focuses on the round headlight of a vintage car, captured in a stylized, high-contrast black and white image. Its raised glass lens is surrounded by a gleaming chrome bezel, which reflects sharp, dazzling highlights, creating a dramatic contrast with the deep, velvety shadows of the dark, polished metal surrounding it. The headlight, motionless and resolute, projects a powerful, focused beam forward, piercing the endless blackness of the frame. Within the cone of light, tiny dust particles suspended in the air dance slowly, lending the light itself a tangible quality. The scene is hushed and imbued with the suspense of film noir and a sense of eternity frozen in time, as if the car is quietly awaiting something in a forgotten corner.", + "seed": 12389, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_10/index_5869_seed_12389_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008421216055285185, + "vlm_logprob": -0.008421216055285185, + "vlm_yes_no_margin": 5.575, + "vlm_token_credit": -0.008421216055285185, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016173940151929855 + ], + "opposite_logprob_tokens": [ + -4.391173839569092 + ], + "logprob_reward": -0.016173940151929855, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016173940151929855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3669992983341217, + 0.11174455285072327, + 0.03293086960911751, + 0.04715218022465706, + 0.08095203340053558, + 0.04254405200481415, + 0.10229583084583282, + 0.06403274834156036, + 0.08295596390962601, + 0.06839243322610855 + ], + "credit_map": [ + 0.03437599539756775, + 0.03288351744413376, + 0.06519610434770584, + 0.05088622495532036, + 0.026380980387330055, + 0.05508194863796234, + 0.03653110936284065, + 0.0656634196639061, + 0.011532390490174294, + 0.008996971882879734, + 0.0199317317456007, + 0.016001731157302856, + 0.008575915358960629, + 0.016560334712266922, + 0.009089886210858822, + 0.021055588498711586, + 0.0031591635197401047, + 0.002732193563133478, + 0.005758629646152258, + 0.005039528477936983, + 0.002464544726535678, + 0.004780421033501625, + 0.002929402282461524, + 0.006066986359655857, + 0.003655251581221819, + 0.004933088552206755, + 0.0051670684479177, + 0.0076993429102003574, + 0.003963095135986805, + 0.0073953899554908276, + 0.005718344356864691, + 0.00862059835344553, + 0.012314596213400364, + 0.0050201923586428165, + 0.011084499768912792, + 0.007346968166530132, + 0.006141188554465771, + 0.012862371280789375, + 0.006063770968466997, + 0.020118437707424164, + 0.0038566729053854942, + 0.004154483322054148, + 0.005035615526139736, + 0.006514302454888821, + 0.0029859496280550957, + 0.006711895577609539, + 0.004148417618125677, + 0.009136714041233063, + 0.017549807205796242, + 0.0072325156070292, + 0.011251047253608704, + 0.011990996077656746, + 0.009984814561903477, + 0.015393555164337158, + 0.010924686677753925, + 0.01796841062605381, + 0.008229962550103664, + 0.006033469457179308, + 0.008628769777715206, + 0.00862890761345625, + 0.006520056631416082, + 0.009124482050538063, + 0.007907463237643242, + 0.008959630504250526, + 0.012219457887113094, + 0.005243601277470589, + 0.011495228856801987, + 0.007287970278412104, + 0.006932889577001333, + 0.012807232327759266, + 0.007990291342139244, + 0.01897929236292839, + 0.010028704069554806, + 0.004716466646641493, + 0.008570155128836632, + 0.00731333252042532, + 0.005349456332623959, + 0.010784345678985119, + 0.005617274437099695, + 0.016012703999876976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.016173940151929855 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019593106117099524 + ], + "opposite_logprob_tokens": [ + -6.251959323883057 + ], + "logprob_reward": -0.0019593106117099524, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019593106117099524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21425075829029083, + 0.09089764952659607, + 0.08151514083147049, + 0.048410505056381226, + 0.19847075641155243, + 0.029320644214749336, + 0.13135342299938202, + 0.04131815582513809, + 0.10018516331911087, + 0.06427773088216782 + ], + "credit_map": [ + 0.024395208805799484, + 0.020757779479026794, + 0.036479245871305466, + 0.028857072815299034, + 0.02054123394191265, + 0.030540738254785538, + 0.023128541186451912, + 0.029550939798355103, + 0.00841896515339613, + 0.008170418441295624, + 0.014107274822890759, + 0.010500611737370491, + 0.007900486700236797, + 0.012850786559283733, + 0.011949333362281322, + 0.01699976995587349, + 0.007765788119286299, + 0.007507143542170525, + 0.01565769873559475, + 0.008492756634950638, + 0.006986220367252827, + 0.012180382385849953, + 0.008147278800606728, + 0.01477786898612976, + 0.005641793366521597, + 0.005274042021483183, + 0.0075205061584711075, + 0.006080280523747206, + 0.004483419936150312, + 0.006102472078055143, + 0.00660336297005415, + 0.006704629398882389, + 0.027851464226841927, + 0.012647244147956371, + 0.021732725203037262, + 0.01737150363624096, + 0.01814945787191391, + 0.03063781186938286, + 0.020595373585820198, + 0.04948516562581062, + 0.0027311109006404877, + 0.0034942806232720613, + 0.004396625328809023, + 0.003665849333629012, + 0.003613031469285488, + 0.004041681531816721, + 0.004144350998103619, + 0.00323371565900743, + 0.01840926893055439, + 0.008834806270897388, + 0.01474519819021225, + 0.013294955715537071, + 0.011251629330217838, + 0.020602060481905937, + 0.014104869216680527, + 0.030110644176602364, + 0.004521356429904699, + 0.0036673187278211117, + 0.007573265116661787, + 0.00390738295391202, + 0.003800775855779648, + 0.005484320688992739, + 0.006288416683673859, + 0.006075318902730942, + 0.017785917967557907, + 0.006867230404168367, + 0.012558862566947937, + 0.009918240830302238, + 0.0072603290900588036, + 0.015803616493940353, + 0.00825362466275692, + 0.021737342700362206, + 0.007512154057621956, + 0.005606353282928467, + 0.008842968381941319, + 0.007316777016967535, + 0.006684974767267704, + 0.009020240977406502, + 0.008644923567771912, + 0.010649340227246284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0019593106117099524 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028208012226969004 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028208012226969004, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028208012226969004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35759755969047546, + 0.10425516217947006, + 0.07668274641036987, + 0.04646814614534378, + 0.09692884236574173, + 0.03820539265871048, + 0.0493522472679615, + 0.0663720890879631, + 0.10225086659193039, + 0.06188692897558212 + ], + "credit_map": [ + 0.06518017500638962, + 0.02927286922931671, + 0.0391930416226387, + 0.0571308396756649, + 0.023816414177417755, + 0.054224442690610886, + 0.03654414787888527, + 0.05223561078310013, + 0.010006848722696304, + 0.012056700885295868, + 0.016184164211153984, + 0.01781807281076908, + 0.008223207667469978, + 0.015130010433495045, + 0.01079505030065775, + 0.014041104353964329, + 0.007105889264494181, + 0.008507011458277702, + 0.016855433583259583, + 0.012046980671584606, + 0.005349045153707266, + 0.011626709252595901, + 0.005830121226608753, + 0.00936155766248703, + 0.005134627688676119, + 0.00433245999738574, + 0.008979531936347485, + 0.006741034798324108, + 0.0044170403853058815, + 0.006237234454602003, + 0.004488409496843815, + 0.006137805059552193, + 0.01090770959854126, + 0.007246433291584253, + 0.015474042855203152, + 0.009443680755794048, + 0.008080790750682354, + 0.014865785837173462, + 0.008866391144692898, + 0.02204400673508644, + 0.0031463387422263622, + 0.0034003444015979767, + 0.005473783239722252, + 0.005393876228481531, + 0.004108398221433163, + 0.005868745967745781, + 0.004377170465886593, + 0.006436737719923258, + 0.004929953254759312, + 0.005162978079169989, + 0.005050713196396828, + 0.008301000110805035, + 0.004666514229029417, + 0.0074423979967832565, + 0.006561763118952513, + 0.007236925885081291, + 0.008270327001810074, + 0.0038998553063720465, + 0.007462762296199799, + 0.006998418364673853, + 0.005009916145354509, + 0.01076525915414095, + 0.0076629589311778545, + 0.016302594915032387, + 0.018597794696688652, + 0.006542915012687445, + 0.01178293488919735, + 0.010037391446530819, + 0.007072633132338524, + 0.016743922606110573, + 0.008495728485286236, + 0.022977545857429504, + 0.007550246547907591, + 0.004261802416294813, + 0.009911414235830307, + 0.006178754847496748, + 0.0038586279843002558, + 0.009803690947592258, + 0.0045880586840212345, + 0.01573433168232441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0028208012226969004 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585110411047935 + ], + "opposite_logprob_tokens": [ + -3.895585060119629 + ], + "logprob_reward": -0.020585110411047935, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585110411047935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3039873540401459, + 0.11397811025381088, + 0.05427734553813934, + 0.07809244841337204, + 0.08375860750675201, + 0.1212005466222763, + 0.10322121530771255, + 0.07755271345376968, + 0.036118634045124054, + 0.027813013643026352 + ], + "credit_map": [ + 0.046918731182813644, + 0.028595540672540665, + 0.034433797001838684, + 0.04399251937866211, + 0.020769476890563965, + 0.04563184827566147, + 0.030928803607821465, + 0.05271662771701813, + 0.012189304456114769, + 0.01091979444026947, + 0.018024712800979614, + 0.02006492391228676, + 0.0071328505873680115, + 0.01777830347418785, + 0.010812492109835148, + 0.017055729404091835, + 0.005978094879537821, + 0.006741954013705254, + 0.011740815825760365, + 0.009119918569922447, + 0.0031679633539170027, + 0.007703372742980719, + 0.003291433909907937, + 0.006533794570714235, + 0.008400307968258858, + 0.005673430860042572, + 0.014496257528662682, + 0.010774275287985802, + 0.006373929791152477, + 0.011777278035879135, + 0.007641862146556377, + 0.01295511331409216, + 0.008381457068026066, + 0.0057008350268006325, + 0.013355684466660023, + 0.009381962940096855, + 0.007821767590939999, + 0.012923447415232658, + 0.009099564515054226, + 0.01709388568997383, + 0.01927531138062477, + 0.008162938058376312, + 0.018511418253183365, + 0.01075047254562378, + 0.008965438231825829, + 0.019484473392367363, + 0.009831948205828667, + 0.02621854841709137, + 0.014381048269569874, + 0.01044908631592989, + 0.014335380867123604, + 0.012033537961542606, + 0.0069567481987178326, + 0.01653369702398777, + 0.007004115264862776, + 0.021527595818042755, + 0.008627260103821754, + 0.004456107970327139, + 0.011107047088444233, + 0.008177374489605427, + 0.0059989155270159245, + 0.012726930901408195, + 0.007256999611854553, + 0.01920207589864731, + 0.004667477682232857, + 0.0030544130131602287, + 0.003168955212458968, + 0.005545116029679775, + 0.004418672528117895, + 0.004783618729561567, + 0.004968022927641869, + 0.005512359086424112, + 0.003611701773479581, + 0.0020205937325954437, + 0.005029814317822456, + 0.003780367085710168, + 0.002120164455845952, + 0.0038628627080470324, + 0.002777017652988434, + 0.0046104928478598595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.020585110411047935 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005669178790412843 + ], + "opposite_logprob_tokens": [ + -7.5005669593811035 + ], + "logprob_reward": -0.0005669178790412843, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005669178790412843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4061616063117981, + 0.12052621692419052, + 0.059751246124506, + 0.052116528153419495, + 0.05357040837407112, + 0.051725421100854874, + 0.07834041863679886, + 0.054705895483493805, + 0.040955882519483566, + 0.08214626461267471 + ], + "credit_map": [ + 0.04857116937637329, + 0.04014182463288307, + 0.05745233967900276, + 0.059469595551490784, + 0.031253691762685776, + 0.06379637122154236, + 0.03225022926926613, + 0.0732264444231987, + 0.01287129707634449, + 0.01262867171317339, + 0.022785214707255363, + 0.020524408668279648, + 0.007467865478247404, + 0.01930384524166584, + 0.00843631662428379, + 0.016508614644408226, + 0.007275862153619528, + 0.005465259309858084, + 0.007308096159249544, + 0.009707314893603325, + 0.0046126884408295155, + 0.009807835333049297, + 0.005015576258301735, + 0.010558614507317543, + 0.008477186784148216, + 0.0043047633953392506, + 0.007030044216662645, + 0.007089560851454735, + 0.002517042448744178, + 0.009209542535245419, + 0.0027279576752334833, + 0.01076043676584959, + 0.006549579557031393, + 0.006638647057116032, + 0.00946966651827097, + 0.00937853753566742, + 0.004064260050654411, + 0.007048742845654488, + 0.004836545325815678, + 0.0055844346061348915, + 0.004098521079868078, + 0.003921222407370806, + 0.006607235409319401, + 0.007701347582042217, + 0.0032786705996841192, + 0.00854509323835373, + 0.005380682181566954, + 0.01219265442341566, + 0.010600029490888119, + 0.006261556874960661, + 0.00840738881379366, + 0.011682369746267796, + 0.007412129547446966, + 0.012000774964690208, + 0.009872754104435444, + 0.012103424407541752, + 0.006263528019189835, + 0.004130116198211908, + 0.006201610900461674, + 0.0064160930924117565, + 0.005394247360527515, + 0.008577335625886917, + 0.006424546241760254, + 0.011298424564301968, + 0.006290982011705637, + 0.0038112220354378223, + 0.00409274036064744, + 0.007076213601976633, + 0.0028834312688559294, + 0.005918289069086313, + 0.004200434312224388, + 0.0066825756803154945, + 0.010267344303429127, + 0.005202688276767731, + 0.010819974355399609, + 0.008569265715777874, + 0.006650107447057962, + 0.012874795123934746, + 0.008580658584833145, + 0.019181443378329277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0005669178790412843 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tightly static, horizontal close-up focuses on the round headlight of a vintage car, captured in a stylized, high-contrast black and white image. Its raised glass lens is surrounded by a gleaming chrome bezel, which reflects sharp, dazzling highlights, creating a dramatic contrast with the deep, velvety shadows of the dark, polished metal surrounding it. The headlight, motionless and resolute, projects a powerful, focused beam forward, piercing the endless blackness of the frame. Within the cone of light, tiny dust particles suspended in the air dance slowly, lending the light itself a tangible quality. The scene is hushed and imbued with the suspense of film noir and a sense of eternity frozen in time, as if the car is quietly awaiting something in a forgotten corner.", + "seed": 12390, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_10/index_5869_seed_12390_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01677110253367573, + "vlm_logprob": -0.01677110253367573, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.01677110253367573, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012907763943076134 + ], + "opposite_logprob_tokens": [ + -5.262907981872559 + ], + "logprob_reward": -0.012907763943076134, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.012907763943076134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2343371957540512, + 0.05038444697856903, + 0.049358442425727844, + 0.04583503678441048, + 0.05991867557168007, + 0.1381799876689911, + 0.10569889843463898, + 0.12723861634731293, + 0.07927072793245316, + 0.10977796465158463 + ], + "credit_map": [ + 0.024668892845511436, + 0.026489287614822388, + 0.043241534382104874, + 0.02706625871360302, + 0.019556131213903427, + 0.034651100635528564, + 0.01760653592646122, + 0.04105745255947113, + 0.00643555261194706, + 0.005845880135893822, + 0.009354537352919579, + 0.006545152049511671, + 0.004641390405595303, + 0.007472928147763014, + 0.003863716498017311, + 0.006225287448614836, + 0.005932964384555817, + 0.0033463812433183193, + 0.011489156633615494, + 0.00582628371194005, + 0.002735565882176161, + 0.007525053806602955, + 0.003757015336304903, + 0.008746018633246422, + 0.0055782366544008255, + 0.002897456055507064, + 0.010378016158938408, + 0.004622216802090406, + 0.0033368594013154507, + 0.007442567963153124, + 0.0034910584799945354, + 0.00808862503618002, + 0.006477195303887129, + 0.004186835139989853, + 0.01799209788441658, + 0.00395665317773819, + 0.004191095940768719, + 0.008876077830791473, + 0.0046406956389546394, + 0.009598019532859325, + 0.014097179286181927, + 0.013812938705086708, + 0.030392495915293694, + 0.0202933382242918, + 0.007215964142233133, + 0.019946467131376266, + 0.010029309429228306, + 0.02239230088889599, + 0.011126750148832798, + 0.008036830462515354, + 0.028317080810666084, + 0.006580851972103119, + 0.009280975908041, + 0.015876252204179764, + 0.009031224064528942, + 0.01744893565773964, + 0.011921857483685017, + 0.013310112059116364, + 0.025870483368635178, + 0.017866401001811028, + 0.007851721718907356, + 0.017888598144054413, + 0.011819456703960896, + 0.020709985867142677, + 0.007980463095009327, + 0.004824117291718721, + 0.023175103589892387, + 0.007482417393475771, + 0.006342906039208174, + 0.013211088255047798, + 0.006450651213526726, + 0.009803976863622665, + 0.01207354199141264, + 0.007067698519676924, + 0.02552291564643383, + 0.00566497165709734, + 0.011146477423608303, + 0.01851058006286621, + 0.008240308612585068, + 0.021551473066210747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.012907763943076134 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004659746307879686 + ], + "opposite_logprob_tokens": [ + -5.379659652709961 + ], + "logprob_reward": -0.004659746307879686, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004659746307879686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23124021291732788, + 0.14652585983276367, + 0.08727304637432098, + 0.046093184500932693, + 0.07130048424005508, + 0.08471760153770447, + 0.03009551763534546, + 0.08929284662008286, + 0.055769193917512894, + 0.1576920747756958 + ], + "credit_map": [ + 0.025715520605444908, + 0.018100829795002937, + 0.05970385670661926, + 0.023389682173728943, + 0.0191879253834486, + 0.036149002611637115, + 0.01804516837000847, + 0.030948232859373093, + 0.01628730818629265, + 0.011382503435015678, + 0.04584281146526337, + 0.015107742510735989, + 0.008758357726037502, + 0.024605298414826393, + 0.008181866258382797, + 0.01635996252298355, + 0.009795824997127056, + 0.006154349539428949, + 0.02815661020576954, + 0.005495474208146334, + 0.006050747353583574, + 0.01306387409567833, + 0.005324485711753368, + 0.013231679797172546, + 0.004848307464271784, + 0.002483774209395051, + 0.012259649112820625, + 0.004292073659598827, + 0.0036416826769709587, + 0.007262532599270344, + 0.003213378367945552, + 0.008091787807643414, + 0.008763809688389301, + 0.006439978256821632, + 0.020835857838392258, + 0.005368797108530998, + 0.0057374476455152035, + 0.01179132517427206, + 0.005459150765091181, + 0.006904120557010174, + 0.009064028970897198, + 0.007424638606607914, + 0.023018771782517433, + 0.0055786799639463425, + 0.00700727105140686, + 0.013687062077224255, + 0.005482005886733532, + 0.013455145992338657, + 0.0035830263514071703, + 0.0021490175276994705, + 0.007436621002852917, + 0.002381320344284177, + 0.00205129012465477, + 0.0050659882836043835, + 0.0024054287932813168, + 0.0050228252075612545, + 0.009713632985949516, + 0.006314648315310478, + 0.020424433052539825, + 0.00739109143614769, + 0.005463481415063143, + 0.01423284038901329, + 0.0067739770747721195, + 0.018978741019964218, + 0.005877966061234474, + 0.0038204509764909744, + 0.015792565420269966, + 0.006071019917726517, + 0.0035732092801481485, + 0.007926513440907001, + 0.004190204665064812, + 0.008517260663211346, + 0.016178259626030922, + 0.0157978106290102, + 0.03512340411543846, + 0.012822987511754036, + 0.016474466770887375, + 0.024922823533415794, + 0.01114832516759634, + 0.025224005803465843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.004659746307879686 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498216673731804 + ], + "opposite_logprob_tokens": [ + -2.9299821853637695 + ], + "logprob_reward": -0.05498216673731804, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498216673731804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12556402385234833, + 0.1441798061132431, + 0.04978817701339722, + 0.07873223721981049, + 0.07229296118021011, + 0.10227231681346893, + 0.08302127569913864, + 0.21664351224899292, + 0.028009934350848198, + 0.09949575364589691 + ], + "credit_map": [ + 0.014379266649484634, + 0.0065319957211613655, + 0.03733840584754944, + 0.013480699621140957, + 0.008674063719809055, + 0.018918048590421677, + 0.008527621626853943, + 0.017713917419314384, + 0.014588535763323307, + 0.010812751948833466, + 0.04545808210968971, + 0.008208475075662136, + 0.010065619833767414, + 0.022187361493706703, + 0.011990915052592754, + 0.02086806111037731, + 0.005932544358074665, + 0.003961177542805672, + 0.011729503981769085, + 0.0039639598689973354, + 0.005067689809948206, + 0.006986554712057114, + 0.004765389021486044, + 0.007381357252597809, + 0.008081814274191856, + 0.005024253390729427, + 0.015952592715620995, + 0.005007890984416008, + 0.006016038823872805, + 0.011743430979549885, + 0.007438845466822386, + 0.019467374309897423, + 0.008206133730709553, + 0.005854280665516853, + 0.023216186091303825, + 0.004319099709391594, + 0.004107598215341568, + 0.012009760364890099, + 0.004125514533370733, + 0.01045438926666975, + 0.00928527768701315, + 0.013714902102947235, + 0.017678530886769295, + 0.016019633039832115, + 0.0067243147641420364, + 0.013594959862530231, + 0.008226282894611359, + 0.017028415575623512, + 0.008853857405483723, + 0.007627786137163639, + 0.019760193303227425, + 0.007801895495504141, + 0.007835139520466328, + 0.01196223869919777, + 0.007314471527934074, + 0.011865688487887383, + 0.01968609355390072, + 0.02748745307326317, + 0.030914494767785072, + 0.03985825181007385, + 0.016675425693392754, + 0.03136836737394333, + 0.02206701971590519, + 0.028586406260728836, + 0.002966859843581915, + 0.002018858678638935, + 0.007536653894931078, + 0.0020943651907145977, + 0.0025324993766844273, + 0.004026522394269705, + 0.002291602548211813, + 0.0045425728894770145, + 0.01026082132011652, + 0.009851702488958836, + 0.015392262488603592, + 0.01257766131311655, + 0.008299418725073338, + 0.015587486326694489, + 0.008638857863843441, + 0.018887540325522423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.05498216673731804 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787428192794323 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.009787428192794323, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787428192794323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10433028638362885, + 0.0619206577539444, + 0.09907516837120056, + 0.06939258426427841, + 0.2061808854341507, + 0.14198338985443115, + 0.08060990273952484, + 0.0701090395450592, + 0.06760159879922867, + 0.09879648685455322 + ], + "credit_map": [ + 0.014681809581816196, + 0.008405068889260292, + 0.02070719189941883, + 0.010162850841879845, + 0.007117650471627712, + 0.017181718721985817, + 0.006594594568014145, + 0.01947939768433571, + 0.008131220936775208, + 0.00487823411822319, + 0.014896228909492493, + 0.005889615043997765, + 0.0033724282402545214, + 0.009083986282348633, + 0.004274347797036171, + 0.011394595727324486, + 0.011403894983232021, + 0.0062543777748942375, + 0.019888775423169136, + 0.010347547940909863, + 0.006278233136981726, + 0.013676865957677364, + 0.008121668361127377, + 0.023103797808289528, + 0.007927574217319489, + 0.0035546845756471157, + 0.018307790160179138, + 0.006499597802758217, + 0.005724803544580936, + 0.010305632837116718, + 0.004680279176682234, + 0.012392223812639713, + 0.022523459047079086, + 0.01106528751552105, + 0.0676952451467514, + 0.016235362738370895, + 0.013357583433389664, + 0.033798083662986755, + 0.013333592563867569, + 0.028172269463539124, + 0.01624743826687336, + 0.008560039103031158, + 0.04415060952305794, + 0.015986906364560127, + 0.007173152640461922, + 0.023946672677993774, + 0.008928508497774601, + 0.016990067437291145, + 0.009698988869786263, + 0.006660198327153921, + 0.022574279457330704, + 0.00460307952016592, + 0.007755705155432224, + 0.012778307311236858, + 0.004986986052244902, + 0.0115523561835289, + 0.006719840224832296, + 0.011489401571452618, + 0.008229917846620083, + 0.014002999290823936, + 0.004004118498414755, + 0.009348055347800255, + 0.005990126635879278, + 0.010324579663574696, + 0.006935902405530214, + 0.0040056719444692135, + 0.01906590349972248, + 0.006319824140518904, + 0.005145382601767778, + 0.010197822004556656, + 0.005279876291751862, + 0.010651219636201859, + 0.0108869643881917, + 0.007388300262391567, + 0.02614637278020382, + 0.006183980032801628, + 0.009849759750068188, + 0.01691960170865059, + 0.007778552360832691, + 0.013642947189509869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.009787428192794323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015184074873104692 + ], + "opposite_logprob_tokens": [ + -6.501518249511719 + ], + "logprob_reward": -0.0015184074873104692, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015184074873104692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2395889014005661, + 0.18860478699207306, + 0.1024906113743782, + 0.0452938936650753, + 0.1357739418745041, + 0.04529127478599548, + 0.069341279566288, + 0.05982779338955879, + 0.07026908546686172, + 0.043518465012311935 + ], + "credit_map": [ + 0.02863352932035923, + 0.014834986999630928, + 0.07584898173809052, + 0.015873830765485764, + 0.017910901457071304, + 0.038308046758174896, + 0.014819419011473656, + 0.033359210938215256, + 0.022415610030293465, + 0.010175571776926517, + 0.0627957135438919, + 0.010302137583494186, + 0.013812904246151447, + 0.03205496817827225, + 0.010780654847621918, + 0.02626722864806652, + 0.011888114735484123, + 0.006145533639937639, + 0.018539823591709137, + 0.011789806187152863, + 0.00673715490847826, + 0.014868796803057194, + 0.007651601918041706, + 0.024869775399565697, + 0.005236166529357433, + 0.00274360878393054, + 0.00924849882721901, + 0.0052275098860263824, + 0.002512999577447772, + 0.007733616512268782, + 0.0034220563247799873, + 0.009169439785182476, + 0.014602327719330788, + 0.008307878859341145, + 0.04405028745532036, + 0.009450414218008518, + 0.00882825069129467, + 0.021775253117084503, + 0.009279999881982803, + 0.019479529932141304, + 0.004421519115567207, + 0.004472634289413691, + 0.006922504864633083, + 0.008644507266581059, + 0.002903703832998872, + 0.00652741314843297, + 0.004534736275672913, + 0.006864258553832769, + 0.007352189626544714, + 0.005177315790206194, + 0.021665658801794052, + 0.004139994736760855, + 0.005477128084748983, + 0.011256633326411247, + 0.004338730592280626, + 0.00993362907320261, + 0.00637709628790617, + 0.0035072765313088894, + 0.015470195561647415, + 0.005828364286571741, + 0.0036032653879374266, + 0.010093757882714272, + 0.0039049768820405006, + 0.011042863130569458, + 0.00710139237344265, + 0.007213796488940716, + 0.015870200470089912, + 0.008431842550635338, + 0.006394435651600361, + 0.00944715365767479, + 0.006122698541730642, + 0.009687570855021477, + 0.0043402258306741714, + 0.003536461153998971, + 0.0076795159839093685, + 0.004792340565472841, + 0.0032447248231619596, + 0.007247194182127714, + 0.0028857688885182142, + 0.009792235679924488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0015184074873104692 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tightly static, horizontal close-up focuses on the round headlight of a vintage car, captured in a stylized, high-contrast black and white image. Its raised glass lens is surrounded by a gleaming chrome bezel, which reflects sharp, dazzling highlights, creating a dramatic contrast with the deep, velvety shadows of the dark, polished metal surrounding it. The headlight, motionless and resolute, projects a powerful, focused beam forward, piercing the endless blackness of the frame. Within the cone of light, tiny dust particles suspended in the air dance slowly, lending the light itself a tangible quality. The scene is hushed and imbued with the suspense of film noir and a sense of eternity frozen in time, as if the car is quietly awaiting something in a forgotten corner.", + "seed": 12391, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_10/index_5869_seed_12391_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02820872897282243, + "vlm_logprob": -0.02820872897282243, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.02820872897282243, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04405871033668518 + ], + "opposite_logprob_tokens": [ + -3.2940587997436523 + ], + "logprob_reward": -0.04405871033668518, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04405871033668518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26125386357307434, + 0.32621127367019653, + 0.058649200946092606, + 0.025115961208939552, + 0.07194261997938156, + 0.056720852851867676, + 0.08890517801046371, + 0.036443471908569336, + 0.036622144281864166, + 0.038135506212711334 + ], + "credit_map": [ + 0.03443723917007446, + 0.023534176871180534, + 0.03573300689458847, + 0.027499645948410034, + 0.018682049587368965, + 0.038942016661167145, + 0.026113038882613182, + 0.05631266161799431, + 0.048118624836206436, + 0.04497726261615753, + 0.04031280428171158, + 0.03064538910984993, + 0.04315553605556488, + 0.04472185671329498, + 0.03459135815501213, + 0.039688415825366974, + 0.0068144123069942, + 0.005504608154296875, + 0.006952168885618448, + 0.004993995651602745, + 0.004948408808559179, + 0.009762397967278957, + 0.005422582384198904, + 0.014250621199607849, + 0.0035842577926814556, + 0.0025868394877761602, + 0.0037676459178328514, + 0.0029760277830064297, + 0.0018754908815026283, + 0.0036075713578611612, + 0.0029498052317649126, + 0.0037683204282075167, + 0.009506331756711006, + 0.007198082283139229, + 0.008106274530291557, + 0.007264369633048773, + 0.0053663793951272964, + 0.01170195359736681, + 0.007257899735122919, + 0.015541322529315948, + 0.008101994171738625, + 0.005233140662312508, + 0.0072261011227965355, + 0.004143540747463703, + 0.005581599660217762, + 0.00954062957316637, + 0.004568746313452721, + 0.012325097806751728, + 0.010774972848594189, + 0.01469285599887371, + 0.009310038760304451, + 0.00980689749121666, + 0.012103685177862644, + 0.013028928078711033, + 0.006724860053509474, + 0.012462933547794819, + 0.004792079329490662, + 0.00352816772647202, + 0.0035518789663910866, + 0.004011490847915411, + 0.0038114802446216345, + 0.005374511703848839, + 0.0040545170195400715, + 0.007319339085370302, + 0.005295202601701021, + 0.005688443314284086, + 0.00495034409686923, + 0.003779762424528599, + 0.0043231346644461155, + 0.005595441907644272, + 0.0031227953732013702, + 0.00386701594106853, + 0.004587188363075256, + 0.003618374466896057, + 0.004856917541474104, + 0.003074015025049448, + 0.003859415650367737, + 0.006324669346213341, + 0.003380746813490987, + 0.008434174582362175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.04405871033668518 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032171655911952257 + ], + "opposite_logprob_tokens": [ + -5.7532172203063965 + ], + "logprob_reward": -0.0032171655911952257, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032171655911952257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31003496050834656, + 0.20393726229667664, + 0.08059205114841461, + 0.07294373214244843, + 0.05347558483481407, + 0.123827263712883, + 0.038981806486845016, + 0.056354936212301254, + 0.026527296751737595, + 0.03332509472966194 + ], + "credit_map": [ + 0.04812103137373924, + 0.03104337677359581, + 0.05119955167174339, + 0.027816936373710632, + 0.025165032595396042, + 0.052431344985961914, + 0.0247491504997015, + 0.04950854554772377, + 0.0350959375500679, + 0.023668508976697922, + 0.028024401515722275, + 0.016313666477799416, + 0.020092012360692024, + 0.031151065602898598, + 0.019162695854902267, + 0.03042898140847683, + 0.0128868967294693, + 0.004503181669861078, + 0.01020151749253273, + 0.006598570849746466, + 0.005216573365032673, + 0.012927897274494171, + 0.008260033093392849, + 0.019997380673885345, + 0.008920367807149887, + 0.005496100056916475, + 0.007975026965141296, + 0.0051933773793280125, + 0.00588242756202817, + 0.012023566290736198, + 0.006167132407426834, + 0.021285736933350563, + 0.0070795463398098946, + 0.00567144388332963, + 0.007157656364142895, + 0.004919430706650019, + 0.006287224590778351, + 0.008501909673213959, + 0.0049875658005476, + 0.00887080654501915, + 0.01534105371683836, + 0.01178833469748497, + 0.015908949077129364, + 0.009003345854580402, + 0.012745758518576622, + 0.020800532773137093, + 0.008961061015725136, + 0.029278241097927094, + 0.00619432283565402, + 0.006137864198535681, + 0.0034104844089597464, + 0.005232078954577446, + 0.004942764528095722, + 0.006124955601990223, + 0.003241388127207756, + 0.003697950392961502, + 0.006946544628590345, + 0.007812552154064178, + 0.00765541335567832, + 0.007112691178917885, + 0.005310453474521637, + 0.008553551509976387, + 0.0047348253428936005, + 0.008228899911046028, + 0.0037143081426620483, + 0.0033947601914405823, + 0.003428491298109293, + 0.0026742129120975733, + 0.003084961324930191, + 0.0037930726539343596, + 0.0023931192699819803, + 0.004044372122734785, + 0.005055370274931192, + 0.003127776086330414, + 0.004573236685246229, + 0.004014573525637388, + 0.0026316458825021982, + 0.004984245169907808, + 0.003496677614748478, + 0.005441567394882441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0032171655911952257 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901675999164581 + ], + "opposite_logprob_tokens": [ + -2.57901668548584 + ], + "logprob_reward": -0.07901675999164581, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901675999164581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21764175593852997, + 0.30132657289505005, + 0.0753178820014, + 0.0653076246380806, + 0.06415066123008728, + 0.11946776509284973, + 0.09280522167682648, + 0.017990723252296448, + 0.026539960876107216, + 0.01945185475051403 + ], + "credit_map": [ + 0.037959542125463486, + 0.014193447306752205, + 0.03747544065117836, + 0.0146480156108737, + 0.012921929359436035, + 0.036251045763492584, + 0.017246011644601822, + 0.04694630205631256, + 0.04887339472770691, + 0.042899083346128464, + 0.03871891275048256, + 0.025837115943431854, + 0.037242788821458817, + 0.044216338545084, + 0.029622025787830353, + 0.0339169055223465, + 0.011252855882048607, + 0.006256473250687122, + 0.009671954438090324, + 0.005989560857415199, + 0.008114784955978394, + 0.011506523936986923, + 0.00721585052087903, + 0.015309879556298256, + 0.010197402909398079, + 0.0045584081672132015, + 0.00795983150601387, + 0.005674457643181086, + 0.004319863859564066, + 0.0107295922935009, + 0.005529907066375017, + 0.0163381639868021, + 0.007583220023661852, + 0.009262940846383572, + 0.007943798787891865, + 0.005012105219066143, + 0.010064597241580486, + 0.009350714273750782, + 0.006325668655335903, + 0.00860761571675539, + 0.016924211755394936, + 0.01139957644045353, + 0.013783089816570282, + 0.008373874239623547, + 0.012059992179274559, + 0.020012952387332916, + 0.008532555773854256, + 0.028381507843732834, + 0.01084419246762991, + 0.01616130955517292, + 0.010044976137578487, + 0.009838802739977837, + 0.012828846462070942, + 0.013997572474181652, + 0.006771378219127655, + 0.012318151071667671, + 0.002162726828828454, + 0.002282425295561552, + 0.0017942694248631597, + 0.0014808040577918291, + 0.0021732980385422707, + 0.0026102482806891203, + 0.0015255657490342855, + 0.003961385693401098, + 0.004664701409637928, + 0.0031886850483715534, + 0.0033072044607251883, + 0.0018448061309754848, + 0.0031301628332585096, + 0.004464137367904186, + 0.001977960579097271, + 0.00396230211481452, + 0.0024935437832027674, + 0.0015734747285023332, + 0.0029305615462362766, + 0.0016734084347262979, + 0.001867512590251863, + 0.0030872856732457876, + 0.0015222551301121712, + 0.0043038115836679935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.07901675999164581 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012548303231596947 + ], + "opposite_logprob_tokens": [ + -4.387548446655273 + ], + "logprob_reward": -0.012548303231596947, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012548303231596947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44183364510536194, + 0.08339737355709076, + 0.07654264569282532, + 0.046411339193582535, + 0.041535668075084686, + 0.029303668066859245, + 0.20943284034729004, + 0.016279738396406174, + 0.037469055503606796, + 0.01779407262802124 + ], + "credit_map": [ + 0.06581582129001617, + 0.055472973734140396, + 0.06426309794187546, + 0.03332152217626572, + 0.04915055260062218, + 0.06462297588586807, + 0.044028542935848236, + 0.06515813618898392, + 0.013833194971084595, + 0.006850023753941059, + 0.01236866693943739, + 0.010477758944034576, + 0.007533155381679535, + 0.01108221523463726, + 0.009663916192948818, + 0.011588442139327526, + 0.013935660012066364, + 0.0061852093786001205, + 0.00962425209581852, + 0.007174226921051741, + 0.008917592465877533, + 0.011415774933993816, + 0.007311543915420771, + 0.011978383176028728, + 0.006306165363639593, + 0.005252703558653593, + 0.0061090546660125256, + 0.0038000510539859533, + 0.005936009809374809, + 0.0071245161816477776, + 0.003910057246685028, + 0.007972780615091324, + 0.004948013927787542, + 0.006573931779712439, + 0.00413372041657567, + 0.003960684407502413, + 0.0064598750323057175, + 0.005780065897852182, + 0.003722052089869976, + 0.0059573291800916195, + 0.0037444636691361666, + 0.002144709462299943, + 0.0036378048826009035, + 0.002634329255670309, + 0.0020839099306613207, + 0.004605901427567005, + 0.0031168682035058737, + 0.007335680071264505, + 0.027149002999067307, + 0.03380317986011505, + 0.02076078951358795, + 0.02198718674480915, + 0.026588108390569687, + 0.033516403287649155, + 0.01705959066748619, + 0.028568582609295845, + 0.0024639342445880175, + 0.001941846334375441, + 0.001973327249288559, + 0.0019718443509191275, + 0.0018079213332384825, + 0.002311736810952425, + 0.0014602437149733305, + 0.002348886104300618, + 0.0055480822920799255, + 0.005634368862956762, + 0.004070806782692671, + 0.003924360498785973, + 0.004770764149725437, + 0.006000090390443802, + 0.0029230653308331966, + 0.004597518127411604, + 0.0027121691964566708, + 0.002502280520275235, + 0.001812893315218389, + 0.0013945194659754634, + 0.002145863138139248, + 0.0026066487189382315, + 0.0015012305229902267, + 0.0031184684485197067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.012548303231596947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the image presented in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round headlight with a raised glass lens surrounded by a chrome bezel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a beam of light projecting forward into darkness with visible dust particles dancing within it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera framing a tightly static horizontal close-up of the headlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there deep, velvety shadows contrasting with the gleaming chrome and sharp highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022027057129889727 + ], + "opposite_logprob_tokens": [ + -6.12720251083374 + ], + "logprob_reward": -0.0022027057129889727, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022027057129889727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28719890117645264, + 0.13733573257923126, + 0.10849875956773758, + 0.06841151416301727, + 0.060427356511354446, + 0.10595227777957916, + 0.08978869765996933, + 0.038771793246269226, + 0.05220494791865349, + 0.051409974694252014 + ], + "credit_map": [ + 0.04013757407665253, + 0.03497961908578873, + 0.03452086076140404, + 0.024383975192904472, + 0.041957538574934006, + 0.038177404552698135, + 0.025179848074913025, + 0.047862108796834946, + 0.01631433330476284, + 0.012894874438643456, + 0.015637699514627457, + 0.01524142175912857, + 0.012921045534312725, + 0.02097151055932045, + 0.013089682906866074, + 0.03026517666876316, + 0.012126740999519825, + 0.015163344331085682, + 0.011158164590597153, + 0.010227765887975693, + 0.01276285108178854, + 0.016005739569664, + 0.008600923232734203, + 0.02245323546230793, + 0.013379758223891258, + 0.0049911728128790855, + 0.008606219664216042, + 0.004677227698266506, + 0.005392054561525583, + 0.01068769209086895, + 0.006304813548922539, + 0.014372576959431171, + 0.008720808662474155, + 0.006655598059296608, + 0.0087598180398345, + 0.006012903060764074, + 0.006747041363269091, + 0.009428992867469788, + 0.005164770409464836, + 0.00893742498010397, + 0.0157600287348032, + 0.010524999350309372, + 0.01253688707947731, + 0.007510945200920105, + 0.011389394290745258, + 0.018010031431913376, + 0.007609346881508827, + 0.02261064387857914, + 0.012638217769563198, + 0.015104963444173336, + 0.008801013231277466, + 0.010132885538041592, + 0.012047250755131245, + 0.014252599328756332, + 0.006869891192764044, + 0.00994187593460083, + 0.005217363592237234, + 0.0065834783017635345, + 0.004565459676086903, + 0.005013121757656336, + 0.004742234945297241, + 0.005375550594180822, + 0.003140046028420329, + 0.0041345409117639065, + 0.007102551870048046, + 0.004949184600263834, + 0.0077486103400588036, + 0.0049148546531796455, + 0.004231900442391634, + 0.008609917014837265, + 0.004212058149278164, + 0.010435874573886395, + 0.008103389292955399, + 0.004226183518767357, + 0.00699650077149272, + 0.004809466190636158, + 0.004737999755889177, + 0.008375925943255424, + 0.004269804339855909, + 0.00989071186631918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0022027057129889727 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_11/index_tensor([9175, 9175])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_11/index_tensor([9175, 9175])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c9bb09d0b96b3d3422b64cb187e4da24c605e20d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_11/index_tensor([9175, 9175])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide-angle shot captures the scene, the camera slowly and steadily panning right, tracking the majestic movement of the ocean on a secluded beach beneath a moonless, starlit night sky. A powerful wave surges from the inky black sea, its translucent interior illuminated by a vibrant network of electric blue bioluminescence, transforming the invisible plankton into a pulsating web of light. As the wave reaches its crest, the light intensifies to a brilliant cyan before crashing violently against the shore, exploding into a spectacular display of luminous foam and sparkling droplets. The water then rapidly spreads over the dark, damp sand, which, like a black mirror, reflects an ethereal glow. As the tide recedes, an intricate, branching pattern of light is left on the beach, like a fading constellation. Individual points of light shimmer and pulsate before being withdrawn into darkness, creating an atmosphere of deep tranquility and magical wonder—a natural spectacle.", + "seed": 6664, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_11/index_9175_seed_6664_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05512777427211404, + "vlm_logprob": -0.05512777427211404, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.05512777427211404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2106677144765854 + ], + "opposite_logprob_tokens": [ + -1.7106677293777466 + ], + "logprob_reward": -0.2106677144765854, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2106677144765854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1451384276151657, + 0.12468690425157547, + 0.31199002265930176, + 0.13585591316223145, + 0.04043632373213768, + 0.03453582525253296, + 0.05166155844926834, + 0.05707879737019539, + 0.048796121031045914, + 0.04982009157538414 + ], + "credit_map": [ + 0.010161642916500568, + 0.013643664307892323, + 0.014790075831115246, + 0.025172512978315353, + 0.018098006024956703, + 0.02076725848019123, + 0.02088727056980133, + 0.021617980673909187, + 0.011175191961228848, + 0.018698550760746002, + 0.026065746322274208, + 0.01617370918393135, + 0.012614996172487736, + 0.011327858082950115, + 0.017707007005810738, + 0.010923843830823898, + 0.015846138820052147, + 0.01872941479086876, + 0.10088307410478592, + 0.03509949520230293, + 0.07994281500577927, + 0.017070554196834564, + 0.0263750609010458, + 0.018043451011180878, + 0.011365622282028198, + 0.016329703852534294, + 0.02963442914187908, + 0.02421431615948677, + 0.011514789424836636, + 0.00858204998075962, + 0.021766366437077522, + 0.012448620982468128, + 0.0017816326580941677, + 0.0019845357164740562, + 0.008881757967174053, + 0.01304194051772356, + 0.0022463169880211353, + 0.002435519127175212, + 0.003442242508754134, + 0.006622373592108488, + 0.0022282118443399668, + 0.0034461061004549265, + 0.005244753323495388, + 0.003194168210029602, + 0.004673386458307505, + 0.002751400228589773, + 0.006730148568749428, + 0.006267646327614784, + 0.0031940338667482138, + 0.004829988814890385, + 0.006178843788802624, + 0.005601990502327681, + 0.00689260708168149, + 0.004422240890562534, + 0.012614624574780464, + 0.007927222177386284, + 0.004556768573820591, + 0.010648710653185844, + 0.0035219143610447645, + 0.005858221556991339, + 0.00829380378127098, + 0.0027381093241274357, + 0.010935376398265362, + 0.010525891557335854, + 0.003402327187359333, + 0.004645348060876131, + 0.005932481959462166, + 0.008970359340310097, + 0.00593479722738266, + 0.0029003319796174765, + 0.009950329549610615, + 0.007060140371322632, + 0.003633442334830761, + 0.0058389524929225445, + 0.00614849803969264, + 0.007099163252860308, + 0.0061493609100580215, + 0.002521644812077284, + 0.00978497602045536, + 0.008644051849842072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.2106677144765854 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011096670292317867 + ], + "opposite_logprob_tokens": [ + -4.511096477508545 + ], + "logprob_reward": -0.011096670292317867, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011096670292317867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1152193546295166, + 0.07748739421367645, + 0.1436191201210022, + 0.1382572203874588, + 0.13598740100860596, + 0.04155787453055382, + 0.052340928465127945, + 0.08895694464445114, + 0.11021628975868225, + 0.09635753929615021 + ], + "credit_map": [ + 0.012485641986131668, + 0.013475300744175911, + 0.017200512811541557, + 0.015552486293017864, + 0.014039425179362297, + 0.01686096377670765, + 0.01353096216917038, + 0.01207406260073185, + 0.005613221321254969, + 0.009314862079918385, + 0.018994905054569244, + 0.009067359380424023, + 0.007220542524009943, + 0.0066717383451759815, + 0.013184230774641037, + 0.007420533802360296, + 0.008940481580793858, + 0.015086254104971886, + 0.02337932214140892, + 0.01377831120043993, + 0.03822268173098564, + 0.01348931435495615, + 0.01882677897810936, + 0.011895970441401005, + 0.01442309096455574, + 0.019769150763750076, + 0.013594325631856918, + 0.027898622676730156, + 0.020327113568782806, + 0.010204874910414219, + 0.01420495007187128, + 0.017835093662142754, + 0.009755208157002926, + 0.01528997253626585, + 0.017718259245157242, + 0.023104069754481316, + 0.023759033530950546, + 0.014774694107472897, + 0.01456963736563921, + 0.017016524448990822, + 0.003068072022870183, + 0.0045235734432935715, + 0.005416453815996647, + 0.004559497814625502, + 0.005469120107591152, + 0.00260773953050375, + 0.01023111306130886, + 0.0056823063641786575, + 0.004629878327250481, + 0.005803956184536219, + 0.00645967572927475, + 0.004788815043866634, + 0.007503564935177565, + 0.0038456148467957973, + 0.011371801607310772, + 0.00793762132525444, + 0.007082005497068167, + 0.0053156521171331406, + 0.006824646145105362, + 0.013228270225226879, + 0.010280181653797626, + 0.004291401710361242, + 0.02786771021783352, + 0.014067074283957481, + 0.009498187340795994, + 0.014692011289298534, + 0.011928184889256954, + 0.023321419954299927, + 0.010234677232801914, + 0.007298543583601713, + 0.01761765591800213, + 0.0156256090849638, + 0.007266949862241745, + 0.008649989031255245, + 0.009836267679929733, + 0.01358703151345253, + 0.010728554800152779, + 0.004401023965328932, + 0.027460850775241852, + 0.014426874928176403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.011096670292317867 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04303942248225212 + ], + "opposite_logprob_tokens": [ + -3.168039321899414 + ], + "logprob_reward": -0.04303942248225212, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04303942248225212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13425922393798828, + 0.06557652354240417, + 0.179691344499588, + 0.14510472118854523, + 0.1243719682097435, + 0.02920502983033657, + 0.18487834930419922, + 0.04999660328030586, + 0.02343316376209259, + 0.06348314136266708 + ], + "credit_map": [ + 0.007810422219336033, + 0.023568663746118546, + 0.01434708945453167, + 0.019824733957648277, + 0.020584959536790848, + 0.015418335795402527, + 0.015869613736867905, + 0.016835393384099007, + 0.006256815977394581, + 0.00958346575498581, + 0.009742839261889458, + 0.009758503176271915, + 0.008692082948982716, + 0.005120274145156145, + 0.009836554527282715, + 0.006585975177586079, + 0.011545306071639061, + 0.02035481110215187, + 0.02582029439508915, + 0.015929725021123886, + 0.05550335720181465, + 0.01562109962105751, + 0.02016698755323887, + 0.014749746769666672, + 0.013004397973418236, + 0.020213967189192772, + 0.027575254440307617, + 0.019034110009670258, + 0.027302898466587067, + 0.011100427247583866, + 0.012735381722450256, + 0.014138265512883663, + 0.00936802290380001, + 0.007467017974704504, + 0.026481332257390022, + 0.013632160611450672, + 0.014617986045777798, + 0.009660215117037296, + 0.011668439954519272, + 0.03147677704691887, + 0.002111756708472967, + 0.0028877216391265392, + 0.0032927682623267174, + 0.003906322177499533, + 0.004426989704370499, + 0.002432383829727769, + 0.0038103656843304634, + 0.006336720194667578, + 0.014738624915480614, + 0.020472979173064232, + 0.02693699114024639, + 0.015487979166209698, + 0.025416649878025055, + 0.01963079534471035, + 0.039591435343027115, + 0.022602874785661697, + 0.005772068165242672, + 0.004157820716500282, + 0.004176557529717684, + 0.0067793033085763454, + 0.0066466000862419605, + 0.0034697793889790773, + 0.011350474320352077, + 0.007643996272236109, + 0.0018727586138993502, + 0.0021885286550968885, + 0.002270815195515752, + 0.005091915372759104, + 0.002432550536468625, + 0.0013737600529566407, + 0.004417954478412867, + 0.003784879809245467, + 0.0038692990783602, + 0.005705441348254681, + 0.009417840279638767, + 0.008039511740207672, + 0.006717501673847437, + 0.003160227322950959, + 0.01688409224152565, + 0.009689223021268845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.04303942248225212 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201278228312731 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.002201278228312731, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201278228312731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08869094401597977, + 0.1585875302553177, + 0.14245423674583435, + 0.184878870844841, + 0.20694339275360107, + 0.04380102455615997, + 0.035539478063583374, + 0.05727274715900421, + 0.04348587244749069, + 0.03834591060876846 + ], + "credit_map": [ + 0.0054112765938043594, + 0.008681402541697025, + 0.009337163530290127, + 0.011859106831252575, + 0.013076748698949814, + 0.016606256365776062, + 0.011532116681337357, + 0.012186873704195023, + 0.01617072895169258, + 0.021723823621869087, + 0.03725038468837738, + 0.01841670833528042, + 0.0164948720484972, + 0.013464352115988731, + 0.021824218332767487, + 0.013242433778941631, + 0.008814282715320587, + 0.01313260942697525, + 0.033489856868982315, + 0.01022146362811327, + 0.045984234660863876, + 0.013535664416849613, + 0.011106147430837154, + 0.006169973406940699, + 0.014267010614275932, + 0.02681879699230194, + 0.02656369097530842, + 0.03281671553850174, + 0.03151679039001465, + 0.013062899000942707, + 0.020666055381298065, + 0.019166911020874977, + 0.008529935963451862, + 0.008498524315655231, + 0.05856577306985855, + 0.06661511957645416, + 0.008915510028600693, + 0.008361700922250748, + 0.019666723906993866, + 0.02779010683298111, + 0.0024641703348606825, + 0.003891333704814315, + 0.006417045835405588, + 0.004998020827770233, + 0.0036879561375826597, + 0.0029926823917776346, + 0.009791092947125435, + 0.009558723308146, + 0.002518296940252185, + 0.0031730218324810266, + 0.0033429646864533424, + 0.003644886426627636, + 0.005693853832781315, + 0.003251174930483103, + 0.008432776667177677, + 0.005482504144310951, + 0.006542532239109278, + 0.0051322090439498425, + 0.007767973933368921, + 0.007332128938287497, + 0.006328849587589502, + 0.003510802984237671, + 0.011278150603175163, + 0.009380101226270199, + 0.004334380850195885, + 0.005686972290277481, + 0.00714260246604681, + 0.006442379206418991, + 0.004356085322797298, + 0.0034902808256447315, + 0.005194072145968676, + 0.006839101202785969, + 0.002673129551112652, + 0.002723584184423089, + 0.004513108637183905, + 0.006758266594260931, + 0.003507559420540929, + 0.0017977702664211392, + 0.010450952686369419, + 0.005921538919210434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.002201278228312731 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633785881102085 + ], + "opposite_logprob_tokens": [ + -4.758633613586426 + ], + "logprob_reward": -0.008633785881102085, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633785881102085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058369286358356476, + 0.03825841844081879, + 0.11643300205469131, + 0.05006241053342819, + 0.12238398939371109, + 0.0398423857986927, + 0.04775504395365715, + 0.2599651515483856, + 0.04419423267245293, + 0.22273601591587067 + ], + "credit_map": [ + 0.005379124544560909, + 0.008332970552146435, + 0.006174004171043634, + 0.008841130882501602, + 0.006420372053980827, + 0.010767950676381588, + 0.006548708770424128, + 0.005905023775994778, + 0.0027917749248445034, + 0.003036402864381671, + 0.008395630866289139, + 0.004537353292107582, + 0.006325938273221254, + 0.0035091431345790625, + 0.005910784006118774, + 0.0037513889838010073, + 0.006148814223706722, + 0.008189951069653034, + 0.032632600516080856, + 0.01610833965241909, + 0.01628860831260681, + 0.008575801737606525, + 0.017078649252653122, + 0.01141023263335228, + 0.004517401102930307, + 0.004807469435036182, + 0.01337155606597662, + 0.008710619062185287, + 0.006604619789868593, + 0.0025055704172700644, + 0.004867518320679665, + 0.004677656572312117, + 0.008914312347769737, + 0.008071488700807095, + 0.03675919398665428, + 0.018072206526994705, + 0.011617736890912056, + 0.00901721604168415, + 0.011386558413505554, + 0.018545275554060936, + 0.002426148159429431, + 0.002555420156568289, + 0.005724882706999779, + 0.0039827353321015835, + 0.0043969289399683475, + 0.003192530246451497, + 0.008469917811453342, + 0.009093819186091423, + 0.002982105128467083, + 0.0046096788719296455, + 0.0056888326071202755, + 0.004400847014039755, + 0.004594616126269102, + 0.005242045968770981, + 0.012775183655321598, + 0.007461732253432274, + 0.01677163504064083, + 0.03456835076212883, + 0.01485060527920723, + 0.03365352377295494, + 0.03584861755371094, + 0.012355866841971874, + 0.07253383100032806, + 0.0393826998770237, + 0.003226613625884056, + 0.003021274460479617, + 0.006079742219299078, + 0.008130209520459175, + 0.004443351179361343, + 0.004553903825581074, + 0.007567300461232662, + 0.007171835284680128, + 0.013419839553534985, + 0.01968252845108509, + 0.027164442464709282, + 0.02981165051460266, + 0.02571261301636696, + 0.009170997887849808, + 0.06656579673290253, + 0.031208161264657974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.008633785881102085 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot captures the scene, the camera slowly and steadily panning right, tracking the majestic movement of the ocean on a secluded beach beneath a moonless, starlit night sky. A powerful wave surges from the inky black sea, its translucent interior illuminated by a vibrant network of electric blue bioluminescence, transforming the invisible plankton into a pulsating web of light. As the wave reaches its crest, the light intensifies to a brilliant cyan before crashing violently against the shore, exploding into a spectacular display of luminous foam and sparkling droplets. The water then rapidly spreads over the dark, damp sand, which, like a black mirror, reflects an ethereal glow. As the tide recedes, an intricate, branching pattern of light is left on the beach, like a fading constellation. Individual points of light shimmer and pulsate before being withdrawn into darkness, creating an atmosphere of deep tranquility and magical wonder—a natural spectacle.", + "seed": 6665, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_11/index_9175_seed_6665_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11197154596447945, + "vlm_logprob": -0.11197154596447945, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.11197154596447945, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.48124563694000244 + ], + "opposite_logprob_tokens": [ + -0.9812456369400024 + ], + "logprob_reward": -0.48124563694000244, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.48124563694000244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.768817126750946, + 0.03359599411487579, + 0.017837893217802048, + 0.04439907148480415, + 0.012880213558673859, + 0.05818256735801697, + 0.014696585014462471, + 0.030846906825900078, + 0.01332277711480856, + 0.00542090879753232 + ], + "credit_map": [ + 0.03856014832854271, + 0.0343068391084671, + 0.033594243228435516, + 0.0932418629527092, + 0.05596575513482094, + 0.24430912733078003, + 0.02463602088391781, + 0.244203120470047, + 0.0016364670591428876, + 0.0017215171828866005, + 0.0038351870607584715, + 0.00498950993642211, + 0.003099286463111639, + 0.005993328057229519, + 0.004038614220917225, + 0.008282084949314594, + 0.001882229233160615, + 0.0013343566097319126, + 0.0023375360760837793, + 0.002879546722397208, + 0.0011104578152298927, + 0.0011727245291694999, + 0.0024866957683116198, + 0.004634345881640911, + 0.0036959440913051367, + 0.006391461472958326, + 0.007755103521049023, + 0.008059436455368996, + 0.002926489105448127, + 0.004891023971140385, + 0.005607977509498596, + 0.005071635823696852, + 0.0011321707861497998, + 0.0012606838718056679, + 0.0016810055822134018, + 0.0015742386458441615, + 0.001327494508586824, + 0.0011076712980866432, + 0.002335743512958288, + 0.002461205469444394, + 0.0035226636100560427, + 0.0031870207749307156, + 0.007093566004186869, + 0.005243739113211632, + 0.009367715567350388, + 0.012531121261417866, + 0.012281176634132862, + 0.004955563694238663, + 0.0008158779819495976, + 0.0009207120747305453, + 0.0008547946927137673, + 0.001929434947669506, + 0.0020224323961883783, + 0.0012350616743788123, + 0.0019780315924435854, + 0.004940240178257227, + 0.003180251456797123, + 0.004097271244972944, + 0.0037814516108483076, + 0.002814220730215311, + 0.0066570402123034, + 0.004622667096555233, + 0.0032065031118690968, + 0.002487502060830593, + 0.0015422437572851777, + 0.002271522767841816, + 0.001293959910981357, + 0.001378827728331089, + 0.002241309965029359, + 0.0014057686785236, + 0.0018596503650769591, + 0.0013294944074004889, + 0.0005257283337414265, + 0.0008411986054852605, + 0.0004450783017091453, + 0.0004740946751553565, + 0.0011967172613367438, + 0.0008448516600765288, + 0.0004911614814773202, + 0.0006020785658620298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.48124563694000244 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608596369624138 + ], + "opposite_logprob_tokens": [ + -4.141086101531982 + ], + "logprob_reward": -0.01608596369624138, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608596369624138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8480948805809021, + 0.0152261508628726, + 0.03454376757144928, + 0.02357417158782482, + 0.01575317233800888, + 0.00824258103966713, + 0.015629854053258896, + 0.02771373838186264, + 0.00380908465012908, + 0.0074124946258962154 + ], + "credit_map": [ + 0.043031442910432816, + 0.03843146935105324, + 0.03635524958372116, + 0.10125887393951416, + 0.06196705624461174, + 0.27098944783210754, + 0.02720588445663452, + 0.26885557174682617, + 0.001021804055199027, + 0.0008201699238270521, + 0.0014513059286400676, + 0.003055470297113061, + 0.00100242393091321, + 0.0024674597661942244, + 0.0013741344446316361, + 0.004033383913338184, + 0.003748363582417369, + 0.0018302942626178265, + 0.005493085365742445, + 0.002894129604101181, + 0.0017409896245226264, + 0.0015094912378117442, + 0.005053134169429541, + 0.012274281121790409, + 0.0018978718435391784, + 0.003016043920069933, + 0.004025171976536512, + 0.0049850656650960445, + 0.0016819603042677045, + 0.002319233724847436, + 0.002867653500288725, + 0.002781173912808299, + 0.0007282804581336677, + 0.0007263391162268817, + 0.0038650943897664547, + 0.0013963564997538924, + 0.0008000502712093294, + 0.0008033909252844751, + 0.005442696623504162, + 0.001990966498851776, + 0.0005418161163106561, + 0.0008195686968974769, + 0.0011978124966844916, + 0.00158979173284024, + 0.0006252112216316164, + 0.0011841119267046452, + 0.0011784459929913282, + 0.001105824252590537, + 0.0013756125699728727, + 0.0016181241953745484, + 0.0014842650853097439, + 0.0028012823313474655, + 0.0015211912104859948, + 0.0014534343499690294, + 0.0024165704380720854, + 0.0029593748040497303, + 0.0027026322204619646, + 0.003531350754201412, + 0.0035231385845690966, + 0.0026385709643363953, + 0.006129701621830463, + 0.004113204311579466, + 0.0027788288425654173, + 0.002296313876286149, + 0.00038967179716564715, + 0.000629316782578826, + 0.000372362817870453, + 0.0003796853416133672, + 0.0005988638149574399, + 0.00045150500955060124, + 0.00048765793326310813, + 0.0005000217352062464, + 0.0007810613024048507, + 0.001084767747670412, + 0.0008158067939803004, + 0.0004517080669756979, + 0.001416817307472229, + 0.0010095305042341352, + 0.0008253851556219161, + 0.0010274184169247746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.01608596369624138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486338846385479 + ], + "opposite_logprob_tokens": [ + -3.0486338138580322 + ], + "logprob_reward": -0.0486338846385479, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486338846385479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40349629521369934, + 0.06162697821855545, + 0.05647711083292961, + 0.07199398428201675, + 0.028713492676615715, + 0.03714146837592125, + 0.01915586367249489, + 0.16943547129631042, + 0.037867188453674316, + 0.11409210413694382 + ], + "credit_map": [ + 0.020587196573615074, + 0.018973197788000107, + 0.017298834398388863, + 0.04853574186563492, + 0.028444744646549225, + 0.12154962867498398, + 0.020961515605449677, + 0.12714549899101257, + 0.0028053377754986286, + 0.002554218750447035, + 0.005334732122719288, + 0.008456479758024216, + 0.0056411814875900745, + 0.013603280298411846, + 0.005632758606225252, + 0.017598988488316536, + 0.006465869955718517, + 0.0032317449804395437, + 0.00692357774823904, + 0.005564176477491856, + 0.0038096969947218895, + 0.002876929473131895, + 0.008955867029726505, + 0.018649248406291008, + 0.005409945733845234, + 0.00918793398886919, + 0.012612123973667622, + 0.01531908754259348, + 0.0049870191141963005, + 0.006958379875868559, + 0.00885755941271782, + 0.008661936968564987, + 0.00222020223736763, + 0.0032347908709198236, + 0.003786579705774784, + 0.0037295189686119556, + 0.0023325292859226465, + 0.002874142723158002, + 0.0039695859886705875, + 0.006566143129020929, + 0.003322176868095994, + 0.003895553993061185, + 0.004596356302499771, + 0.004264956805855036, + 0.0029899945948272943, + 0.007951747626066208, + 0.00471739424392581, + 0.0054032872430980206, + 0.0014070161851122975, + 0.0016348495846614242, + 0.0023888854775577784, + 0.0028432910330593586, + 0.0020972387865185738, + 0.0021054805256426334, + 0.0032926842104643583, + 0.003386419266462326, + 0.015668081119656563, + 0.023608878254890442, + 0.019967813044786453, + 0.015447390265762806, + 0.0396001935005188, + 0.025019347667694092, + 0.016642626374959946, + 0.013481135480105877, + 0.004322718363255262, + 0.007313109003007412, + 0.0037974787410348654, + 0.004541626200079918, + 0.0057905386202037334, + 0.004635641351342201, + 0.003600991563871503, + 0.0038650829810649157, + 0.008542332798242569, + 0.020051486790180206, + 0.010482918471097946, + 0.007052678149193525, + 0.026841837912797928, + 0.017608478665351868, + 0.010969477705657482, + 0.01254289597272873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.0486338846385479 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028211576864123344 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028211576864123344, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028211576864123344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16725707054138184, + 0.06064026802778244, + 0.08209090679883957, + 0.035363778471946716, + 0.19728964567184448, + 0.15872076153755188, + 0.04514702409505844, + 0.1557689905166626, + 0.06703391671180725, + 0.030687624588608742 + ], + "credit_map": [ + 0.013863236643373966, + 0.012907426804304123, + 0.01213437132537365, + 0.012344028800725937, + 0.015957221388816833, + 0.05058083310723305, + 0.016150806099176407, + 0.03331916406750679, + 0.0036451101768761873, + 0.003471911419183016, + 0.00570815522223711, + 0.00790873821824789, + 0.005240324884653091, + 0.012846558354794979, + 0.00740865059196949, + 0.014410826377570629, + 0.008078630082309246, + 0.005234123673290014, + 0.013750690035521984, + 0.0086008096113801, + 0.004784639924764633, + 0.004697782453149557, + 0.010937376879155636, + 0.026006856933236122, + 0.002750169252976775, + 0.004472001921385527, + 0.006347186863422394, + 0.00622586440294981, + 0.0019955988973379135, + 0.003651136066764593, + 0.0049818214029073715, + 0.004939999897032976, + 0.010295259766280651, + 0.010925662703812122, + 0.046090058982372284, + 0.019932754337787628, + 0.010444088838994503, + 0.012452371418476105, + 0.06475339084863663, + 0.02239607647061348, + 0.010353343561291695, + 0.007874096743762493, + 0.034069400280714035, + 0.0120726078748703, + 0.011258939281105995, + 0.019219718873500824, + 0.050101496279239655, + 0.013771168887615204, + 0.002279441338032484, + 0.0030291308648884296, + 0.0021647540852427483, + 0.005344192963093519, + 0.005299811717122793, + 0.0032407892867922783, + 0.008374324068427086, + 0.01541458535939455, + 0.012688616290688515, + 0.0212992113083601, + 0.01755637116730213, + 0.013848548755049706, + 0.03367644175887108, + 0.021298931911587715, + 0.01837955228984356, + 0.017021318897604942, + 0.006104520056396723, + 0.012930027209222317, + 0.005859431345015764, + 0.005708160810172558, + 0.011531681753695011, + 0.0072934296913445, + 0.009191193617880344, + 0.00841547828167677, + 0.0029263535980135202, + 0.005375183653086424, + 0.003269990673288703, + 0.0027331861201673746, + 0.005196476820856333, + 0.003933620639145374, + 0.0032805376686155796, + 0.003972275648266077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.0028211576864123344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107108686119318 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.01107108686119318, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107108686119318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5950105786323547, + 0.0776624083518982, + 0.03987397998571396, + 0.05418726056814194, + 0.03578289970755577, + 0.11061380058526993, + 0.011361599899828434, + 0.039789024740457535, + 0.017217999324202538, + 0.018500415608286858 + ], + "credit_map": [ + 0.03052111156284809, + 0.028210509568452835, + 0.025979559868574142, + 0.07380029559135437, + 0.04208888113498688, + 0.18014124035835266, + 0.02527175098657608, + 0.18899725377559662, + 0.006508590653538704, + 0.00643389904871583, + 0.010452591814100742, + 0.010395778343081474, + 0.013569921255111694, + 0.007633340544998646, + 0.010594946332275867, + 0.012073342688381672, + 0.0033413823693990707, + 0.0019801175221800804, + 0.009143935516476631, + 0.0030835794750601053, + 0.0019018237944692373, + 0.0017237754072993994, + 0.005079117603600025, + 0.01362025085836649, + 0.00438926974311471, + 0.005951350089162588, + 0.009544950909912586, + 0.011514008976519108, + 0.003511398797854781, + 0.004385335370898247, + 0.008034847676753998, + 0.006856101099401712, + 0.0016999266808852553, + 0.001657462096773088, + 0.007802385371178389, + 0.003939023707062006, + 0.0018104914342984557, + 0.0024328562431037426, + 0.010807552374899387, + 0.005633203312754631, + 0.007998446002602577, + 0.007001468446105719, + 0.013182399794459343, + 0.014032411389052868, + 0.009064524434506893, + 0.021254468709230423, + 0.027582529932260513, + 0.010497547686100006, + 0.0007984747062437236, + 0.0011621542507782578, + 0.0009516136487945914, + 0.0018144914647564292, + 0.0009395918459631503, + 0.001268792781047523, + 0.0021897396072745323, + 0.0022367420606315136, + 0.004178246483206749, + 0.0045384857803583145, + 0.005823611281812191, + 0.004287732299417257, + 0.007365788333117962, + 0.00587142538279295, + 0.004055560100823641, + 0.0036681729834526777, + 0.0016110797878354788, + 0.002450033091008663, + 0.0014335676096379757, + 0.001583003206178546, + 0.0026893075555562973, + 0.0021715317852795124, + 0.0031319239642471075, + 0.0021475525572896004, + 0.0017152460059151053, + 0.002977243158966303, + 0.0011057547526434064, + 0.0011551121715456247, + 0.00301106134429574, + 0.00420433608815074, + 0.00173595838714391, + 0.0025957049801945686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01107108686119318 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot captures the scene, the camera slowly and steadily panning right, tracking the majestic movement of the ocean on a secluded beach beneath a moonless, starlit night sky. A powerful wave surges from the inky black sea, its translucent interior illuminated by a vibrant network of electric blue bioluminescence, transforming the invisible plankton into a pulsating web of light. As the wave reaches its crest, the light intensifies to a brilliant cyan before crashing violently against the shore, exploding into a spectacular display of luminous foam and sparkling droplets. The water then rapidly spreads over the dark, damp sand, which, like a black mirror, reflects an ethereal glow. As the tide recedes, an intricate, branching pattern of light is left on the beach, like a fading constellation. Individual points of light shimmer and pulsate before being withdrawn into darkness, creating an atmosphere of deep tranquility and magical wonder—a natural spectacle.", + "seed": 6666, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_11/index_9175_seed_6666_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08615634804591536, + "vlm_logprob": -0.08615634804591536, + "vlm_yes_no_margin": 3.8999999880790712, + "vlm_token_credit": -0.08615634804591536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.36100488901138306 + ], + "opposite_logprob_tokens": [ + -1.2360048294067383 + ], + "logprob_reward": -0.36100488901138306, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.36100488901138306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08095809817314148, + 0.020208211615681648, + 0.07966761291027069, + 0.08384009450674057, + 0.02185993827879429, + 0.02262129634618759, + 0.0619669109582901, + 0.02560899406671524, + 0.029012931510806084, + 0.5742558240890503 + ], + "credit_map": [ + 0.005977683700621128, + 0.007142098154872656, + 0.008699093945324421, + 0.010093547403812408, + 0.018514176830649376, + 0.009730163030326366, + 0.0080324811860919, + 0.012768859975039959, + 0.0017579452833160758, + 0.0020303146447986364, + 0.003142254427075386, + 0.003513769945129752, + 0.0027171073015779257, + 0.001876454334706068, + 0.0020379398483783007, + 0.003132427576929331, + 0.009074561297893524, + 0.011269145645201206, + 0.013667995110154152, + 0.011146479286253452, + 0.010290049016475677, + 0.008845489472150803, + 0.006644891109317541, + 0.008729007095098495, + 0.0071270652115345, + 0.007095042150467634, + 0.008500385098159313, + 0.013481910340487957, + 0.020885135978460312, + 0.007965147495269775, + 0.008120502345263958, + 0.010664907284080982, + 0.00230972352437675, + 0.0012711723102256656, + 0.0029361455235630274, + 0.004503161646425724, + 0.0017082261620089412, + 0.001850106636993587, + 0.0032692262902855873, + 0.004012176301330328, + 0.0018356358632445335, + 0.001878573908470571, + 0.004020883701741695, + 0.0037595259491354227, + 0.0025380938313901424, + 0.002432696521282196, + 0.0026086587458848953, + 0.003547228407114744, + 0.005888187792152166, + 0.005570453125983477, + 0.005049168597906828, + 0.008289806544780731, + 0.005002710502594709, + 0.007754687685519457, + 0.009186255745589733, + 0.015225645154714584, + 0.0019101033685728908, + 0.002286065835505724, + 0.002217237139120698, + 0.0054553113877773285, + 0.002052607014775276, + 0.0018684598617255688, + 0.005490435753017664, + 0.004328772891312838, + 0.002190405735746026, + 0.0030335874762386084, + 0.003498351201415062, + 0.00592842185869813, + 0.002530975267291069, + 0.0016083733644336462, + 0.005807352252304554, + 0.004415465984493494, + 0.03751487657427788, + 0.16130435466766357, + 0.02801715023815632, + 0.0507773794233799, + 0.11380582302808762, + 0.03921525925397873, + 0.05311459302902222, + 0.09050645679235458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.36100488901138306 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014223506674170494 + ], + "opposite_logprob_tokens": [ + -4.264223575592041 + ], + "logprob_reward": -0.014223506674170494, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014223506674170494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20432043075561523, + 0.06762735545635223, + 0.23378494381904602, + 0.09230653941631317, + 0.047717414796352386, + 0.060481470078229904, + 0.028821764513850212, + 0.01544476393610239, + 0.04512076452374458, + 0.20437456667423248 + ], + "credit_map": [ + 0.01731032319366932, + 0.04064098000526428, + 0.02562662586569786, + 0.028597718104720116, + 0.0319143645465374, + 0.0279551949352026, + 0.01637990027666092, + 0.01589532382786274, + 0.004713969770818949, + 0.007866967469453812, + 0.008143133483827114, + 0.009731938131153584, + 0.010698473080992699, + 0.006944207940250635, + 0.009776520542800426, + 0.009752144105732441, + 0.022871902212500572, + 0.02547590062022209, + 0.04326215013861656, + 0.022650932893157005, + 0.025893088430166245, + 0.025806071236729622, + 0.03843391314148903, + 0.02939099445939064, + 0.007012827787548304, + 0.007628647610545158, + 0.01087452843785286, + 0.018193509429693222, + 0.016136061400175095, + 0.008639656938612461, + 0.009551769122481346, + 0.014269531704485416, + 0.005608825478702784, + 0.003675404703244567, + 0.007946047000586987, + 0.005976642947643995, + 0.005174667574465275, + 0.0063116345554590225, + 0.0055898139253258705, + 0.007434379309415817, + 0.005242567975074053, + 0.006277692969888449, + 0.008873620070517063, + 0.010483861900866032, + 0.0051253377459943295, + 0.005681665148586035, + 0.008107451722025871, + 0.010689269751310349, + 0.002423524623736739, + 0.002986285835504532, + 0.003729418385773897, + 0.003231710521504283, + 0.00405013095587492, + 0.002259508240967989, + 0.003881029551848769, + 0.006260154768824577, + 0.0013568048598244786, + 0.0016348863719031215, + 0.002340917242690921, + 0.0026681330054998398, + 0.0013527467381209135, + 0.0012660699430853128, + 0.0022716596722602844, + 0.002553545869886875, + 0.0030351029708981514, + 0.004361521918326616, + 0.004895784426480532, + 0.008971013128757477, + 0.0036284830421209335, + 0.003254367969930172, + 0.009811066091060638, + 0.00716342544183135, + 0.013192969374358654, + 0.05511860176920891, + 0.011147653684020042, + 0.019739016890525818, + 0.038296476006507874, + 0.013691853731870651, + 0.02167646400630474, + 0.03151153028011322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.014223506674170494 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863763228058815 + ], + "opposite_logprob_tokens": [ + -3.048637628555298 + ], + "logprob_reward": -0.04863763228058815, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863763228058815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031118012964725494, + 0.02438645251095295, + 0.05527719855308533, + 0.10519044101238251, + 0.030403127893805504, + 0.02817453071475029, + 0.03285311907529831, + 0.0531882680952549, + 0.03283878043293953, + 0.6065701246261597 + ], + "credit_map": [ + 0.002099929843097925, + 0.004284089896827936, + 0.002884757472202182, + 0.004002521280199289, + 0.0035271155647933483, + 0.0022204818669706583, + 0.005804905202239752, + 0.0062942118383944035, + 0.002983604557812214, + 0.0041976249776780605, + 0.0032201630529016256, + 0.0032663969323039055, + 0.0034082664642482996, + 0.002007079077884555, + 0.002981052966788411, + 0.002322264015674591, + 0.005282215774059296, + 0.006544889882206917, + 0.008035054430365562, + 0.007085425313562155, + 0.004935465753078461, + 0.0053618233650922775, + 0.007790544535964727, + 0.010241778567433357, + 0.007580206263810396, + 0.0075754765421152115, + 0.009665161371231079, + 0.01630731113255024, + 0.026391709223389626, + 0.009456782601773739, + 0.008320285007357597, + 0.019893508404493332, + 0.0028433827683329582, + 0.0031816300470381975, + 0.0036139688454568386, + 0.004790648352354765, + 0.0031791680958122015, + 0.004942302592098713, + 0.00378995668143034, + 0.004062071908265352, + 0.0021612076088786125, + 0.003053105901926756, + 0.004568712320178747, + 0.005682166200131178, + 0.0026389018166810274, + 0.002190967323258519, + 0.0033885911107063293, + 0.0044908770360052586, + 0.0031934839207679033, + 0.0032725997734814882, + 0.004462213255465031, + 0.0047574350610375404, + 0.004118775017559528, + 0.0031876612920314074, + 0.003889995627105236, + 0.005970954895019531, + 0.004099342506378889, + 0.004864726215600967, + 0.0042993840761482716, + 0.010800089687108994, + 0.003936456516385078, + 0.004845231305807829, + 0.01118221040815115, + 0.009160825982689857, + 0.002650390611961484, + 0.00600772351026535, + 0.003523260587826371, + 0.007178314495831728, + 0.0033893880899995565, + 0.0017125258455052972, + 0.003846465377137065, + 0.004530710633844137, + 0.03891364112496376, + 0.1690225452184677, + 0.03189071640372276, + 0.0537814162671566, + 0.11898573487997055, + 0.04021046683192253, + 0.05824500322341919, + 0.09552060812711716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.04863763228058815 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028236540965735912 + ], + "opposite_logprob_tokens": [ + -5.877823829650879 + ], + "logprob_reward": -0.0028236540965735912, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028236540965735912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10087200999259949, + 0.05648798495531082, + 0.3484964072704315, + 0.09540103375911713, + 0.06461799889802933, + 0.057263005524873734, + 0.03849399834871292, + 0.04377536475658417, + 0.052949562668800354, + 0.14164265990257263 + ], + "credit_map": [ + 0.009590639732778072, + 0.019813578575849533, + 0.014440414495766163, + 0.01229817420244217, + 0.010623783804476261, + 0.013357911258935928, + 0.008518065325915813, + 0.01222944725304842, + 0.004968963097780943, + 0.005146848503500223, + 0.008291056379675865, + 0.007585327606648207, + 0.010122217237949371, + 0.006925771478563547, + 0.007114833686500788, + 0.006332967896014452, + 0.03557220473885536, + 0.05389958247542381, + 0.06257724016904831, + 0.03929077461361885, + 0.03878103941679001, + 0.03889147564768791, + 0.04393598437309265, + 0.035548098385334015, + 0.0057722716592252254, + 0.006155198905616999, + 0.004661205690354109, + 0.004301976412534714, + 0.03785065561532974, + 0.009828747250139713, + 0.005123159848153591, + 0.021707819774746895, + 0.006151651497930288, + 0.006089504808187485, + 0.012155131436884403, + 0.008916826918721199, + 0.006058506667613983, + 0.008595503866672516, + 0.007983158342540264, + 0.008667712099850178, + 0.00441307108849287, + 0.004865352995693684, + 0.009831414557993412, + 0.01196279563009739, + 0.004917355254292488, + 0.0042092204093933105, + 0.007154491730034351, + 0.009909301996231079, + 0.00429140031337738, + 0.0030540525913238525, + 0.003744665300473571, + 0.003609594888985157, + 0.003588930005207658, + 0.002619443228468299, + 0.006286298856139183, + 0.011299611069262028, + 0.0040882951579988, + 0.004518652334809303, + 0.005427652969956398, + 0.0055166129022836685, + 0.005091526545584202, + 0.004840622656047344, + 0.0073548052459955215, + 0.006937197409570217, + 0.004000757355242968, + 0.010668504051864147, + 0.005936475470662117, + 0.009954527951776981, + 0.006176295690238476, + 0.0030599709134548903, + 0.006470391992479563, + 0.006682639010250568, + 0.010537543334066868, + 0.04019474983215332, + 0.00724426144734025, + 0.014333739876747131, + 0.0265224389731884, + 0.01064904686063528, + 0.010695108212530613, + 0.021465763449668884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.0028236540965735912 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092058166861534 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.004092058166861534, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092058166861534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07864876836538315, + 0.03621332719922066, + 0.04822605103254318, + 0.029129400849342346, + 0.15372298657894135, + 0.14710566401481628, + 0.023542184382677078, + 0.027836782857775688, + 0.061113301664590836, + 0.3944615423679352 + ], + "credit_map": [ + 0.00818604975938797, + 0.015634216368198395, + 0.010084199719130993, + 0.009919364005327225, + 0.008150298148393631, + 0.010608004406094551, + 0.006186340004205704, + 0.00988029595464468, + 0.003939911723136902, + 0.0032713154796510935, + 0.005815592128783464, + 0.0039190794341266155, + 0.005273862276226282, + 0.003783673280850053, + 0.0038691451773047447, + 0.006340747233480215, + 0.005189511459320784, + 0.004365865141153336, + 0.00846956018358469, + 0.006847011856734753, + 0.006930373143404722, + 0.004613589029759169, + 0.006382224150002003, + 0.005427921190857887, + 0.002270330209285021, + 0.0024105492047965527, + 0.0029750110115855932, + 0.0032022367231547832, + 0.006789201870560646, + 0.0020243139006197453, + 0.003666691714897752, + 0.005791065748780966, + 0.01691264845430851, + 0.015218597836792469, + 0.018397977575659752, + 0.020437609404325485, + 0.0166359581053257, + 0.02108980529010296, + 0.025362320244312286, + 0.019668055698275566, + 0.011271354742348194, + 0.014662284404039383, + 0.0239290539175272, + 0.0306303221732378, + 0.013986104167997837, + 0.01125424262136221, + 0.017079073935747147, + 0.024293236434459686, + 0.001816156436689198, + 0.002676627831533551, + 0.0026693851687014103, + 0.0044007617980241776, + 0.0020442120730876923, + 0.0020555774681270123, + 0.0035133566707372665, + 0.004366107285022736, + 0.0019494668813422322, + 0.003089592093601823, + 0.0036141215823590755, + 0.004549086559563875, + 0.0029225540347397327, + 0.0021984868217259645, + 0.004379857797175646, + 0.005133618135005236, + 0.004257774446159601, + 0.005503218155354261, + 0.006638878025114536, + 0.012017994187772274, + 0.005590064451098442, + 0.004797343164682388, + 0.013606423512101173, + 0.008701604790985584, + 0.02564381994307041, + 0.11054861545562744, + 0.020423268899321556, + 0.034640904515981674, + 0.0776478573679924, + 0.027088697999715805, + 0.03616594523191452, + 0.06230244040489197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.004092058166861534 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot captures the scene, the camera slowly and steadily panning right, tracking the majestic movement of the ocean on a secluded beach beneath a moonless, starlit night sky. A powerful wave surges from the inky black sea, its translucent interior illuminated by a vibrant network of electric blue bioluminescence, transforming the invisible plankton into a pulsating web of light. As the wave reaches its crest, the light intensifies to a brilliant cyan before crashing violently against the shore, exploding into a spectacular display of luminous foam and sparkling droplets. The water then rapidly spreads over the dark, damp sand, which, like a black mirror, reflects an ethereal glow. As the tide recedes, an intricate, branching pattern of light is left on the beach, like a fading constellation. Individual points of light shimmer and pulsate before being withdrawn into darkness, creating an atmosphere of deep tranquility and magical wonder—a natural spectacle.", + "seed": 6667, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_11/index_9175_seed_6667_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08196565983816981, + "vlm_logprob": -0.08196565983816981, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.08196565983816981, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35823535919189453 + ], + "opposite_logprob_tokens": [ + -1.2332353591918945 + ], + "logprob_reward": -0.35823535919189453, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35823535919189453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20138335227966309, + 0.0909697487950325, + 0.10125569999217987, + 0.10836298018693924, + 0.2464030534029007, + 0.07668165117502213, + 0.04804353415966034, + 0.06271026283502579, + 0.0493289977312088, + 0.014860644936561584 + ], + "credit_map": [ + 0.01818528212606907, + 0.011441670358181, + 0.012875281274318695, + 0.013051843270659447, + 0.012759091332554817, + 0.04050804674625397, + 0.025405289605259895, + 0.06715685874223709, + 0.005082146264612675, + 0.01256382092833519, + 0.01421397365629673, + 0.01234996784478426, + 0.012366480194032192, + 0.014201466925442219, + 0.007737643551081419, + 0.012454252690076828, + 0.01765267178416252, + 0.005688515957444906, + 0.006701127626001835, + 0.005331275053322315, + 0.0072714886628091335, + 0.005868609994649887, + 0.017434606328606606, + 0.03530741110444069, + 0.013602262362837791, + 0.004893566016107798, + 0.02323906123638153, + 0.0054811532609164715, + 0.0055808681063354015, + 0.006028675474226475, + 0.011642981320619583, + 0.03789442405104637, + 0.018283212557435036, + 0.02728518657386303, + 0.027454374358057976, + 0.03659237548708916, + 0.024175245314836502, + 0.04378122091293335, + 0.03418950363993645, + 0.03464195877313614, + 0.008404769003391266, + 0.006155535578727722, + 0.010438844561576843, + 0.007806429639458656, + 0.006826986093074083, + 0.003982933238148689, + 0.024286944419145584, + 0.00877921748906374, + 0.007696761749684811, + 0.005263344384729862, + 0.006955000571906567, + 0.006503764074295759, + 0.005075553897768259, + 0.003805172862485051, + 0.00627474719658494, + 0.006469196639955044, + 0.0050977314822375774, + 0.018999038264155388, + 0.005092384293675423, + 0.01851744018495083, + 0.003928437829017639, + 0.003053896827623248, + 0.0038600966800004244, + 0.004161235876381397, + 0.012011593207716942, + 0.0030791747849434614, + 0.005668066907674074, + 0.002654177835211158, + 0.0016581850359216332, + 0.0038363321218639612, + 0.0028880273457616568, + 0.017533447593450546, + 0.001394746475853026, + 0.0018846836173906922, + 0.00173049361910671, + 0.0027430537156760693, + 0.0010913527803495526, + 0.001230123220011592, + 0.0021634979639202356, + 0.0026226956397295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.35823535919189453 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012579145841300488 + ], + "opposite_logprob_tokens": [ + -4.387578964233398 + ], + "logprob_reward": -0.012579145841300488, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012579145841300488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07091803848743439, + 0.3602200448513031, + 0.059925466775894165, + 0.05237355828285217, + 0.1442679911851883, + 0.029827997088432312, + 0.02789914980530739, + 0.043714314699172974, + 0.09378203749656677, + 0.11707133799791336 + ], + "credit_map": [ + 0.0049209012649953365, + 0.0037095854058861732, + 0.004833052400499582, + 0.003974782768636942, + 0.00582848209887743, + 0.02245902270078659, + 0.0072102295234799385, + 0.01798197813332081, + 0.014286170713603497, + 0.030444025993347168, + 0.016967907547950745, + 0.1059582307934761, + 0.0357428677380085, + 0.043307382613420486, + 0.023815276101231575, + 0.0896981880068779, + 0.007586423307657242, + 0.0036586641799658537, + 0.008117280900478363, + 0.006039785221219063, + 0.006021611392498016, + 0.006039002910256386, + 0.008671117946505547, + 0.013791583478450775, + 0.008996382355690002, + 0.003116469131782651, + 0.0178834218531847, + 0.006759266834706068, + 0.0030503221787512302, + 0.00267074815928936, + 0.0045441025868058205, + 0.005352849140763283, + 0.010332703590393066, + 0.015795864164829254, + 0.016071166843175888, + 0.02274392731487751, + 0.013520077802240849, + 0.026647333055734634, + 0.019778450950980186, + 0.01937847025692463, + 0.0032842839136719704, + 0.0034479997120797634, + 0.003992795012891293, + 0.0034027721267193556, + 0.00332993664778769, + 0.002183204283937812, + 0.004943022038787603, + 0.005243981257081032, + 0.004958503879606724, + 0.003418854670599103, + 0.0033030721824616194, + 0.0038878985214978456, + 0.002461968222633004, + 0.002770615741610527, + 0.003594722133129835, + 0.0035035153850913048, + 0.0063918232917785645, + 0.015030939131975174, + 0.0029090195894241333, + 0.009663436561822891, + 0.002208366058766842, + 0.002090600086376071, + 0.002382158301770687, + 0.00303797353990376, + 0.0274514053016901, + 0.0059067727997899055, + 0.006610286887735128, + 0.00546991778537631, + 0.002728902269154787, + 0.006538573652505875, + 0.006199121009558439, + 0.032877061516046524, + 0.011897549033164978, + 0.011563845910131931, + 0.01655689999461174, + 0.017678461968898773, + 0.011410839855670929, + 0.014708255417644978, + 0.014331966638565063, + 0.018923519179224968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.012579145841300488 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029787791892886162 + ], + "opposite_logprob_tokens": [ + -3.52978777885437 + ], + "logprob_reward": -0.029787791892886162, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029787791892886162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20229017734527588, + 0.10455042868852615, + 0.06060509756207466, + 0.06535778194665909, + 0.04095343500375748, + 0.04147273674607277, + 0.052766475826501846, + 0.08893522620201111, + 0.26036596298217773, + 0.0827026441693306 + ], + "credit_map": [ + 0.017483694478869438, + 0.009200409054756165, + 0.005681243259459734, + 0.0055971513502299786, + 0.012007446959614754, + 0.050613440573215485, + 0.024943729862570763, + 0.07676307111978531, + 0.00419717188924551, + 0.008795487694442272, + 0.006976776756346226, + 0.02337544411420822, + 0.008971953764557838, + 0.00956145953387022, + 0.010984801687300205, + 0.03168732300400734, + 0.007672000676393509, + 0.005225791595876217, + 0.006864398717880249, + 0.005313058849424124, + 0.004128511529415846, + 0.004492486827075481, + 0.009731310419738293, + 0.017177540808916092, + 0.003575837705284357, + 0.002992836758494377, + 0.019462985917925835, + 0.004697481635957956, + 0.004021184518933296, + 0.0038434125017374754, + 0.00734670227393508, + 0.01941733993589878, + 0.0036536871921271086, + 0.004491098690778017, + 0.00525414664298296, + 0.0054670674726367, + 0.006067442242056131, + 0.004850984085351229, + 0.004429774824529886, + 0.006739235483109951, + 0.003708509262651205, + 0.004285407718271017, + 0.006025299429893494, + 0.004867776297032833, + 0.003960053902119398, + 0.0026418380439281464, + 0.01025589369237423, + 0.005727958865463734, + 0.014329995959997177, + 0.005603496450930834, + 0.004762132652103901, + 0.014707512222230434, + 0.0030126087367534637, + 0.0032183181028813124, + 0.00433643301948905, + 0.002795979380607605, + 0.019013652577996254, + 0.03067004680633545, + 0.005768307950347662, + 0.009537022560834885, + 0.0077317142859101295, + 0.004924502223730087, + 0.005494813434779644, + 0.005795166362076998, + 0.06675595045089722, + 0.019709745422005653, + 0.022284405305981636, + 0.016302555799484253, + 0.007750409189611673, + 0.02031807042658329, + 0.019566161558032036, + 0.08767867833375931, + 0.008113943040370941, + 0.007925409823656082, + 0.011588300578296185, + 0.009059322066605091, + 0.008113800548017025, + 0.012662568129599094, + 0.013481459580361843, + 0.011757840402424335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.029787791892886162 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024922997690737247 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024922997690737247, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024922997690737247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5426321625709534, + 0.095687136054039, + 0.05639318749308586, + 0.10329514741897583, + 0.06334193050861359, + 0.04892420023679733, + 0.02469552494585514, + 0.020617462694644928, + 0.028064453974366188, + 0.01634872891008854 + ], + "credit_map": [ + 0.03109145164489746, + 0.035275328904390335, + 0.02280469238758087, + 0.02402130886912346, + 0.049537889659404755, + 0.17261582612991333, + 0.049361735582351685, + 0.15792395174503326, + 0.0037263042759150267, + 0.008586824871599674, + 0.013282323256134987, + 0.02979840710759163, + 0.011090213432908058, + 0.012701996602118015, + 0.009502648375928402, + 0.006998422555625439, + 0.009813586249947548, + 0.00468340003862977, + 0.005674059968441725, + 0.006469207815825939, + 0.004368141759186983, + 0.004208442755043507, + 0.007667617406696081, + 0.013508731499314308, + 0.010686879977583885, + 0.0039747185073792934, + 0.024904875084757805, + 0.005967613775283098, + 0.005649303086102009, + 0.005559738725423813, + 0.010558625683188438, + 0.03599339723587036, + 0.005508955102413893, + 0.006829143036156893, + 0.006502611096948385, + 0.00842791423201561, + 0.007510302122682333, + 0.00819653645157814, + 0.00726382015272975, + 0.01310265064239502, + 0.004538260400295258, + 0.004300815984606743, + 0.007421101443469524, + 0.005521108862012625, + 0.004848221316933632, + 0.002788929734379053, + 0.013167047873139381, + 0.006338716484606266, + 0.00397784449160099, + 0.002484425902366638, + 0.004545764997601509, + 0.0032733529806137085, + 0.002319104503840208, + 0.0014785146340727806, + 0.0028905777726322412, + 0.0037259398959577084, + 0.002316672122105956, + 0.0036301929503679276, + 0.0024381575640290976, + 0.0036577568389475346, + 0.0028931305278092623, + 0.001340080751106143, + 0.0023578228428959846, + 0.001983648631721735, + 0.005600591190159321, + 0.0017264756606891751, + 0.003445227863267064, + 0.0027821736875921488, + 0.001192170544527471, + 0.002210564911365509, + 0.003362140152603388, + 0.007745110895484686, + 0.001587736769579351, + 0.0015634770970791578, + 0.0024538354482501745, + 0.0028226268477737904, + 0.0014783126534894109, + 0.0018163727363571525, + 0.001995400059968233, + 0.0026309669483453035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0024922997690737247 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a secluded beach at a moonless, starlit night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle with a wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a powerful wave shown surging with electric blue bioluminescence that turns cyan at the crest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the receding water leave an intricate, branching pattern of light on the dark sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067337024956941605 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.0067337024956941605, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067337024956941605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14633934199810028, + 0.07691322267055511, + 0.04826481640338898, + 0.03536036238074303, + 0.22982124984264374, + 0.07238451391458511, + 0.23673945665359497, + 0.09593556076288223, + 0.04319605231285095, + 0.01504544448107481 + ], + "credit_map": [ + 0.0085978414863348, + 0.012927671894431114, + 0.008121010847389698, + 0.008046133443713188, + 0.015016783960163593, + 0.04419751465320587, + 0.012486618012189865, + 0.036945752799510956, + 0.004617359489202499, + 0.010301105678081512, + 0.011351070366799831, + 0.015826016664505005, + 0.009599564597010612, + 0.012320051901042461, + 0.005493403412401676, + 0.007404644042253494, + 0.0046772132627666, + 0.003081303322687745, + 0.006329013966023922, + 0.0034488760866224766, + 0.0035223502200096846, + 0.0031291006598621607, + 0.009023887105286121, + 0.015053068287670612, + 0.002455892274156213, + 0.0018298927461728454, + 0.010777034796774387, + 0.0033272725995630026, + 0.0021920609287917614, + 0.0019699246622622013, + 0.004535994492471218, + 0.008272288367152214, + 0.016283901408314705, + 0.025532590225338936, + 0.028466621413826942, + 0.04048510268330574, + 0.022619707509875298, + 0.04618847742676735, + 0.0225807037204504, + 0.027664119377732277, + 0.00807847362011671, + 0.00512396264821291, + 0.007998731918632984, + 0.00711346510797739, + 0.006170409265905619, + 0.004619847517460585, + 0.022267485037446022, + 0.01101213414222002, + 0.04920770972967148, + 0.02671026811003685, + 0.035608068108558655, + 0.030123921111226082, + 0.021297970786690712, + 0.017479388043284416, + 0.02577492780983448, + 0.030537199229002, + 0.02213597483932972, + 0.015553483739495277, + 0.006981806829571724, + 0.015598696656525135, + 0.014052134938538074, + 0.00536738708615303, + 0.009042368270456791, + 0.00720370514318347, + 0.010721920989453793, + 0.002980566117912531, + 0.0037368214689195156, + 0.0030284281820058823, + 0.0013638113159686327, + 0.002905344357714057, + 0.004284839611500502, + 0.014174319803714752, + 0.0016363146714866161, + 0.001390910823829472, + 0.0016966518014669418, + 0.0020504193380475044, + 0.0013563568936660886, + 0.0013806376373395324, + 0.002858842955902219, + 0.0026753093115985394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.0067337024956941605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_12/index_tensor([2563, 2563])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_12/index_tensor([2563, 2563])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..73d076fbf1d3d5e8d3aad0b7b33326c7f94a1117 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_12/index_tensor([2563, 2563])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle close-up captures a sturdy Siberian husky, its head thrown back high, as it lets out a long, powerful howl into the grayish-white sky. The husky's thick, shaggy black and white coat is speckled with glittering snowflakes. Its icy blue eyes narrow slightly as it howls with concentration, a call full of feral energy. As the howl cascades, a distinct, thick, white mist gushes continuously from its mouth, stretched wide open by the effort. This mist quickly condenses in the bitterly cold air, forming a fleeting, spectacular cloud before slowly dissipating. It finds itself in a vast expanse of ice and snow, a thick, untouched blanket of pristine white beneath its feet, and a sparse, frost-covered coniferous forest in the background. The low-slanting winter sun casts a cold light, tinging the entire scene with a light blue hue. At the same time, it outlines a circle of bright contour light on the edge of the Husky's fur and the white mist exhaled. The whole picture is quiet and powerful, conveying a sense of primitive, lonely and vital grandeur.", + "seed": 18233, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_12/index_2563_seed_18233_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.003934847866185009, + "vlm_logprob": -0.003934847866185009, + "vlm_yes_no_margin": 5.675, + "vlm_token_credit": -0.003934847866185009, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0043342225253582 + ], + "opposite_logprob_tokens": [ + -5.629334449768066 + ], + "logprob_reward": -0.0043342225253582, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0043342225253582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15639397501945496, + 0.08683747053146362, + 0.09046068042516708, + 0.13705143332481384, + 0.2657471001148224, + 0.024022817611694336, + 0.04515628144145012, + 0.047363683581352234, + 0.08956865221261978, + 0.05739787593483925 + ], + "credit_map": [ + 0.030113710090517998, + 0.01652832143008709, + 0.018350956961512566, + 0.018414411693811417, + 0.017914453521370888, + 0.01792154274880886, + 0.016824575141072273, + 0.020326009020209312, + 0.017277928069233894, + 0.008351403288543224, + 0.007890468463301659, + 0.007752466481178999, + 0.0133830476552248, + 0.010080277919769287, + 0.01264361385256052, + 0.0094582699239254, + 0.007411056663841009, + 0.006998718250542879, + 0.007515632081776857, + 0.013776270672678947, + 0.012591270729899406, + 0.01065102033317089, + 0.02119482308626175, + 0.010321884416043758, + 0.011842708103358746, + 0.01391149777919054, + 0.016286153346300125, + 0.014920340850949287, + 0.017543397843837738, + 0.017981979995965958, + 0.027898140251636505, + 0.01666722446680069, + 0.03670859336853027, + 0.029603812843561172, + 0.03369666635990143, + 0.032030899077653885, + 0.0375986248254776, + 0.04441528394818306, + 0.031806375831365585, + 0.019886866211891174, + 0.003127759788185358, + 0.002916211960837245, + 0.002073317300528288, + 0.0032986202277243137, + 0.004225684329867363, + 0.003311700187623501, + 0.0027110374066978693, + 0.00235848524607718, + 0.006856210064142942, + 0.004060540348291397, + 0.004000933840870857, + 0.004322418477386236, + 0.006282305344939232, + 0.008396643213927746, + 0.005925366654992104, + 0.005311866290867329, + 0.006916735786944628, + 0.004345351830124855, + 0.005154657643288374, + 0.005054638721048832, + 0.0065389638766646385, + 0.009004207327961922, + 0.005066730547696352, + 0.00528239831328392, + 0.009504793211817741, + 0.005890384782105684, + 0.003867963794618845, + 0.004436715971678495, + 0.00926012173295021, + 0.027676796540617943, + 0.009298413060605526, + 0.019633464515209198, + 0.008894517086446285, + 0.0059776888228952885, + 0.0048997956328094006, + 0.0084823714569211, + 0.010081210173666477, + 0.0077187614515423775, + 0.006132055073976517, + 0.005211472045630217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0043342225253582 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001947650802321732 + ], + "opposite_logprob_tokens": [ + -6.25194787979126 + ], + "logprob_reward": -0.001947650802321732, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001947650802321732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09110984206199646, + 0.21735088527202606, + 0.11472547799348831, + 0.12555056810379028, + 0.07866135984659195, + 0.039108891040086746, + 0.04601350799202919, + 0.06631752103567123, + 0.14884310960769653, + 0.07231885194778442 + ], + "credit_map": [ + 0.020250922068953514, + 0.012109953910112381, + 0.013808236457407475, + 0.008553115651011467, + 0.008721795864403248, + 0.007892767898738384, + 0.01002917718142271, + 0.009743875823915005, + 0.028514176607131958, + 0.027490733191370964, + 0.022310813888907433, + 0.020590197294950485, + 0.02884470857679844, + 0.030467020347714424, + 0.03782143443822861, + 0.021311812102794647, + 0.006662738509476185, + 0.009690046310424805, + 0.009566655382514, + 0.018008224666118622, + 0.01621430553495884, + 0.013642141595482826, + 0.02803897298872471, + 0.012902401387691498, + 0.01290242001414299, + 0.013532663695514202, + 0.01848415471613407, + 0.014165040105581284, + 0.017440559342503548, + 0.015860579907894135, + 0.02010534517467022, + 0.013059809803962708, + 0.012256599962711334, + 0.007726807147264481, + 0.011334152892231941, + 0.009673832915723324, + 0.009292823262512684, + 0.012225301936268806, + 0.008589951321482658, + 0.00756189227104187, + 0.004270793870091438, + 0.0037815487012267113, + 0.004607755690813065, + 0.007155574392527342, + 0.00507302675396204, + 0.006052298471331596, + 0.004234728869050741, + 0.003933166153728962, + 0.007961447350680828, + 0.004628049209713936, + 0.003478582249954343, + 0.005715066567063332, + 0.005510480143129826, + 0.00831183884292841, + 0.005815877579152584, + 0.0045921690762043, + 0.010334976017475128, + 0.006411680951714516, + 0.00685871159657836, + 0.006510043516755104, + 0.006936203222721815, + 0.010525933466851711, + 0.009719275869429111, + 0.009020697325468063, + 0.010958921164274216, + 0.008497142232954502, + 0.004992356989532709, + 0.008168332278728485, + 0.012450953014194965, + 0.04581707715988159, + 0.019645627588033676, + 0.03831269592046738, + 0.010526786558330059, + 0.007477555889636278, + 0.006365044973790646, + 0.007185090333223343, + 0.014494209550321102, + 0.011592414230108261, + 0.008487497456371784, + 0.006190252490341663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.001947650802321732 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256047938019037 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256047938019037, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256047938019037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13029012084007263, + 0.1255406141281128, + 0.08615749329328537, + 0.0906183123588562, + 0.21603989601135254, + 0.04201273247599602, + 0.04382934048771858, + 0.0898021012544632, + 0.12221553176641464, + 0.05349384620785713 + ], + "credit_map": [ + 0.031494490802288055, + 0.016802914440631866, + 0.015653766691684723, + 0.014725536108016968, + 0.011057456023991108, + 0.013264346867799759, + 0.011526335030794144, + 0.01576528511941433, + 0.010674865916371346, + 0.015423976816236973, + 0.009614592418074608, + 0.011444802395999432, + 0.024959880858659744, + 0.013157534413039684, + 0.025023387745022774, + 0.01524156704545021, + 0.009507459588348866, + 0.007897641509771347, + 0.01024845615029335, + 0.01149804424494505, + 0.01197489257901907, + 0.011800159700214863, + 0.012511536478996277, + 0.010719305835664272, + 0.006990464869886637, + 0.010400151833891869, + 0.008607870899140835, + 0.011703197844326496, + 0.014161483384668827, + 0.011771134100854397, + 0.018701255321502686, + 0.008282750844955444, + 0.02687249891459942, + 0.024902338162064552, + 0.029011299833655357, + 0.025301320478320122, + 0.032841652631759644, + 0.03345264494419098, + 0.02784969098865986, + 0.015808461233973503, + 0.007191641256213188, + 0.004775297362357378, + 0.004471659194678068, + 0.0045662522315979, + 0.005223048385232687, + 0.006406842730939388, + 0.004748128820210695, + 0.004629860632121563, + 0.006043699570000172, + 0.003900958923622966, + 0.003579420503228903, + 0.005607172846794128, + 0.0072753215208649635, + 0.006496079731732607, + 0.006127322092652321, + 0.004799361806362867, + 0.012721640057861805, + 0.008582551032304764, + 0.01108256634324789, + 0.01057406421750784, + 0.010267660953104496, + 0.01297096349298954, + 0.013074859976768494, + 0.010527790524065495, + 0.021242238581180573, + 0.008192206732928753, + 0.007741691544651985, + 0.016599074006080627, + 0.00997543428093195, + 0.025991670787334442, + 0.014578883536159992, + 0.017894336953759193, + 0.00945067498832941, + 0.00480084540322423, + 0.005750695709139109, + 0.009780892170965672, + 0.005706643220037222, + 0.006832708604633808, + 0.005998766981065273, + 0.005172616336494684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.005256047938019037 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942771211266518 + ], + "opposite_logprob_tokens": [ + -5.1309428215026855 + ], + "logprob_reward": -0.005942771211266518, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942771211266518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11754010617733002, + 0.07443223148584366, + 0.05057919770479202, + 0.043552178889513016, + 0.05892304703593254, + 0.04566178098320961, + 0.046192727982997894, + 0.04084830731153488, + 0.2904227375984192, + 0.23184770345687866 + ], + "credit_map": [ + 0.016236649826169014, + 0.015459118410944939, + 0.012468717992305756, + 0.01320584025233984, + 0.016887828707695007, + 0.012588105164468288, + 0.014957183040678501, + 0.0157366581261158, + 0.014566746540367603, + 0.007700613234192133, + 0.008600338362157345, + 0.00882903952151537, + 0.012030376121401787, + 0.009846202097833157, + 0.006996185518801212, + 0.005862733814865351, + 0.006364717613905668, + 0.0037591522559523582, + 0.005189657211303711, + 0.005930168088525534, + 0.006180149968713522, + 0.005844557657837868, + 0.011512636207044125, + 0.005798159167170525, + 0.0074858288280665874, + 0.0038959768135100603, + 0.009806940332055092, + 0.004496836569160223, + 0.0037229505833238363, + 0.006010028067976236, + 0.003714877413585782, + 0.004418742842972279, + 0.009831592440605164, + 0.005978431552648544, + 0.006360443774610758, + 0.008341907523572445, + 0.007777401711791754, + 0.01000976748764515, + 0.005523699335753918, + 0.005099801812320948, + 0.009862896054983139, + 0.004800313152372837, + 0.003662906354293227, + 0.004658076912164688, + 0.005751328077167273, + 0.0068475487641990185, + 0.005787327419966459, + 0.004291385877877474, + 0.007301262114197016, + 0.004499494563788176, + 0.004725923761725426, + 0.005294749978929758, + 0.006958825513720512, + 0.006691549438983202, + 0.006028902716934681, + 0.0046920208260416985, + 0.006582810543477535, + 0.0037569915875792503, + 0.004752374719828367, + 0.004096528515219688, + 0.004544488154351711, + 0.005929657258093357, + 0.006267451215535402, + 0.004918004386126995, + 0.03226596117019653, + 0.020922541618347168, + 0.011368106119334698, + 0.01200258731842041, + 0.025029156357049942, + 0.08763828128576279, + 0.03247905895113945, + 0.06871702522039413, + 0.043729715049266815, + 0.0195864699780941, + 0.01705065742135048, + 0.018501469865441322, + 0.04763437807559967, + 0.0382961742579937, + 0.023375485092401505, + 0.023673348128795624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.005942771211266518 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021935468539595604 + ], + "opposite_logprob_tokens": [ + -6.127193450927734 + ], + "logprob_reward": -0.0021935468539595604, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021935468539595604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16161304712295532, + 0.14446236193180084, + 0.05797848105430603, + 0.043728746473789215, + 0.10382801294326782, + 0.04483639821410179, + 0.06853675842285156, + 0.060963548719882965, + 0.21174877882003784, + 0.10230383276939392 + ], + "credit_map": [ + 0.0292058065533638, + 0.02409827709197998, + 0.022429727017879486, + 0.018373947590589523, + 0.019010700285434723, + 0.01838579773902893, + 0.01303273905068636, + 0.01707606576383114, + 0.025616293773055077, + 0.016207892447710037, + 0.017325878143310547, + 0.01174256019294262, + 0.019581645727157593, + 0.01823127269744873, + 0.02045910805463791, + 0.015297708101570606, + 0.005948726553469896, + 0.005348083097487688, + 0.006889082957059145, + 0.008415406569838524, + 0.008502195589244366, + 0.00695457449182868, + 0.010352998040616512, + 0.0055674160830676556, + 0.006397245917469263, + 0.005150237120687962, + 0.006105166859924793, + 0.005829592235386372, + 0.004409974440932274, + 0.005986477714031935, + 0.0053020971827209, + 0.0045479582622647285, + 0.012634656392037868, + 0.01190148014575243, + 0.013843432068824768, + 0.013689469546079636, + 0.014211989007890224, + 0.01640513353049755, + 0.012614866718649864, + 0.008527001366019249, + 0.007115548010915518, + 0.004158186726272106, + 0.004520525690168142, + 0.0057864016853272915, + 0.005647046957165003, + 0.006614288315176964, + 0.0066249435767531395, + 0.0043694619089365005, + 0.010941093787550926, + 0.005610684398561716, + 0.005227676592767239, + 0.008507668040692806, + 0.011170604266226292, + 0.011070326901972294, + 0.009578902274370193, + 0.006429806351661682, + 0.011159062385559082, + 0.005189420655369759, + 0.00747386459261179, + 0.00850950088351965, + 0.005795640405267477, + 0.009361037984490395, + 0.0063926042057573795, + 0.007082420401275158, + 0.022866306826472282, + 0.015834152698516846, + 0.011389446444809437, + 0.013193276710808277, + 0.021558519452810287, + 0.06193729490041733, + 0.020304130390286446, + 0.044665656983852386, + 0.01693580113351345, + 0.0070276339538395405, + 0.010589527897536755, + 0.013046883046627045, + 0.011335814371705055, + 0.01822829432785511, + 0.01401696540415287, + 0.011122921481728554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0021935468539595604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up captures a sturdy Siberian husky, its head thrown back high, as it lets out a long, powerful howl into the grayish-white sky. The husky's thick, shaggy black and white coat is speckled with glittering snowflakes. Its icy blue eyes narrow slightly as it howls with concentration, a call full of feral energy. As the howl cascades, a distinct, thick, white mist gushes continuously from its mouth, stretched wide open by the effort. This mist quickly condenses in the bitterly cold air, forming a fleeting, spectacular cloud before slowly dissipating. It finds itself in a vast expanse of ice and snow, a thick, untouched blanket of pristine white beneath its feet, and a sparse, frost-covered coniferous forest in the background. The low-slanting winter sun casts a cold light, tinging the entire scene with a light blue hue. At the same time, it outlines a circle of bright contour light on the edge of the Husky's fur and the white mist exhaled. The whole picture is quiet and powerful, conveying a sense of primitive, lonely and vital grandeur.", + "seed": 18234, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_12/index_2563_seed_18234_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.002342010603751987, + "vlm_logprob": -0.002342010603751987, + "vlm_yes_no_margin": 6.45, + "vlm_token_credit": -0.002342010603751987, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003456453327089548 + ], + "opposite_logprob_tokens": [ + -6.5034565925598145 + ], + "logprob_reward": -0.003456453327089548, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.003456453327089548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11975059658288956, + 0.06404124200344086, + 0.12920567393302917, + 0.055125992745161057, + 0.16432391107082367, + 0.09105561673641205, + 0.05184054374694824, + 0.11583102494478226, + 0.1704121083021164, + 0.03841330483555794 + ], + "credit_map": [ + 0.019530467689037323, + 0.017661917954683304, + 0.011516190133988857, + 0.013548845425248146, + 0.013768360950052738, + 0.013635714538395405, + 0.011364871636033058, + 0.01872422732412815, + 0.010977471247315407, + 0.007577759213745594, + 0.005619545001536608, + 0.00801867339760065, + 0.00661089550703764, + 0.008511567488312721, + 0.0048345476388931274, + 0.011890782043337822, + 0.026327062398195267, + 0.012063590809702873, + 0.009628897532820702, + 0.008746493607759476, + 0.01223045028746128, + 0.017579127103090286, + 0.008479205891489983, + 0.034150853753089905, + 0.007737657520920038, + 0.007819670252501965, + 0.004807416815310717, + 0.008427360095083714, + 0.00799927581101656, + 0.005282910540699959, + 0.006383311469107866, + 0.0066683911718428135, + 0.02352731116116047, + 0.023832328617572784, + 0.01366764958947897, + 0.024438127875328064, + 0.02564183436334133, + 0.014346635900437832, + 0.022063307464122772, + 0.016806710511446, + 0.01306022796779871, + 0.012463259510695934, + 0.010943750850856304, + 0.013596192933619022, + 0.011354212649166584, + 0.009086865931749344, + 0.009315446950495243, + 0.01123566273599863, + 0.007686820346862078, + 0.010018739849328995, + 0.0051360721699893475, + 0.004723333287984133, + 0.008327658288180828, + 0.004416777286678553, + 0.005675668362528086, + 0.005855473689734936, + 0.017516225576400757, + 0.01533826719969511, + 0.01108668465167284, + 0.019810203462839127, + 0.008922538720071316, + 0.014016618020832539, + 0.00941444467753172, + 0.019726039841771126, + 0.03183551877737045, + 0.01276841014623642, + 0.015609282068908215, + 0.008922026492655277, + 0.011382799595594406, + 0.023752178996801376, + 0.01758413016796112, + 0.048557769507169724, + 0.00673640426248312, + 0.003193250158801675, + 0.004842950962483883, + 0.0030695032328367233, + 0.004280019551515579, + 0.006038009189069271, + 0.00258997012861073, + 0.0076632010750472546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.003456453327089548 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011879300000146031 + ], + "opposite_logprob_tokens": [ + -6.751187801361084 + ], + "logprob_reward": -0.0011879300000146031, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011879300000146031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07308486849069595, + 0.1344972401857376, + 0.14998088777065277, + 0.17430655658245087, + 0.05686187371611595, + 0.05429285019636154, + 0.13998670876026154, + 0.1132052093744278, + 0.05948245897889137, + 0.04430130496621132 + ], + "credit_map": [ + 0.01297500915825367, + 0.010711169801652431, + 0.009874938987195492, + 0.0076520079746842384, + 0.010281180031597614, + 0.007386939134448767, + 0.006804992910474539, + 0.007398633286356926, + 0.01903490349650383, + 0.019607823342084885, + 0.016684509813785553, + 0.017964649945497513, + 0.0182205718010664, + 0.015135256573557854, + 0.011674970388412476, + 0.016174549236893654, + 0.03228849545121193, + 0.01082804799079895, + 0.010757760144770145, + 0.00860194955021143, + 0.012006869539618492, + 0.022178547456860542, + 0.010174578055739403, + 0.043144650757312775, + 0.0263966117054224, + 0.030691130086779594, + 0.011127663776278496, + 0.02318762242794037, + 0.040339481085538864, + 0.013440418057143688, + 0.016623957082629204, + 0.012499681673943996, + 0.009111919440329075, + 0.007823946885764599, + 0.005814654752612114, + 0.006785571575164795, + 0.007496388163417578, + 0.004572881385684013, + 0.00454140966758132, + 0.010715105570852757, + 0.007561865262687206, + 0.008983858861029148, + 0.006760431919246912, + 0.0066078281961381435, + 0.00620768778026104, + 0.005971614737063646, + 0.006112121976912022, + 0.006087436806410551, + 0.020479843020439148, + 0.02138650044798851, + 0.014572888612747192, + 0.016080761328339577, + 0.022133510559797287, + 0.010661310516297817, + 0.018527766689658165, + 0.01614413410425186, + 0.01760370470583439, + 0.013235442340373993, + 0.010038516484200954, + 0.013075310736894608, + 0.011505522765219212, + 0.012585706077516079, + 0.01729314774274826, + 0.017867857590317726, + 0.009445021860301495, + 0.008547322824597359, + 0.007427195552736521, + 0.006215576082468033, + 0.007604664191603661, + 0.006177323870360851, + 0.007523676380515099, + 0.0065416814759373665, + 0.006632332690060139, + 0.007936548441648483, + 0.006801614072173834, + 0.0051201991736888885, + 0.0069152251817286015, + 0.003156829858198762, + 0.0037741726264357567, + 0.003964381292462349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.0011879300000146031 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464016804471612 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.00464016804471612, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464016804471612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09543837606906891, + 0.06291785091161728, + 0.14757147431373596, + 0.2020525485277176, + 0.100045345723629, + 0.032425474375486374, + 0.12028404325246811, + 0.08211687952280045, + 0.1345611959695816, + 0.022586777806282043 + ], + "credit_map": [ + 0.014118636958301067, + 0.013336882926523685, + 0.007472506258636713, + 0.01251189224421978, + 0.014417862519621849, + 0.008643798530101776, + 0.010378934442996979, + 0.014557857066392899, + 0.0093613900244236, + 0.008835501968860626, + 0.008132090792059898, + 0.008275477215647697, + 0.007513856049627066, + 0.006600420922040939, + 0.006620694883167744, + 0.007578419521450996, + 0.027762165293097496, + 0.014103610999882221, + 0.014098639599978924, + 0.012250253930687904, + 0.015301907435059547, + 0.021593963727355003, + 0.00981023721396923, + 0.03265069052577019, + 0.030465688556432724, + 0.03792756050825119, + 0.013440032489597797, + 0.0267210491001606, + 0.043077316135168076, + 0.016346855089068413, + 0.01678745076060295, + 0.017286591231822968, + 0.01475444994866848, + 0.01457129418849945, + 0.006798170506954193, + 0.015304431319236755, + 0.015090136788785458, + 0.009045159444212914, + 0.012523533776402473, + 0.011958174407482147, + 0.005013854242861271, + 0.004751773085445166, + 0.004316478502005339, + 0.00392674095928669, + 0.004005069844424725, + 0.0035119401291012764, + 0.0033389225136488676, + 0.0035606957972049713, + 0.017346695065498352, + 0.02266150526702404, + 0.011532260105013847, + 0.01744406670331955, + 0.01690591126680374, + 0.010727938264608383, + 0.01282783318310976, + 0.010837827809154987, + 0.015701042488217354, + 0.007350130472332239, + 0.007758428808301687, + 0.005529152229428291, + 0.006047816015779972, + 0.011342165991663933, + 0.007602074649184942, + 0.02078607678413391, + 0.02603474259376526, + 0.01144112553447485, + 0.011145045049488544, + 0.007745533250272274, + 0.008998818695545197, + 0.018578050658106804, + 0.010756067931652069, + 0.039861805737018585, + 0.003501678118482232, + 0.0019051707349717617, + 0.005019808188080788, + 0.0030830674804747105, + 0.0018975009443238378, + 0.0022758450359106064, + 0.0021686037071049213, + 0.0027351034805178642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.00464016804471612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000715714821126312 + ], + "opposite_logprob_tokens": [ + -7.250715732574463 + ], + "logprob_reward": -0.000715714821126312, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.000715714821126312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10353238135576248, + 0.21011827886104584, + 0.09981939196586609, + 0.04917549341917038, + 0.07119770348072052, + 0.06782492995262146, + 0.04928124323487282, + 0.2669637203216553, + 0.04803340509533882, + 0.03405342996120453 + ], + "credit_map": [ + 0.01752137392759323, + 0.009370272047817707, + 0.012684808112680912, + 0.008959419094026089, + 0.007659535855054855, + 0.014159295707941055, + 0.007315908093005419, + 0.025861769914627075, + 0.0359567366540432, + 0.01993427239358425, + 0.019331108778715134, + 0.021326297894120216, + 0.022855697199702263, + 0.031824834644794464, + 0.016661960631608963, + 0.0422273650765419, + 0.016832219436764717, + 0.014644834212958813, + 0.00901492778211832, + 0.007759059779345989, + 0.022168844938278198, + 0.00958565715700388, + 0.0081087751314044, + 0.011705072596669197, + 0.007057112641632557, + 0.007967548444867134, + 0.004085296764969826, + 0.005926811136305332, + 0.007909572683274746, + 0.00537273520603776, + 0.0044403993524611, + 0.006416018586605787, + 0.009784025140106678, + 0.009096985682845116, + 0.009500919841229916, + 0.007865461520850658, + 0.008615122176706791, + 0.008702883496880531, + 0.006240211892873049, + 0.01139209233224392, + 0.010914829559624195, + 0.007965399883687496, + 0.006912405602633953, + 0.007389952894300222, + 0.010117966681718826, + 0.007173130288720131, + 0.005831078626215458, + 0.011520159430801868, + 0.007722511421889067, + 0.00627601332962513, + 0.005759915802627802, + 0.004827795084565878, + 0.005191562697291374, + 0.005462891422212124, + 0.004268747754395008, + 0.009771806187927723, + 0.05592332035303116, + 0.016306716948747635, + 0.021216392517089844, + 0.014710037037730217, + 0.012489370070397854, + 0.045334793627262115, + 0.01242667157202959, + 0.08855641633272171, + 0.009186089038848877, + 0.003383021801710129, + 0.004468587227165699, + 0.0032174985390156507, + 0.00258769653737545, + 0.0071381754241883755, + 0.0036902478896081448, + 0.014362086541950703, + 0.00701608881354332, + 0.002207377692684531, + 0.0037835403345525265, + 0.0021310527808964252, + 0.0019463237840682268, + 0.005409492179751396, + 0.00214761053211987, + 0.009411945939064026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.000715714821126312 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001709786825813353 + ], + "opposite_logprob_tokens": [ + -6.376709938049316 + ], + "logprob_reward": -0.001709786825813353, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001709786825813353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1354427933692932, + 0.05693720653653145, + 0.10629970580339432, + 0.041803695261478424, + 0.11314172297716141, + 0.04147094488143921, + 0.062106046825647354, + 0.1107247993350029, + 0.2828260362148285, + 0.049246981739997864 + ], + "credit_map": [ + 0.021111732348799706, + 0.012120774015784264, + 0.01776970736682415, + 0.018362965434789658, + 0.009937215596437454, + 0.018094923347234726, + 0.01543371006846428, + 0.022611763328313828, + 0.010522784665226936, + 0.006015297956764698, + 0.004123545251786709, + 0.00558081828057766, + 0.004708603024482727, + 0.007853731513023376, + 0.0036235188599675894, + 0.014508903957903385, + 0.02109174244105816, + 0.012249945662915707, + 0.008990359492599964, + 0.007060805801302195, + 0.014469963498413563, + 0.013477170839905739, + 0.008086242713034153, + 0.020873483270406723, + 0.006332071963697672, + 0.00516898138448596, + 0.0046686334535479546, + 0.004821307025849819, + 0.008336330763995647, + 0.004461789503693581, + 0.0034568379633128643, + 0.004557744599878788, + 0.01882023550570011, + 0.020550183951854706, + 0.011059893295168877, + 0.011195976287126541, + 0.021099962294101715, + 0.008163905702531338, + 0.008808145299553871, + 0.013443420641124249, + 0.006442133337259293, + 0.004927137400954962, + 0.00493254791945219, + 0.00520862964913249, + 0.004098384641110897, + 0.004212580155581236, + 0.003299573203548789, + 0.00834995973855257, + 0.00958251953125, + 0.008799472823739052, + 0.005283785052597523, + 0.006287865806370974, + 0.008570880629122257, + 0.006017325446009636, + 0.008859211578965187, + 0.008704984560608864, + 0.018112285062670708, + 0.011620249599218369, + 0.014134335331618786, + 0.01198336947709322, + 0.008830985985696316, + 0.01402110792696476, + 0.009770111180841923, + 0.022252345457673073, + 0.056336015462875366, + 0.015786834061145782, + 0.027148429304361343, + 0.013458752073347569, + 0.011024799197912216, + 0.044112104922533035, + 0.020113427191972733, + 0.09484566003084183, + 0.009254281409084797, + 0.0031292575877159834, + 0.006150119472295046, + 0.0028727613389492035, + 0.004170124884694815, + 0.0070479693822562695, + 0.0031620420049875975, + 0.013460428453981876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.001709786825813353 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up captures a sturdy Siberian husky, its head thrown back high, as it lets out a long, powerful howl into the grayish-white sky. The husky's thick, shaggy black and white coat is speckled with glittering snowflakes. Its icy blue eyes narrow slightly as it howls with concentration, a call full of feral energy. As the howl cascades, a distinct, thick, white mist gushes continuously from its mouth, stretched wide open by the effort. This mist quickly condenses in the bitterly cold air, forming a fleeting, spectacular cloud before slowly dissipating. It finds itself in a vast expanse of ice and snow, a thick, untouched blanket of pristine white beneath its feet, and a sparse, frost-covered coniferous forest in the background. The low-slanting winter sun casts a cold light, tinging the entire scene with a light blue hue. At the same time, it outlines a circle of bright contour light on the edge of the Husky's fur and the white mist exhaled. The whole picture is quiet and powerful, conveying a sense of primitive, lonely and vital grandeur.", + "seed": 18235, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_12/index_2563_seed_18235_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0018679224886000156, + "vlm_logprob": -0.0018679224886000156, + "vlm_yes_no_margin": 6.65, + "vlm_token_credit": -0.0018679224886000156, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0033812513574957848 + ], + "opposite_logprob_tokens": [ + -6.753381252288818 + ], + "logprob_reward": -0.0033812513574957848, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0033812513574957848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08476363867521286, + 0.08970050513744354, + 0.0622517466545105, + 0.05880378931760788, + 0.168406680226326, + 0.1668616682291031, + 0.0669160708785057, + 0.05859699845314026, + 0.11770176142454147, + 0.12599708139896393 + ], + "credit_map": [ + 0.00954436231404543, + 0.009764231741428375, + 0.012967297807335854, + 0.012422007508575916, + 0.00790328998118639, + 0.009987319819629192, + 0.009916892275214195, + 0.012258228845894337, + 0.007984667085111141, + 0.02005685493350029, + 0.010576658882200718, + 0.013397550210356712, + 0.013070674613118172, + 0.007862089201807976, + 0.008517610840499401, + 0.008234404027462006, + 0.005996754392981529, + 0.010545852594077587, + 0.006485292688012123, + 0.006768316030502319, + 0.010198197327554226, + 0.007940663956105709, + 0.00632117222994566, + 0.00799549650400877, + 0.009057510644197464, + 0.005494591314345598, + 0.006550658028572798, + 0.009235086850821972, + 0.004471160937100649, + 0.009929093532264233, + 0.003988087642937899, + 0.010077601298689842, + 0.037509460002183914, + 0.018175272271037102, + 0.015546627342700958, + 0.025976937264204025, + 0.016612453386187553, + 0.02467598021030426, + 0.013998424634337425, + 0.015911512076854706, + 0.02469535358250141, + 0.014626417309045792, + 0.02127898298203945, + 0.025371422991156578, + 0.016018012538552284, + 0.021474871784448624, + 0.029599769040942192, + 0.013796831481158733, + 0.007345902267843485, + 0.005853768438100815, + 0.008201033808290958, + 0.01029219850897789, + 0.01096807699650526, + 0.00801530759781599, + 0.005981591995805502, + 0.01025819405913353, + 0.009376056492328644, + 0.00558923976495862, + 0.005983831360936165, + 0.008038320578634739, + 0.009856509044766426, + 0.009657529182732105, + 0.005049020983278751, + 0.005046489182859659, + 0.010700537823140621, + 0.014732941053807735, + 0.013739529997110367, + 0.013261192478239536, + 0.026495566591620445, + 0.020814690738916397, + 0.008820895105600357, + 0.009136403910815716, + 0.011026001535356045, + 0.02173716016113758, + 0.01654399000108242, + 0.01797056943178177, + 0.020010441541671753, + 0.012443632818758488, + 0.01518872007727623, + 0.01107657328248024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.0033812513574957848 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006413786904886365 + ], + "opposite_logprob_tokens": [ + -7.375641345977783 + ], + "logprob_reward": -0.0006413786904886365, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006413786904886365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09416396170854568, + 0.08891913294792175, + 0.06540485471487045, + 0.06131107732653618, + 0.14784041047096252, + 0.032343316823244095, + 0.0729721188545227, + 0.14571857452392578, + 0.13815699517726898, + 0.1531696319580078 + ], + "credit_map": [ + 0.017220890149474144, + 0.011246394366025925, + 0.013548572547733784, + 0.009723193943500519, + 0.010282764211297035, + 0.011838855221867561, + 0.009080067276954651, + 0.01122321654111147, + 0.013242091983556747, + 0.017924554646015167, + 0.011463767848908901, + 0.009935508482158184, + 0.01030683796852827, + 0.008568291552364826, + 0.0066476184874773026, + 0.010830446146428585, + 0.006719494238495827, + 0.01101768296211958, + 0.008055203594267368, + 0.007153449580073357, + 0.010272057726979256, + 0.008322883397340775, + 0.00620685052126646, + 0.007657230366021395, + 0.008462953381240368, + 0.007803242653608322, + 0.006161773577332497, + 0.007007690612226725, + 0.006811601109802723, + 0.012286308221518993, + 0.004346712492406368, + 0.008430791087448597, + 0.032170116901397705, + 0.017065422609448433, + 0.01247285958379507, + 0.021403813734650612, + 0.01349076721817255, + 0.020634599030017853, + 0.014030402526259422, + 0.016572417691349983, + 0.005030554253607988, + 0.00445630494505167, + 0.002570204436779022, + 0.004876964725553989, + 0.005936300382018089, + 0.004938150756061077, + 0.002181062940508127, + 0.0023537722881883383, + 0.007880873046815395, + 0.006423118524253368, + 0.010142214596271515, + 0.0120727913454175, + 0.010848725214600563, + 0.009618332609534264, + 0.005530421156436205, + 0.010455640964210033, + 0.009634768590331078, + 0.01822110451757908, + 0.013854606077075005, + 0.02755439467728138, + 0.02725648134946823, + 0.018215734511613846, + 0.01765969954431057, + 0.013321757316589355, + 0.015647796913981438, + 0.014514777809381485, + 0.006729412358254194, + 0.01929461769759655, + 0.03652804344892502, + 0.019833670929074287, + 0.01453691627830267, + 0.011071749031543732, + 0.012997477315366268, + 0.040346838533878326, + 0.012853390537202358, + 0.026555126532912254, + 0.019254036247730255, + 0.01758364960551262, + 0.01165776140987873, + 0.01192134153097868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.0006413786904886365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201278228312731 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.002201278228312731, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201278228312731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05574909597635269, + 0.04630273953080177, + 0.12739071249961853, + 0.02647191658616066, + 0.041094519197940826, + 0.01824011839926243, + 0.10016059875488281, + 0.11040456593036652, + 0.44974368810653687, + 0.024442052468657494 + ], + "credit_map": [ + 0.007480288855731487, + 0.005515436176210642, + 0.00967137049883604, + 0.012796100229024887, + 0.004198428243398666, + 0.005380427930504084, + 0.005390545818954706, + 0.005316498689353466, + 0.007778739091008902, + 0.006385397631675005, + 0.004885273054242134, + 0.006075840909034014, + 0.005950840190052986, + 0.006436959374696016, + 0.004032621160149574, + 0.004757072310894728, + 0.010044913738965988, + 0.023333631455898285, + 0.015475359745323658, + 0.020874321460723877, + 0.015746645629405975, + 0.011023995466530323, + 0.01721072569489479, + 0.013681132346391678, + 0.005200749263167381, + 0.0022163859102874994, + 0.0025066279340535402, + 0.0022041499614715576, + 0.002224014140665531, + 0.006412767339497805, + 0.002411783905699849, + 0.0032954385969787836, + 0.008086305111646652, + 0.004431451670825481, + 0.003549354849383235, + 0.006960035767406225, + 0.0037318808026611805, + 0.005387554876506329, + 0.004416999872773886, + 0.004530936945229769, + 0.003573137102648616, + 0.002665742067620158, + 0.0017086915904656053, + 0.0018212436698377132, + 0.0023565248120576143, + 0.0024549581576138735, + 0.0015720835654065013, + 0.002087736502289772, + 0.010904482565820217, + 0.008429309353232384, + 0.016247611492872238, + 0.015547902323305607, + 0.01444587204605341, + 0.012658122926950455, + 0.008245236240327358, + 0.013682057149708271, + 0.015921171754598618, + 0.011535133235156536, + 0.010170754045248032, + 0.013372542336583138, + 0.02153703011572361, + 0.018901869654655457, + 0.010905447416007519, + 0.008060612715780735, + 0.04578792676329613, + 0.05318097770214081, + 0.027461491525173187, + 0.046014416962862015, + 0.13087202608585358, + 0.07556609064340591, + 0.03943466767668724, + 0.03142610937356949, + 0.005672204773873091, + 0.0030890582129359245, + 0.002367905806750059, + 0.002430625492706895, + 0.0024952800013124943, + 0.0033662784844636917, + 0.002784872893244028, + 0.002235825639218092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.002201278228312731 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178285456262529 + ], + "opposite_logprob_tokens": [ + -6.75117826461792 + ], + "logprob_reward": -0.001178285456262529, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178285456262529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052832819521427155, + 0.07850334793329239, + 0.04575600102543831, + 0.07088852673768997, + 0.04955046996474266, + 0.051786430180072784, + 0.1696045994758606, + 0.039151355624198914, + 0.3247242271900177, + 0.11720225214958191 + ], + "credit_map": [ + 0.008747785352170467, + 0.00683686975389719, + 0.007625572383403778, + 0.007237498182803392, + 0.005200409330427647, + 0.007292852271348238, + 0.00422705477103591, + 0.005664777476340532, + 0.010585012845695019, + 0.0119757279753685, + 0.009240051731467247, + 0.009587778709828854, + 0.010037743486464024, + 0.010933979414403439, + 0.006502917036414146, + 0.009640131145715714, + 0.006557026412338018, + 0.007121099159121513, + 0.004958873149007559, + 0.005005366168916225, + 0.005405598785728216, + 0.006126612890511751, + 0.00465457933023572, + 0.005926840007305145, + 0.011802135035395622, + 0.006624983157962561, + 0.006827031262218952, + 0.008772630244493484, + 0.006334326695650816, + 0.01584750786423683, + 0.003901346353814006, + 0.010778557509183884, + 0.010752084665000439, + 0.005483668297529221, + 0.0038955355994403362, + 0.009570208378136158, + 0.0050178412348032, + 0.00673720333725214, + 0.004050317220389843, + 0.0040436056442558765, + 0.01329292543232441, + 0.0069947917945683, + 0.006523548159748316, + 0.004408891312777996, + 0.0053375838324427605, + 0.006930993869900703, + 0.0038691468071192503, + 0.00442854268476367, + 0.02098168060183525, + 0.017948176711797714, + 0.026743043214082718, + 0.02272125519812107, + 0.02196182683110237, + 0.027946213260293007, + 0.012114180251955986, + 0.019188201054930687, + 0.0043151890859007835, + 0.004803959745913744, + 0.0037284879945218563, + 0.00460553914308548, + 0.008697226643562317, + 0.005724721122533083, + 0.0038273714017122984, + 0.003448860254138708, + 0.03279302269220352, + 0.040519438683986664, + 0.02005426399409771, + 0.03376971185207367, + 0.09344938397407532, + 0.05467883124947548, + 0.027009600773453712, + 0.022449981421232224, + 0.019676093012094498, + 0.021513557061553, + 0.011905730701982975, + 0.014703569933772087, + 0.011548823677003384, + 0.017744317650794983, + 0.010239381343126297, + 0.009870770387351513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.001178285456262529 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019374187104403973 + ], + "opposite_logprob_tokens": [ + -6.251937389373779 + ], + "logprob_reward": -0.0019374187104403973, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019374187104403973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06644409149885178, + 0.07327534258365631, + 0.05159887298941612, + 0.05847697705030441, + 0.16676677763462067, + 0.0849132314324379, + 0.06976647675037384, + 0.03739088028669357, + 0.2614491581916809, + 0.1299182027578354 + ], + "credit_map": [ + 0.009565540589392185, + 0.007332096341997385, + 0.009305939078330994, + 0.00975601002573967, + 0.008541344664990902, + 0.00809081457555294, + 0.006059390492737293, + 0.007792952936142683, + 0.0070647126995027065, + 0.01662648469209671, + 0.009173242375254631, + 0.011440908536314964, + 0.008409391157329082, + 0.008544187992811203, + 0.005647395271807909, + 0.006369025446474552, + 0.007390429265797138, + 0.008629292249679565, + 0.00523683475330472, + 0.007172641344368458, + 0.006612132769078016, + 0.005325519014149904, + 0.004210283048450947, + 0.007021739613264799, + 0.009525329805910587, + 0.004085906781256199, + 0.005281781777739525, + 0.011261921375989914, + 0.004322071094065905, + 0.010695910081267357, + 0.004150220658630133, + 0.009153836406767368, + 0.03456634283065796, + 0.018923357129096985, + 0.014663802459836006, + 0.02649020589888096, + 0.01689336635172367, + 0.023951208218932152, + 0.01569511368870735, + 0.015583382919430733, + 0.018947673961520195, + 0.009800623171031475, + 0.010775553062558174, + 0.008318427950143814, + 0.009950809180736542, + 0.012867718935012817, + 0.006602994166314602, + 0.007649434730410576, + 0.007349995896220207, + 0.007495314814150333, + 0.008550688624382019, + 0.010845805518329144, + 0.010546885430812836, + 0.011924460530281067, + 0.0042912825010716915, + 0.008762041106820107, + 0.0038247613701969385, + 0.004034302197396755, + 0.00386389484629035, + 0.007151392288506031, + 0.004933356773108244, + 0.005008149426430464, + 0.004302435554563999, + 0.004272588528692722, + 0.025935979560017586, + 0.033387377858161926, + 0.0168742835521698, + 0.02626684121787548, + 0.07587993144989014, + 0.044532615691423416, + 0.020634034648537636, + 0.01793810911476612, + 0.011314298026263714, + 0.024284031242132187, + 0.009177852421998978, + 0.011309061199426651, + 0.0309382826089859, + 0.024523481726646423, + 0.008244500495493412, + 0.010126687586307526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.0019374187104403973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle close-up captures a sturdy Siberian husky, its head thrown back high, as it lets out a long, powerful howl into the grayish-white sky. The husky's thick, shaggy black and white coat is speckled with glittering snowflakes. Its icy blue eyes narrow slightly as it howls with concentration, a call full of feral energy. As the howl cascades, a distinct, thick, white mist gushes continuously from its mouth, stretched wide open by the effort. This mist quickly condenses in the bitterly cold air, forming a fleeting, spectacular cloud before slowly dissipating. It finds itself in a vast expanse of ice and snow, a thick, untouched blanket of pristine white beneath its feet, and a sparse, frost-covered coniferous forest in the background. The low-slanting winter sun casts a cold light, tinging the entire scene with a light blue hue. At the same time, it outlines a circle of bright contour light on the edge of the Husky's fur and the white mist exhaled. The whole picture is quiet and powerful, conveying a sense of primitive, lonely and vital grandeur.", + "seed": 18236, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_12/index_2563_seed_18236_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0017493066261522472, + "vlm_logprob": -0.0017493066261522472, + "vlm_yes_no_margin": 6.675, + "vlm_token_credit": -0.0017493066261522472, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021018334664404392 + ], + "opposite_logprob_tokens": [ + -7.002101898193359 + ], + "logprob_reward": -0.0021018334664404392, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0021018334664404392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22459685802459717, + 0.0643848180770874, + 0.07496615499258041, + 0.06801052391529083, + 0.06561390310525894, + 0.11246632784605026, + 0.1378990262746811, + 0.06490243971347809, + 0.0425388365983963, + 0.1446211040019989 + ], + "credit_map": [ + 0.03445373848080635, + 0.02139381133019924, + 0.026958797127008438, + 0.021005935966968536, + 0.020667046308517456, + 0.037922702729701996, + 0.0385267473757267, + 0.02366807870566845, + 0.006219613831490278, + 0.008180026896297932, + 0.008025371469557285, + 0.006887969095259905, + 0.006520003546029329, + 0.009205910377204418, + 0.00896546896547079, + 0.01038045808672905, + 0.009004303254187107, + 0.016475122421979904, + 0.011027607135474682, + 0.004711401183158159, + 0.009494596160948277, + 0.01065327413380146, + 0.007089549675583839, + 0.006510297767817974, + 0.013354451395571232, + 0.0069614555686712265, + 0.01119246520102024, + 0.004413166083395481, + 0.0072839823551476, + 0.010004282929003239, + 0.00541536184027791, + 0.009385359473526478, + 0.005859732162207365, + 0.0072862389497458935, + 0.009974689222872257, + 0.005808261223137379, + 0.008444768376648426, + 0.01071532815694809, + 0.007446266710758209, + 0.010078618302941322, + 0.01783827133476734, + 0.01870109513401985, + 0.016954205930233, + 0.007048854138702154, + 0.01128091849386692, + 0.016508745029568672, + 0.009899789467453957, + 0.014234446920454502, + 0.017045166343450546, + 0.016890594735741615, + 0.01910686120390892, + 0.013589995913207531, + 0.010989164933562279, + 0.025579536333680153, + 0.013790739700198174, + 0.02090696059167385, + 0.009231481701135635, + 0.007857663556933403, + 0.009303553961217403, + 0.006311087869107723, + 0.005972613114863634, + 0.012666038237512112, + 0.006087564397603273, + 0.007472436409443617, + 0.005245898850262165, + 0.0036866096779704094, + 0.006627222057431936, + 0.004014737904071808, + 0.003460071049630642, + 0.00800244975835085, + 0.0037564386148005724, + 0.007745410315692425, + 0.018706459552049637, + 0.014335314743220806, + 0.02114642783999443, + 0.012757592834532261, + 0.017224982380867004, + 0.028168458491563797, + 0.014184260740876198, + 0.01809760555624962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.0021018334664404392 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005656072753481567 + ], + "opposite_logprob_tokens": [ + -7.500565528869629 + ], + "logprob_reward": -0.0005656072753481567, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005656072753481567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42255404591560364, + 0.06897526234388351, + 0.04444853588938713, + 0.03922845795750618, + 0.16360324621200562, + 0.06796527653932571, + 0.04769343510270119, + 0.04423113912343979, + 0.05546743795275688, + 0.045833222568035126 + ], + "credit_map": [ + 0.04503772035241127, + 0.040193744003772736, + 0.04756621643900871, + 0.046915266662836075, + 0.07272934168577194, + 0.05779104307293892, + 0.06921510398387909, + 0.04310557618737221, + 0.005627000704407692, + 0.007747368421405554, + 0.008623712696135044, + 0.006685545668005943, + 0.011464876122772694, + 0.009889804758131504, + 0.009252956137061119, + 0.009683995507657528, + 0.0057169850915670395, + 0.008016486652195454, + 0.00648385938256979, + 0.0027558389119803905, + 0.004066916182637215, + 0.006354144308716059, + 0.0042805722914636135, + 0.0067737349309027195, + 0.003454294754192233, + 0.0035147820599377155, + 0.00503942696377635, + 0.0045887744054198265, + 0.006993560120463371, + 0.0036817272193729877, + 0.0071551017463207245, + 0.004800789523869753, + 0.016505073755979538, + 0.017596131190657616, + 0.02270369976758957, + 0.014044282957911491, + 0.028053386136889458, + 0.02407299354672432, + 0.019973328337073326, + 0.020654363557696342, + 0.0064154379069805145, + 0.009561517275869846, + 0.008085107430815697, + 0.00788558553904295, + 0.009707258082926273, + 0.009069669060409069, + 0.009055942296981812, + 0.008184758014976978, + 0.006172392517328262, + 0.006690583191812038, + 0.006768651772290468, + 0.003647923469543457, + 0.004808061756193638, + 0.007629941683262587, + 0.004328708630055189, + 0.0076471720822155476, + 0.006051918957382441, + 0.008295958861708641, + 0.006848062388598919, + 0.002901952713727951, + 0.0029398505575954914, + 0.007057648617774248, + 0.0032428724225610495, + 0.006892872974276543, + 0.007521416060626507, + 0.004387808032333851, + 0.009152506478130817, + 0.00469540199264884, + 0.004593541380017996, + 0.010076481848955154, + 0.004414423834532499, + 0.010625860653817654, + 0.006250798236578703, + 0.004015044774860144, + 0.007354495581239462, + 0.004214010201394558, + 0.00503219710662961, + 0.009223410859704018, + 0.004429379478096962, + 0.005313885398209095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0005656072753481567 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031928059179335833 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031928059179335833, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031928059179335833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45407456159591675, + 0.06520634144544601, + 0.04473862051963806, + 0.06825578957796097, + 0.11307398229837418, + 0.068276047706604, + 0.029628688469529152, + 0.07297530770301819, + 0.03555857762694359, + 0.04821205884218216 + ], + "credit_map": [ + 0.049580421298742294, + 0.040336765348911285, + 0.055530987679958344, + 0.04268672317266464, + 0.06767968088388443, + 0.061060115694999695, + 0.0852922797203064, + 0.05190758407115936, + 0.006601560860872269, + 0.008806613273918629, + 0.007739587686955929, + 0.006979748141020536, + 0.00662525137886405, + 0.009085673838853836, + 0.00905529409646988, + 0.010312609374523163, + 0.004419086035341024, + 0.007684648502618074, + 0.006824244745075703, + 0.003239960642531514, + 0.007246852852404118, + 0.005653357598930597, + 0.004727716092020273, + 0.004942751489579678, + 0.007452525198459625, + 0.0053809876553714275, + 0.009422804228961468, + 0.00471752742305398, + 0.01374703086912632, + 0.008204746060073376, + 0.008079967461526394, + 0.011250199750065804, + 0.01040207501500845, + 0.015426215715706348, + 0.01612231135368347, + 0.00814851839095354, + 0.017247911542654037, + 0.02121029980480671, + 0.012638342566788197, + 0.01187831163406372, + 0.007726434618234634, + 0.012992016039788723, + 0.009904094971716404, + 0.005016244016587734, + 0.007569630164653063, + 0.009135445579886436, + 0.006976453587412834, + 0.008955729193985462, + 0.0031282941345125437, + 0.004095681942999363, + 0.003807166125625372, + 0.0030158834997564554, + 0.0033637513406574726, + 0.005025631748139858, + 0.002812961582094431, + 0.004379319492727518, + 0.008737867698073387, + 0.010711753740906715, + 0.011516114696860313, + 0.004450166132301092, + 0.00444651348516345, + 0.014131861738860607, + 0.006127781234681606, + 0.012853249907493591, + 0.0041635893285274506, + 0.0034018056467175484, + 0.005421456415206194, + 0.003350248094648123, + 0.0025417543947696686, + 0.00709344120696187, + 0.0028692171908915043, + 0.006717063952237368, + 0.006426653359085321, + 0.004554038867354393, + 0.007456901948899031, + 0.004336366895586252, + 0.005457659717649221, + 0.009500900283455849, + 0.004818011075258255, + 0.005661528557538986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.0031928059179335833 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017093108035624027 + ], + "opposite_logprob_tokens": [ + -6.376709461212158 + ], + "logprob_reward": -0.0017093108035624027, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017093108035624027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21304462850093842, + 0.04309385269880295, + 0.03964412957429886, + 0.08203639090061188, + 0.11659631133079529, + 0.2128707766532898, + 0.023002544417977333, + 0.12030914425849915, + 0.089119553565979, + 0.06028270348906517 + ], + "credit_map": [ + 0.0228718351572752, + 0.015957282856106758, + 0.026440203189849854, + 0.022418461740016937, + 0.029123203828930855, + 0.023980610072612762, + 0.04479097202420235, + 0.027462054044008255, + 0.004360507242381573, + 0.005050502717494965, + 0.005291349720209837, + 0.004000426735728979, + 0.0066075981594622135, + 0.005661457777023315, + 0.00588647834956646, + 0.006235529202967882, + 0.004690088797360659, + 0.006105971056967974, + 0.005462361499667168, + 0.0031800006981939077, + 0.0046762884594500065, + 0.0055687581188976765, + 0.004564942326396704, + 0.005395715124905109, + 0.01584606245160103, + 0.010176165029406548, + 0.013850738294422626, + 0.005986085627228022, + 0.011070693843066692, + 0.011046173050999641, + 0.004964642226696014, + 0.009095819666981697, + 0.01081047486513853, + 0.014841841533780098, + 0.016511844471096992, + 0.008560444228351116, + 0.018818434327840805, + 0.019342903047800064, + 0.01401817798614502, + 0.013692180626094341, + 0.026144934818148613, + 0.04128648340702057, + 0.0287970919162035, + 0.018701739609241486, + 0.0208686925470829, + 0.040650129318237305, + 0.012596126645803452, + 0.023825544863939285, + 0.0030999924056231976, + 0.002554547041654587, + 0.0030872311908751726, + 0.002178680384531617, + 0.0020463941618800163, + 0.003910496365278959, + 0.002973945578560233, + 0.0031512558925896883, + 0.01595274545252323, + 0.017776919528841972, + 0.0178572628647089, + 0.008591868914663792, + 0.007874760776758194, + 0.027517298236489296, + 0.008259729482233524, + 0.01647854782640934, + 0.010111411102116108, + 0.008228455670177937, + 0.013699669390916824, + 0.007874533534049988, + 0.008214478380978107, + 0.017052877694368362, + 0.007030796725302935, + 0.016907330602407455, + 0.008466307073831558, + 0.006398910656571388, + 0.009111815132200718, + 0.005318829789757729, + 0.007705173920840025, + 0.010780112817883492, + 0.005733119789510965, + 0.006768428720533848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.0017093108035624027 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Siberian husky with a thick, shaggy black and white coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the husky captured in a low-angle close-up with its head thrown back high?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the husky emit a visible, thick white mist from its mouth that condenses into a cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast expanse of ice and snow with a sparse, frost-covered coniferous forest in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a low-slanting winter sun casting a light blue hue and bright contour light on the husky and mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001176975667476654 + ], + "opposite_logprob_tokens": [ + -6.751176834106445 + ], + "logprob_reward": -0.001176975667476654, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001176975667476654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19399479031562805, + 0.14301545917987823, + 0.05489033833146095, + 0.04200504347681999, + 0.11275091767311096, + 0.08368042856454849, + 0.03922351822257042, + 0.09451199322938919, + 0.07451199740171432, + 0.16141542792320251 + ], + "credit_map": [ + 0.022069765254855156, + 0.01859583705663681, + 0.023857971653342247, + 0.020282898098230362, + 0.028345417231321335, + 0.023471718654036522, + 0.03248478099703789, + 0.024886444211006165, + 0.01929340697824955, + 0.018548117950558662, + 0.017308026552200317, + 0.013249658048152924, + 0.011970496736466885, + 0.024853499606251717, + 0.016732079908251762, + 0.021060196682810783, + 0.006870411802083254, + 0.005697943270206451, + 0.008157452568411827, + 0.004826951771974564, + 0.007577976677566767, + 0.007818099111318588, + 0.0060258423909544945, + 0.007915670983493328, + 0.006154877599328756, + 0.00422388780862093, + 0.005870410706847906, + 0.00368759548291564, + 0.00792321003973484, + 0.004468503873795271, + 0.005071172956377268, + 0.004605396650731564, + 0.008635615929961205, + 0.015216529369354248, + 0.01515184249728918, + 0.009867928922176361, + 0.01965954899787903, + 0.01787421852350235, + 0.014977761544287205, + 0.0113674933090806, + 0.011958862654864788, + 0.01121450588107109, + 0.012065526098012924, + 0.0072377217002213, + 0.008470246568322182, + 0.013324201107025146, + 0.007393779698759317, + 0.012015603482723236, + 0.004699454642832279, + 0.005828327964991331, + 0.005502685438841581, + 0.003042330266907811, + 0.00375987752340734, + 0.008158702403306961, + 0.0037100170738995075, + 0.0045221298933029175, + 0.01202417816966772, + 0.015312819741666317, + 0.01495120394974947, + 0.005374493543058634, + 0.007703482639044523, + 0.01856970600783825, + 0.006762870587408543, + 0.013813259080052376, + 0.008894229307770729, + 0.00636259512975812, + 0.011687844060361385, + 0.008192941546440125, + 0.006425401195883751, + 0.01365751400589943, + 0.006410537753254175, + 0.012880947440862656, + 0.020267440006136894, + 0.015361210331320763, + 0.025384215638041496, + 0.01341789960861206, + 0.019192058593034744, + 0.03243131563067436, + 0.014883989468216896, + 0.020477332174777985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.001176975667476654 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_13/index_tensor([4248, 4248])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_13/index_tensor([4248, 4248])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..52e22ba5a0b7dac2984a0b725ee09a6ebe31e527 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_13/index_tensor([4248, 4248])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a high angle, the camera held still, focused on a sheet of fine, slightly yellowed rice paper. The paper's subtle fibers are clearly visible in the soft sidelight. A wolf-hair brush, its barrel gleaming with the natural sheen of bamboo knots, its full tip saturated with rich, jet-black ink, shimmers with a wet sheen, as if ready to drip. Guided by a steady and graceful hand (not fully captured), the tip touches the paper at a precise angle, instantly creating a tiny dot of ink. Then, with the fluidity of flowing water, the brush moves steadily and deliberately across the paper, lifting and pressing the tip downwards to create a rhythmic and powerful black line. This continuous, undulating line precisely depicts the majestic contours of the distant mountains. The light of the entire scene is soft and even, as if coming from an ancient window lattice. The main colors are the black of ink, the slightly yellow of paper, and the plain white space, creating a tranquil atmosphere of extreme concentration, full of Zen and the beauty of classical art.", + "seed": 24470, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_13/index_4248_seed_24470_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.041769815236330034, + "vlm_logprob": -0.041769815236330034, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.041769815236330034, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04964635521173477 + ], + "opposite_logprob_tokens": [ + -3.4246463775634766 + ], + "logprob_reward": -0.04964635521173477, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04964635521173477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17286905646324158, + 0.06381378322839737, + 0.08486722409725189, + 0.2212962955236435, + 0.10847105830907822, + 0.031104253605008125, + 0.02447810024023056, + 0.08706800639629364, + 0.15262870490550995, + 0.05340353772044182 + ], + "credit_map": [ + 0.021283218637108803, + 0.024041777476668358, + 0.02122609317302704, + 0.025311864912509918, + 0.017935387790203094, + 0.026556359604001045, + 0.01938965730369091, + 0.01712469384074211, + 0.008850457146763802, + 0.006911113392561674, + 0.008522996678948402, + 0.01038439478725195, + 0.007456788793206215, + 0.0070480844005942345, + 0.009181703440845013, + 0.005458245053887367, + 0.009912531822919846, + 0.007721212226897478, + 0.01764984428882599, + 0.009774083271622658, + 0.016693197190761566, + 0.008923067711293697, + 0.006214743014425039, + 0.007978550158441067, + 0.021295050159096718, + 0.02611459232866764, + 0.03562038764357567, + 0.02603750303387642, + 0.03460608050227165, + 0.02884172834455967, + 0.015770042315125465, + 0.03301091492176056, + 0.01041663158684969, + 0.010924521833658218, + 0.012496552430093288, + 0.016251005232334137, + 0.011690964922308922, + 0.017848648130893707, + 0.01408544834703207, + 0.014757283963263035, + 0.004916455131024122, + 0.006153831258416176, + 0.0048128766939044, + 0.004702083300799131, + 0.0033788636792451143, + 0.002210358390584588, + 0.002406069543212652, + 0.0025237128138542175, + 0.004175656009465456, + 0.0049486225470900536, + 0.0026445938274264336, + 0.002509105484932661, + 0.002449367893859744, + 0.0031858906149864197, + 0.001965498086065054, + 0.002599368803203106, + 0.010471337474882603, + 0.010428540408611298, + 0.017524665221571922, + 0.010723522864282131, + 0.009922508150339127, + 0.010332156904041767, + 0.010499252006411552, + 0.007166014984250069, + 0.02828819863498211, + 0.03480188548564911, + 0.027026290073990822, + 0.0164240300655365, + 0.012985833920538425, + 0.013592395931482315, + 0.009973603300750256, + 0.009536470286548138, + 0.004915005061775446, + 0.004994564224034548, + 0.009510436095297337, + 0.006256203167140484, + 0.011035480536520481, + 0.005467382725328207, + 0.005947466939687729, + 0.005277000833302736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.04964635521173477 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023339200764894485 + ], + "opposite_logprob_tokens": [ + -3.77333927154541 + ], + "logprob_reward": -0.023339200764894485, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023339200764894485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3858884572982788, + 0.06080654636025429, + 0.10766322910785675, + 0.050725702196359634, + 0.06906656920909882, + 0.06078186631202698, + 0.06344342976808548, + 0.021662944927811623, + 0.046926066279411316, + 0.13303519785404205 + ], + "credit_map": [ + 0.042031556367874146, + 0.055018212646245956, + 0.05314305052161217, + 0.04467204958200455, + 0.05390390753746033, + 0.06413934379816055, + 0.034072596579790115, + 0.038907740265131, + 0.007079069036990404, + 0.00596203887835145, + 0.009484563954174519, + 0.007091342471539974, + 0.0059873019345104694, + 0.013653719797730446, + 0.005621292162686586, + 0.0059272171929478645, + 0.014719607308506966, + 0.015702348202466965, + 0.011484778486192226, + 0.011714374646544456, + 0.015285950154066086, + 0.01653825119137764, + 0.009898586198687553, + 0.012319335713982582, + 0.0051225884817540646, + 0.006975158117711544, + 0.006529143080115318, + 0.005663753021508455, + 0.00694319186732173, + 0.007714540231972933, + 0.003866360289976001, + 0.007910965010523796, + 0.009058788418769836, + 0.007750802673399448, + 0.009946138598024845, + 0.010009042918682098, + 0.007054136134684086, + 0.009391664527356625, + 0.007735033053904772, + 0.008120964281260967, + 0.0072227767668664455, + 0.008817763067781925, + 0.00873546488583088, + 0.011547342874109745, + 0.0074428534135222435, + 0.005637770518660545, + 0.00586173590272665, + 0.0055161574855446815, + 0.008654198609292507, + 0.006789869163185358, + 0.010282674804329872, + 0.00872963760048151, + 0.007313769776374102, + 0.00964932981878519, + 0.007545126136392355, + 0.00447881780564785, + 0.002485345583409071, + 0.003041577758267522, + 0.0025212520267814398, + 0.0031870342791080475, + 0.002568519674241543, + 0.0031456141732633114, + 0.0024162691552191973, + 0.0022973325103521347, + 0.006476808805018663, + 0.0064934310503304005, + 0.00713950814679265, + 0.006952101364731789, + 0.004665776621550322, + 0.005401029717177153, + 0.006155526265501976, + 0.0036418817471712828, + 0.010739960707724094, + 0.01464039832353592, + 0.01813552901148796, + 0.02003556303679943, + 0.01843317411839962, + 0.013262128457427025, + 0.020619485527276993, + 0.017168959602713585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.023339200764894485 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649622090160847 + ], + "opposite_logprob_tokens": [ + -4.758649826049805 + ], + "logprob_reward": -0.008649622090160847, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649622090160847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3563399612903595, + 0.03914102539420128, + 0.10512425005435944, + 0.09855757653713226, + 0.15377290546894073, + 0.036877721548080444, + 0.06366755068302155, + 0.02660074084997177, + 0.0385810062289238, + 0.08133729547262192 + ], + "credit_map": [ + 0.0447782427072525, + 0.05077105760574341, + 0.044204894453287125, + 0.038213860243558884, + 0.04770476371049881, + 0.06752871721982956, + 0.03060409612953663, + 0.03253432363271713, + 0.004880511201918125, + 0.004195350222289562, + 0.00485098036006093, + 0.004820110742002726, + 0.004325068090111017, + 0.0063067651353776455, + 0.00621445057913661, + 0.0035477904602885246, + 0.01332212332636118, + 0.014908114448189735, + 0.012483544647693634, + 0.009297281503677368, + 0.020519470795989037, + 0.01417736429721117, + 0.00919424556195736, + 0.011222107335925102, + 0.009091710671782494, + 0.010786982253193855, + 0.01999172382056713, + 0.01149340346455574, + 0.016194619238376617, + 0.012340213172137737, + 0.007145929150283337, + 0.011512997560203075, + 0.017613425850868225, + 0.017773248255252838, + 0.019307833164930344, + 0.02003040909767151, + 0.017828600481152534, + 0.022716831415891647, + 0.020375488325953484, + 0.018127072602510452, + 0.003571019973605871, + 0.003289362881332636, + 0.00861908309161663, + 0.004171925596892834, + 0.007225465029478073, + 0.0044137658551335335, + 0.0030135929118841887, + 0.0025735069066286087, + 0.009045311249792576, + 0.006671056617051363, + 0.010504716075956821, + 0.008577687665820122, + 0.007866074331104755, + 0.009758306667208672, + 0.0067498101852834225, + 0.004494586493819952, + 0.0031252619810402393, + 0.0035738369915634394, + 0.003174192039296031, + 0.0037866306956857443, + 0.003020673291757703, + 0.004777467343956232, + 0.0028096064925193787, + 0.00233307178132236, + 0.005239602643996477, + 0.006752621848136187, + 0.007793625351041555, + 0.003917086869478226, + 0.00422306451946497, + 0.005208482034504414, + 0.002833473961800337, + 0.00261304690502584, + 0.007942537777125835, + 0.009244031272828579, + 0.011010480113327503, + 0.012976001016795635, + 0.009772009216248989, + 0.00962877832353115, + 0.010365440510213375, + 0.010398016311228275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.008649622090160847 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420446764677763 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.01420446764677763, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420446764677763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18863733112812042, + 0.060161445289850235, + 0.0545223131775856, + 0.15394049882888794, + 0.1724092662334442, + 0.03779875859618187, + 0.011023661121726036, + 0.04037204012274742, + 0.06625211983919144, + 0.21488259732723236 + ], + "credit_map": [ + 0.026871245354413986, + 0.024455731734633446, + 0.025749972090125084, + 0.02136710286140442, + 0.020720887929201126, + 0.03600716590881348, + 0.0173637792468071, + 0.016101427376270294, + 0.005845773033797741, + 0.006517669185996056, + 0.00915113277733326, + 0.008089193142950535, + 0.005675255320966244, + 0.010678659193217754, + 0.007900554686784744, + 0.006303201895207167, + 0.0073715802282094955, + 0.008167427033185959, + 0.0069488584995269775, + 0.006095764227211475, + 0.004902205895632505, + 0.008178606629371643, + 0.006226974073797464, + 0.006630892399698496, + 0.01462475024163723, + 0.01534742210060358, + 0.030355118215084076, + 0.01748618856072426, + 0.029723625630140305, + 0.0168361347168684, + 0.012253834865987301, + 0.01731341890990734, + 0.016064956784248352, + 0.01846424490213394, + 0.02703450806438923, + 0.018550187349319458, + 0.025602703914046288, + 0.02562112547457218, + 0.021606024354696274, + 0.019465506076812744, + 0.004496391396969557, + 0.006373582407832146, + 0.005266423802822828, + 0.004806692246347666, + 0.0049963747151196, + 0.005122894886881113, + 0.0033120682928711176, + 0.003424327587708831, + 0.0014410997973755002, + 0.001461379579268396, + 0.001843913458287716, + 0.0011532598873600364, + 0.0018314720364287496, + 0.0016768513014540076, + 0.0008048832532949746, + 0.0008108008769340813, + 0.003960342612117529, + 0.005682929418981075, + 0.007336786948144436, + 0.005087010562419891, + 0.004050901159644127, + 0.006553145125508308, + 0.00430690124630928, + 0.0033940200228244066, + 0.01069829985499382, + 0.013067713938653469, + 0.0115183861926198, + 0.007522592321038246, + 0.00636214530095458, + 0.007289277855306864, + 0.00503845838829875, + 0.0047552394680678844, + 0.01826273836195469, + 0.021819816902279854, + 0.028223278000950813, + 0.03401828929781914, + 0.023667508736252785, + 0.026630986481904984, + 0.032104168087244034, + 0.030155787244439125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.01420446764677763 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300943046808243 + ], + "opposite_logprob_tokens": [ + -2.238009452819824 + ], + "logprob_reward": -0.11300943046808243, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300943046808243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4919799268245697, + 0.11059869825839996, + 0.04183392599225044, + 0.03929201140999794, + 0.06404159963130951, + 0.03958119451999664, + 0.05742418393492699, + 0.054861441254615784, + 0.0417490154504776, + 0.05863797292113304 + ], + "credit_map": [ + 0.06342015415430069, + 0.0737898200750351, + 0.08862364292144775, + 0.04690653085708618, + 0.060550786554813385, + 0.07850012928247452, + 0.04583139717578888, + 0.03435749560594559, + 0.015220477245748043, + 0.013853799551725388, + 0.018479086458683014, + 0.014430277980864048, + 0.011871034279465675, + 0.01351773552596569, + 0.01401556096971035, + 0.00921072531491518, + 0.00423137191683054, + 0.004191221669316292, + 0.010286612436175346, + 0.003668364603072405, + 0.008820085786283016, + 0.004445601254701614, + 0.0028187474235892296, + 0.0033719190396368504, + 0.005633099935948849, + 0.006822157651185989, + 0.006445348262786865, + 0.0037159218918532133, + 0.005732033867388964, + 0.004496845882385969, + 0.0031600596848875284, + 0.0032865440007299185, + 0.008161209523677826, + 0.005272528622299433, + 0.012845834717154503, + 0.006544563919305801, + 0.008381122723221779, + 0.012861577793955803, + 0.004673006944358349, + 0.005301754456013441, + 0.004832176025956869, + 0.006279657129198313, + 0.006051809061318636, + 0.006656592711806297, + 0.004449756350368261, + 0.003586558159440756, + 0.0037423388566821814, + 0.003982308320701122, + 0.007770531810820103, + 0.00704277865588665, + 0.00964320357888937, + 0.009029178880155087, + 0.0071030971594154835, + 0.0069432067684829235, + 0.00614825589582324, + 0.003743934677913785, + 0.0079762889072299, + 0.007217614445835352, + 0.012875312939286232, + 0.006762062199413776, + 0.005980783142149448, + 0.005436323583126068, + 0.00471479631960392, + 0.003898261347785592, + 0.007716371677815914, + 0.007101761177182198, + 0.005972892977297306, + 0.00678485631942749, + 0.003781188279390335, + 0.004340392537415028, + 0.003244308289140463, + 0.002807243261486292, + 0.005372915416955948, + 0.005908191669732332, + 0.01046009175479412, + 0.008442889899015427, + 0.0100864227861166, + 0.006013755686581135, + 0.006273769307881594, + 0.006079936400055885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.11300943046808243 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a high angle, the camera held still, focused on a sheet of fine, slightly yellowed rice paper. The paper's subtle fibers are clearly visible in the soft sidelight. A wolf-hair brush, its barrel gleaming with the natural sheen of bamboo knots, its full tip saturated with rich, jet-black ink, shimmers with a wet sheen, as if ready to drip. Guided by a steady and graceful hand (not fully captured), the tip touches the paper at a precise angle, instantly creating a tiny dot of ink. Then, with the fluidity of flowing water, the brush moves steadily and deliberately across the paper, lifting and pressing the tip downwards to create a rhythmic and powerful black line. This continuous, undulating line precisely depicts the majestic contours of the distant mountains. The light of the entire scene is soft and even, as if coming from an ancient window lattice. The main colors are the black of ink, the slightly yellow of paper, and the plain white space, creating a tranquil atmosphere of extreme concentration, full of Zen and the beauty of classical art.", + "seed": 24471, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_13/index_4248_seed_24471_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02918442147783935, + "vlm_logprob": -0.02918442147783935, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.02918442147783935, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.049672674387693405 + ], + "opposite_logprob_tokens": [ + -3.4246726036071777 + ], + "logprob_reward": -0.049672674387693405, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.049672674387693405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1063147559762001, + 0.15382599830627441, + 0.10870146751403809, + 0.07865025848150253, + 0.12005271762609482, + 0.16295289993286133, + 0.056309353560209274, + 0.03197844326496124, + 0.09935405850410461, + 0.0818600133061409 + ], + "credit_map": [ + 0.00981299951672554, + 0.011629682034254074, + 0.013646059669554234, + 0.013913990929722786, + 0.02295808680355549, + 0.012615328654646873, + 0.01076025702059269, + 0.01097834948450327, + 0.013771742582321167, + 0.010927216149866581, + 0.01434326358139515, + 0.017133299261331558, + 0.018833067268133163, + 0.035553157329559326, + 0.021941352635622025, + 0.02132290229201317, + 0.006427019834518433, + 0.01150768343359232, + 0.009725676849484444, + 0.011766375042498112, + 0.01800493709743023, + 0.02710581198334694, + 0.009794041514396667, + 0.014369922690093517, + 0.004636439960449934, + 0.014907664619386196, + 0.006473811808973551, + 0.0055426619946956635, + 0.02667761780321598, + 0.006183142773807049, + 0.003983086906373501, + 0.010245838202536106, + 0.01331856194883585, + 0.017940353602170944, + 0.010368812829256058, + 0.01771640032529831, + 0.021164635196328163, + 0.018864816054701805, + 0.008737878873944283, + 0.011941276490688324, + 0.014471612870693207, + 0.030080394819378853, + 0.013684920035302639, + 0.01901058480143547, + 0.0400308333337307, + 0.0204994585365057, + 0.010434827767312527, + 0.014740274287760258, + 0.00689056096598506, + 0.006829200778156519, + 0.00822120439261198, + 0.008851258084177971, + 0.008128565736114979, + 0.006930328905582428, + 0.005407800432294607, + 0.005050435662269592, + 0.0023965921718627214, + 0.005408278666436672, + 0.0030824122950434685, + 0.003957404755055904, + 0.007608451414853334, + 0.0027371232863515615, + 0.0031404613982886076, + 0.003647720208391547, + 0.01057301927357912, + 0.013662614859640598, + 0.015247460454702377, + 0.014160982333123684, + 0.01303967647254467, + 0.01048742700368166, + 0.011630932800471783, + 0.01055195089429617, + 0.004677848424762487, + 0.013239007443189621, + 0.009344223886728287, + 0.01153555791825056, + 0.017229346558451653, + 0.009603843092918396, + 0.0067149922251701355, + 0.009515209123492241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.049672674387693405 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009807259775698185 + ], + "opposite_logprob_tokens": [ + -4.634807109832764 + ], + "logprob_reward": -0.009807259775698185, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009807259775698185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07404406368732452, + 0.0793074443936348, + 0.14366132020950317, + 0.24005432426929474, + 0.08902991563081741, + 0.05155669525265694, + 0.045405514538288116, + 0.05480623617768288, + 0.0718931034207344, + 0.1502414494752884 + ], + "credit_map": [ + 0.007675324101001024, + 0.007723812013864517, + 0.008578264154493809, + 0.008270518854260445, + 0.011119416914880276, + 0.012064656242728233, + 0.00785293709486723, + 0.010759134776890278, + 0.010500723496079445, + 0.007818961516022682, + 0.012532508932054043, + 0.009681441821157932, + 0.009495752863585949, + 0.010395318269729614, + 0.00960446149110794, + 0.009278275072574615, + 0.007597050629556179, + 0.01622241549193859, + 0.012003572657704353, + 0.01212386041879654, + 0.027472760528326035, + 0.03705212473869324, + 0.011247752234339714, + 0.01994178257882595, + 0.018517037853598595, + 0.027007484808564186, + 0.03767680749297142, + 0.04018087312579155, + 0.026105979457497597, + 0.03107266314327717, + 0.030460530892014503, + 0.02903294749557972, + 0.008234065026044846, + 0.009905263781547546, + 0.007433626335114241, + 0.01333682145923376, + 0.014512005262076855, + 0.01544725801795721, + 0.009269271977245808, + 0.010891602374613285, + 0.005317907780408859, + 0.007985621690750122, + 0.008846264332532883, + 0.006001978646963835, + 0.009261959232389927, + 0.00587368244305253, + 0.004034338518977165, + 0.004234943073242903, + 0.0046495189890265465, + 0.008578280918300152, + 0.0046087391674518585, + 0.0044742911122739315, + 0.01236218586564064, + 0.0037941820919513702, + 0.0038534975610673428, + 0.0030848169699311256, + 0.007322449702769518, + 0.007961265742778778, + 0.005248727276921272, + 0.0068109710700809956, + 0.012850621715188026, + 0.004508331883698702, + 0.005301249213516712, + 0.004802617710083723, + 0.009142129682004452, + 0.008891062811017036, + 0.006596927531063557, + 0.008344948291778564, + 0.014769962057471275, + 0.007032732479274273, + 0.009089939296245575, + 0.008025401271879673, + 0.0081331767141819, + 0.03817892447113991, + 0.009028050117194653, + 0.012848425656557083, + 0.05376233160495758, + 0.013180645182728767, + 0.007032491732388735, + 0.00807739607989788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.009807259775698185 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028208012226969004 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028208012226969004, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028208012226969004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10394234210252762, + 0.08438277244567871, + 0.11417614668607712, + 0.09785927087068558, + 0.15935680270195007, + 0.0720202699303627, + 0.041042353957891464, + 0.09630421549081802, + 0.05656278505921364, + 0.17435303330421448 + ], + "credit_map": [ + 0.008611216209828854, + 0.009898476302623749, + 0.014433316886425018, + 0.013510541059076786, + 0.015742529183626175, + 0.014861930161714554, + 0.012870107777416706, + 0.01401421520859003, + 0.007065759506076574, + 0.008563480339944363, + 0.008504101075232029, + 0.009864900261163712, + 0.009521490894258022, + 0.01819889433681965, + 0.009487507864832878, + 0.013176633976399899, + 0.007552810478955507, + 0.011276771314442158, + 0.010277791880071163, + 0.01357224304229021, + 0.015474656596779823, + 0.02974764071404934, + 0.01081205252557993, + 0.015462169423699379, + 0.006708381697535515, + 0.01021441537886858, + 0.012250707484781742, + 0.015134312212467194, + 0.012953685596585274, + 0.017459893599152565, + 0.01083179097622633, + 0.012306082993745804, + 0.014867017976939678, + 0.01971163973212242, + 0.013035092502832413, + 0.02257184311747551, + 0.02698361873626709, + 0.025060245767235756, + 0.01596239022910595, + 0.02116495557129383, + 0.008435549214482307, + 0.011805996298789978, + 0.010993902571499348, + 0.007633509114384651, + 0.014639043249189854, + 0.008120376616716385, + 0.005119622685015202, + 0.005272264592349529, + 0.003727397182956338, + 0.007288780529052019, + 0.0038810819387435913, + 0.007741064764559269, + 0.0068175308406353, + 0.004260482732206583, + 0.0032029009889811277, + 0.004123109392821789, + 0.007427693344652653, + 0.018718821927905083, + 0.007228800095617771, + 0.013035273179411888, + 0.024897415190935135, + 0.009053078480064869, + 0.010342145338654518, + 0.005600978620350361, + 0.007633529603481293, + 0.007387272547930479, + 0.004871715791523457, + 0.006796874105930328, + 0.010686159133911133, + 0.005481650121510029, + 0.007827398367226124, + 0.005878182593733072, + 0.008559300564229488, + 0.04228462278842926, + 0.013447780162096024, + 0.018484322354197502, + 0.05713574215769768, + 0.014901071786880493, + 0.009820162318646908, + 0.009720017202198505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.0028208012226969004 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463898153975606 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.00463898153975606, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463898153975606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07629254460334778, + 0.04529865086078644, + 0.04533287137746811, + 0.15166866779327393, + 0.07402973622083664, + 0.07907084375619888, + 0.09570877254009247, + 0.12518861889839172, + 0.20999716222286224, + 0.09741213172674179 + ], + "credit_map": [ + 0.00957665964961052, + 0.010434018447995186, + 0.007792633958160877, + 0.009081774391233921, + 0.012174593284726143, + 0.009650206193327904, + 0.008700349368155003, + 0.008882309310138226, + 0.0030448660254478455, + 0.0048393732868134975, + 0.0061727650463581085, + 0.007259116508066654, + 0.005497233010828495, + 0.005945256445556879, + 0.006483545061200857, + 0.006056495476514101, + 0.004033451899886131, + 0.004405034705996513, + 0.004501738119870424, + 0.00512954406440258, + 0.007909846492111683, + 0.008519184775650501, + 0.004349189810454845, + 0.006484881974756718, + 0.008946969173848629, + 0.027034912258386612, + 0.01485501229763031, + 0.01445971429347992, + 0.04242366552352905, + 0.015096635557711124, + 0.01093209721148014, + 0.017919661477208138, + 0.006520745810121298, + 0.0074630603194236755, + 0.00827872660011053, + 0.012207302264869213, + 0.00869433581829071, + 0.015150078572332859, + 0.006298742722719908, + 0.009416749700903893, + 0.009387211874127388, + 0.00963503960520029, + 0.015548723749816418, + 0.009448624216020107, + 0.008095934987068176, + 0.010789256542921066, + 0.007744107395410538, + 0.008421942591667175, + 0.007464400492608547, + 0.02320491522550583, + 0.0071609425358474255, + 0.009665142744779587, + 0.026390476152300835, + 0.008988378569483757, + 0.005104957614094019, + 0.007729565259069204, + 0.009553779847919941, + 0.029103288426995277, + 0.007836034521460533, + 0.012399330735206604, + 0.03844483569264412, + 0.011650050058960915, + 0.009308257140219212, + 0.006893046200275421, + 0.02090832032263279, + 0.036965999752283096, + 0.01640787161886692, + 0.023555627092719078, + 0.05444345623254776, + 0.016551483422517776, + 0.024686196818947792, + 0.016478218138217926, + 0.008357876911759377, + 0.011591962538659573, + 0.014290916733443737, + 0.015137852169573307, + 0.010114150121808052, + 0.011305668391287327, + 0.011844167485833168, + 0.014769529923796654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.00463898153975606 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789823904633522 + ], + "opposite_logprob_tokens": [ + -2.578982353210449 + ], + "logprob_reward": -0.0789823904633522, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789823904633522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07917178422212601, + 0.051521968096494675, + 0.10893645882606506, + 0.3700745403766632, + 0.0679590106010437, + 0.10220056027173996, + 0.11684408038854599, + 0.036724794656038284, + 0.016115237027406693, + 0.05045149847865105 + ], + "credit_map": [ + 0.008849229663610458, + 0.008785039186477661, + 0.012214467860758305, + 0.008497027680277824, + 0.012517877854406834, + 0.009409409016370773, + 0.00986488163471222, + 0.009033869951963425, + 0.0047565666027367115, + 0.0043480414897203445, + 0.006615169811993837, + 0.006331474520266056, + 0.005166217219084501, + 0.008671954274177551, + 0.008417103439569473, + 0.007215446326881647, + 0.00790493655949831, + 0.011721193790435791, + 0.010214943438768387, + 0.009610074572265148, + 0.01842495985329151, + 0.0279864314943552, + 0.008824784308671951, + 0.014249145053327084, + 0.017795853316783905, + 0.10560774058103561, + 0.012437354773283005, + 0.015347511507570744, + 0.1509002298116684, + 0.018071716651320457, + 0.009204195812344551, + 0.04070999100804329, + 0.0055809011682868, + 0.006802883464843035, + 0.006825177930295467, + 0.010553887113928795, + 0.006925742607563734, + 0.014798855409026146, + 0.0073383948765695095, + 0.00913318246603012, + 0.015439596027135849, + 0.011623530648648739, + 0.015950147062540054, + 0.015062570571899414, + 0.012874896638095379, + 0.01264062337577343, + 0.008515656925737858, + 0.010093548335134983, + 0.010865322314202785, + 0.030117075890302658, + 0.010887011885643005, + 0.007060543168336153, + 0.038520947098731995, + 0.009127975441515446, + 0.0047443280927836895, + 0.005520904436707497, + 0.0049686674028635025, + 0.005428939126431942, + 0.006184636615216732, + 0.0044407714158296585, + 0.004333190154284239, + 0.004772563464939594, + 0.0036939806304872036, + 0.002902054460719228, + 0.0017259439919143915, + 0.0023426865227520466, + 0.0026533787604421377, + 0.0022824076004326344, + 0.001834533060900867, + 0.0018493585521355271, + 0.0019861061591655016, + 0.0014408254064619541, + 0.002835269086062908, + 0.013662964105606079, + 0.003027289407327771, + 0.0024270019493997097, + 0.01842140406370163, + 0.004760321695357561, + 0.0019078452605754137, + 0.0034094129223376513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.0789823904633522 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a high angle, the camera held still, focused on a sheet of fine, slightly yellowed rice paper. The paper's subtle fibers are clearly visible in the soft sidelight. A wolf-hair brush, its barrel gleaming with the natural sheen of bamboo knots, its full tip saturated with rich, jet-black ink, shimmers with a wet sheen, as if ready to drip. Guided by a steady and graceful hand (not fully captured), the tip touches the paper at a precise angle, instantly creating a tiny dot of ink. Then, with the fluidity of flowing water, the brush moves steadily and deliberately across the paper, lifting and pressing the tip downwards to create a rhythmic and powerful black line. This continuous, undulating line precisely depicts the majestic contours of the distant mountains. The light of the entire scene is soft and even, as if coming from an ancient window lattice. The main colors are the black of ink, the slightly yellow of paper, and the plain white space, creating a tranquil atmosphere of extreme concentration, full of Zen and the beauty of classical art.", + "seed": 24472, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_13/index_4248_seed_24472_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020562657294794916, + "vlm_logprob": -0.020562657294794916, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.020562657294794916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022823037579655647 + ], + "opposite_logprob_tokens": [ + -5.022822856903076 + ], + "logprob_reward": -0.022823037579655647, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.022823037579655647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5758586525917053, + 0.03170439228415489, + 0.03932274132966995, + 0.027899933978915215, + 0.020810279995203018, + 0.036475177854299545, + 0.049060530960559845, + 0.032332487404346466, + 0.1620834767818451, + 0.02445238083600998 + ], + "credit_map": [ + 0.0441867858171463, + 0.20062197744846344, + 0.047844722867012024, + 0.027530288323760033, + 0.030438773334026337, + 0.029928145930171013, + 0.026468677446246147, + 0.1688392013311386, + 0.0037800457794219255, + 0.004182499833405018, + 0.0045544481836259365, + 0.0025581801310181618, + 0.0038352024275809526, + 0.0068975817412137985, + 0.002813731087371707, + 0.0030827035661786795, + 0.004710722714662552, + 0.0050420016050338745, + 0.0067194243893027306, + 0.002496554749086499, + 0.005973873194307089, + 0.007177851628512144, + 0.002815484069287777, + 0.004386828280985355, + 0.004728391300886869, + 0.0034904645290225744, + 0.005003799684345722, + 0.0015778322704136372, + 0.0038538931403309107, + 0.0037208236753940582, + 0.0016160523518919945, + 0.003908677026629448, + 0.0035565400030463934, + 0.002444242127239704, + 0.003788976464420557, + 0.0016908594407141209, + 0.004157878924161196, + 0.0025937221944332123, + 0.0010800627060234547, + 0.001497998251579702, + 0.005287908483296633, + 0.0051751756109297276, + 0.010078496299684048, + 0.003269309177994728, + 0.004390816669911146, + 0.0029847407713532448, + 0.002133741043508053, + 0.0031549902632832527, + 0.009087680839002132, + 0.006182547193020582, + 0.011716749519109726, + 0.005802697502076626, + 0.007196095306426287, + 0.004280636552721262, + 0.0020636783447116613, + 0.002730443375185132, + 0.007452536839991808, + 0.0040737115778028965, + 0.005422936752438545, + 0.0030968941282480955, + 0.00507820351049304, + 0.0025892944540828466, + 0.002196428133174777, + 0.002422480843961239, + 0.03922717645764351, + 0.021702436730265617, + 0.028432976454496384, + 0.01248494628816843, + 0.02480972558259964, + 0.01553850807249546, + 0.010183900594711304, + 0.009703797288239002, + 0.005438565742224455, + 0.003469057846814394, + 0.004528956953436136, + 0.003121328307315707, + 0.0027017465326935053, + 0.0020768840331584215, + 0.0017592946533113718, + 0.0013565471163019538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.022823037579655647 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036361785605549812 + ], + "opposite_logprob_tokens": [ + -5.628636360168457 + ], + "logprob_reward": -0.0036361785605549812, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036361785605549812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15029147267341614, + 0.09259835630655289, + 0.09409346431493759, + 0.0728767067193985, + 0.04821616783738136, + 0.09655144810676575, + 0.1205073669552803, + 0.056612905114889145, + 0.21141064167022705, + 0.056841373443603516 + ], + "credit_map": [ + 0.021026136353611946, + 0.01946316473186016, + 0.022627707570791245, + 0.01447594165802002, + 0.019370628520846367, + 0.021584272384643555, + 0.014381707645952702, + 0.017361940816044807, + 0.01705310493707657, + 0.010776435025036335, + 0.013468651100993156, + 0.00786627922207117, + 0.012451195158064365, + 0.014489402063190937, + 0.0068077086471021175, + 0.009685594588518143, + 0.016603121533989906, + 0.010278495028614998, + 0.01328742690384388, + 0.008719313889741898, + 0.014839235693216324, + 0.015333171933889389, + 0.005209672264754772, + 0.009823036380112171, + 0.014233163557946682, + 0.0068758889101445675, + 0.010829216800630093, + 0.008479423820972443, + 0.015190912410616875, + 0.008333600126206875, + 0.0036351210437715054, + 0.005299384240061045, + 0.009822309948503971, + 0.0058426144532859325, + 0.007594890892505646, + 0.00558155681937933, + 0.008897978812456131, + 0.004845304414629936, + 0.0020926508586853743, + 0.0035388614051043987, + 0.015296646393835545, + 0.012299900874495506, + 0.02266336977481842, + 0.011506629176437855, + 0.015231860801577568, + 0.00809494499117136, + 0.004440879914909601, + 0.007017225027084351, + 0.01838093064725399, + 0.015735553577542305, + 0.0246023740619421, + 0.010226601734757423, + 0.017193013802170753, + 0.01724638044834137, + 0.00815755594521761, + 0.008964970707893372, + 0.008831904269754887, + 0.007980028167366982, + 0.011364861391484737, + 0.00514282938092947, + 0.007895183749496937, + 0.003825208405032754, + 0.006604181136935949, + 0.004968713968992233, + 0.048121675848960876, + 0.02838090993463993, + 0.037960346788167953, + 0.015478791669011116, + 0.030609434470534325, + 0.02278180979192257, + 0.015394928865134716, + 0.012682747095823288, + 0.010783752426505089, + 0.005922163370996714, + 0.007231302093714476, + 0.007483947556465864, + 0.01001522783190012, + 0.006880070548504591, + 0.0043983845971524715, + 0.004126529674977064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.0036361785605549812 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003199935657903552 + ], + "opposite_logprob_tokens": [ + -5.753200054168701 + ], + "logprob_reward": -0.003199935657903552, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003199935657903552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.515145480632782, + 0.06221610680222511, + 0.0669369325041771, + 0.05496890842914581, + 0.04730378836393356, + 0.037943556904792786, + 0.11434083431959152, + 0.024334505200386047, + 0.051815345883369446, + 0.024994555860757828 + ], + "credit_map": [ + 0.031630344688892365, + 0.20681039988994598, + 0.01882842928171158, + 0.012986190617084503, + 0.024366507306694984, + 0.02293090894818306, + 0.015829462558031082, + 0.18176323175430298, + 0.010540513321757317, + 0.00699304323643446, + 0.008758354932069778, + 0.004993597976863384, + 0.008990620262920856, + 0.00774597842246294, + 0.005863569211214781, + 0.008330429904162884, + 0.009541241452097893, + 0.008929361589252949, + 0.0129463579505682, + 0.0064343782141804695, + 0.008173090405762196, + 0.009515214711427689, + 0.00455515505746007, + 0.006842128466814756, + 0.010137242265045643, + 0.006884366739541292, + 0.011120694689452648, + 0.003917130175977945, + 0.008451223373413086, + 0.005313873756676912, + 0.0028304271399974823, + 0.006313952151685953, + 0.007199926767498255, + 0.0064756982028484344, + 0.008320493623614311, + 0.0045710355043411255, + 0.006000596098601818, + 0.00933176837861538, + 0.0020876291673630476, + 0.003316640853881836, + 0.004960579331964254, + 0.0048501393757760525, + 0.009076051414012909, + 0.004062078893184662, + 0.005204067099839449, + 0.00343743315897882, + 0.0025930937845259905, + 0.003760112216696143, + 0.015507198870182037, + 0.014969117008149624, + 0.024359697476029396, + 0.01107074599713087, + 0.015698572620749474, + 0.014974924735724926, + 0.008787189610302448, + 0.008973386138677597, + 0.003605647711083293, + 0.0031288762111216784, + 0.0037247431464493275, + 0.0024676602333784103, + 0.0034665255807340145, + 0.0020758272148668766, + 0.0030131277162581682, + 0.0028520964551717043, + 0.009094709530472755, + 0.007050908636301756, + 0.010595222003757954, + 0.005520169623196125, + 0.006625373847782612, + 0.005256077740341425, + 0.0037818914279341698, + 0.00389099377207458, + 0.0050768484361469746, + 0.0033213431015610695, + 0.004616031888872385, + 0.002359732985496521, + 0.004119295161217451, + 0.0026909448206424713, + 0.001281591597944498, + 0.0015287675196304917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.003199935657903552 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190698698163033 + ], + "opposite_logprob_tokens": [ + -4.018190860748291 + ], + "logprob_reward": -0.018190698698163033, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018190698698163033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5293872952461243, + 0.031782057136297226, + 0.03491637855768204, + 0.024143997579813004, + 0.0500275082886219, + 0.03277267515659332, + 0.10481827706098557, + 0.018544398248195648, + 0.15764634311199188, + 0.015961043536663055 + ], + "credit_map": [ + 0.03262827917933464, + 0.21258080005645752, + 0.020543940365314484, + 0.013624439015984535, + 0.02344251610338688, + 0.02388094924390316, + 0.0158856064081192, + 0.1868007332086563, + 0.004793090745806694, + 0.0035046155098825693, + 0.006103586405515671, + 0.002936416072770953, + 0.004773410502821207, + 0.004294386133551598, + 0.0026969758328050375, + 0.002679575001820922, + 0.00535427825525403, + 0.0044874157756567, + 0.005928001366555691, + 0.0025874613784253597, + 0.0046487972140312195, + 0.006738258525729179, + 0.0020935514476150274, + 0.003078613430261612, + 0.0031576629262417555, + 0.0031651207245886326, + 0.00448705954477191, + 0.0022611955646425486, + 0.004046064801514149, + 0.002486069453880191, + 0.00198856252245605, + 0.0025522620417177677, + 0.00698579428717494, + 0.005656959023326635, + 0.007549004629254341, + 0.0032888511195778847, + 0.009865772910416126, + 0.009591000154614449, + 0.0033229293767362833, + 0.0037671979516744614, + 0.004334716126322746, + 0.004141488578170538, + 0.004842284135520458, + 0.004892055876553059, + 0.0038356115110218525, + 0.0031376760452985764, + 0.004092332907021046, + 0.003496508114039898, + 0.015350468456745148, + 0.01326071098446846, + 0.021602436900138855, + 0.00943767186254263, + 0.013989260420203209, + 0.016711169853806496, + 0.006392157636582851, + 0.008074403740465641, + 0.0036675273440778255, + 0.002500097965821624, + 0.003445548005402088, + 0.0015529750380665064, + 0.003445519832894206, + 0.0010056364117190242, + 0.0015053601237013936, + 0.0014217337593436241, + 0.03234805539250374, + 0.020281968638300896, + 0.02776503935456276, + 0.011840561404824257, + 0.022702952846884727, + 0.018840063363313675, + 0.013218286447227001, + 0.010649423114955425, + 0.003061724593862891, + 0.0018043848685920238, + 0.0026034237816929817, + 0.002723261946812272, + 0.002608619397506118, + 0.001378623186610639, + 0.0008824296528473496, + 0.0008985765744000673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.018190698698163033 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496343597769737 + ], + "opposite_logprob_tokens": [ + -2.9299633502960205 + ], + "logprob_reward": -0.05496343597769737, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496343597769737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6881155967712402, + 0.03346846252679825, + 0.0331064835190773, + 0.01995694264769554, + 0.038723479956388474, + 0.022035356611013412, + 0.024761728942394257, + 0.023083427920937538, + 0.07688825577497482, + 0.039860326796770096 + ], + "credit_map": [ + 0.04369959607720375, + 0.2765701413154602, + 0.03310975804924965, + 0.018402379006147385, + 0.028407644480466843, + 0.026577770709991455, + 0.01932055689394474, + 0.24202771484851837, + 0.0053214430809021, + 0.004597197752445936, + 0.004829815123230219, + 0.004156135953962803, + 0.0036072968505322933, + 0.003943750634789467, + 0.0031241783872246742, + 0.0038886452093720436, + 0.005283412989228964, + 0.003986692056059837, + 0.00532984035089612, + 0.0029656188562512398, + 0.004585433751344681, + 0.005245028994977474, + 0.0022961576469242573, + 0.003414296545088291, + 0.002982393139973283, + 0.00269244029186666, + 0.0037222993560135365, + 0.001355079817585647, + 0.003470686497166753, + 0.0028919121250510216, + 0.0009509701048955321, + 0.001891160849481821, + 0.006369472946971655, + 0.004708404187113047, + 0.007059886120259762, + 0.004174873698502779, + 0.004026815760880709, + 0.004810475744307041, + 0.0035538224037736654, + 0.004019728396087885, + 0.003501082304865122, + 0.0029185861349105835, + 0.004907091613858938, + 0.0021900124847888947, + 0.0032727729994803667, + 0.0017958665266633034, + 0.0012587105156853795, + 0.0021912348456680775, + 0.003731196979060769, + 0.0033532928209751844, + 0.005959438160061836, + 0.002805118914693594, + 0.0031724250875413418, + 0.0025066444650292397, + 0.0015783008420839906, + 0.0016553106252104044, + 0.004411466885358095, + 0.0033593839034438133, + 0.0032414463348686695, + 0.0019972047302871943, + 0.0037454913835972548, + 0.002192307496443391, + 0.0022741216234862804, + 0.0018620052142068744, + 0.016848085448145866, + 0.010098427534103394, + 0.013542475178837776, + 0.005934359505772591, + 0.011757676489651203, + 0.00840432196855545, + 0.005247507244348526, + 0.0050553991459310055, + 0.007370511535555124, + 0.005718811880797148, + 0.006395210977643728, + 0.0029032896272838116, + 0.0078633027151227, + 0.00509323924779892, + 0.0018170931143686175, + 0.0026988692115992308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.05496343597769737 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a high angle, the camera held still, focused on a sheet of fine, slightly yellowed rice paper. The paper's subtle fibers are clearly visible in the soft sidelight. A wolf-hair brush, its barrel gleaming with the natural sheen of bamboo knots, its full tip saturated with rich, jet-black ink, shimmers with a wet sheen, as if ready to drip. Guided by a steady and graceful hand (not fully captured), the tip touches the paper at a precise angle, instantly creating a tiny dot of ink. Then, with the fluidity of flowing water, the brush moves steadily and deliberately across the paper, lifting and pressing the tip downwards to create a rhythmic and powerful black line. This continuous, undulating line precisely depicts the majestic contours of the distant mountains. The light of the entire scene is soft and even, as if coming from an ancient window lattice. The main colors are the black of ink, the slightly yellow of paper, and the plain white space, creating a tranquil atmosphere of extreme concentration, full of Zen and the beauty of classical art.", + "seed": 24473, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_13/index_4248_seed_24473_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01843478102236986, + "vlm_logprob": -0.01843478102236986, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.01843478102236986, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02508770488202572 + ], + "opposite_logprob_tokens": [ + -4.400087833404541 + ], + "logprob_reward": -0.02508770488202572, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02508770488202572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20843586325645447, + 0.02810688130557537, + 0.03457972779870033, + 0.046238113194704056, + 0.10664787143468857, + 0.07509881258010864, + 0.046733882278203964, + 0.14583221077919006, + 0.16404002904891968, + 0.14428657293319702 + ], + "credit_map": [ + 0.019893886521458626, + 0.02674483321607113, + 0.013315574266016483, + 0.014433076605200768, + 0.03447479009628296, + 0.040230248123407364, + 0.025316594168543816, + 0.0340268611907959, + 0.003905618330463767, + 0.003297936636954546, + 0.0025857556611299515, + 0.0025552818551659584, + 0.0037691702600568533, + 0.004658111836761236, + 0.003324175253510475, + 0.0040108319371938705, + 0.004998298361897469, + 0.004974767565727234, + 0.0047193788923323154, + 0.002122078789398074, + 0.004753266926854849, + 0.0045904554426670074, + 0.004242409951984882, + 0.004179072566330433, + 0.00497348653152585, + 0.007624300196766853, + 0.008700432255864143, + 0.004393373616039753, + 0.003862776793539524, + 0.004778014961630106, + 0.004647791385650635, + 0.007257941644638777, + 0.01163581945002079, + 0.02538049779832363, + 0.00902701448649168, + 0.009097708389163017, + 0.011013333685696125, + 0.008129792287945747, + 0.00800611823797226, + 0.02435758337378502, + 0.00596854742616415, + 0.018782921135425568, + 0.006140869576483965, + 0.0030118937138468027, + 0.029069149866700172, + 0.005441498011350632, + 0.002954677678644657, + 0.0037292581982910633, + 0.005232116673141718, + 0.012165337800979614, + 0.0023857736960053444, + 0.002179495757445693, + 0.016102636232972145, + 0.0043048132210969925, + 0.0014211476081982255, + 0.002942560240626335, + 0.010734746232628822, + 0.04407922551035881, + 0.006881489884108305, + 0.007659997325390577, + 0.059973832219839096, + 0.008830993436276913, + 0.004368339665234089, + 0.003303587669506669, + 0.013446524739265442, + 0.06558463722467422, + 0.008717342279851437, + 0.004115108400583267, + 0.0099826380610466, + 0.054264847189188004, + 0.0037054400891065598, + 0.004223507363349199, + 0.02368781343102455, + 0.039303626865148544, + 0.019109414890408516, + 0.009504836052656174, + 0.006076883990317583, + 0.026953382417559624, + 0.008486101403832436, + 0.011164508759975433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.02508770488202572 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00465986505150795 + ], + "opposite_logprob_tokens": [ + -5.379659652709961 + ], + "logprob_reward": -0.00465986505150795, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00465986505150795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13311178982257843, + 0.0699557214975357, + 0.06468652933835983, + 0.06380288302898407, + 0.1212853491306305, + 0.0906873494386673, + 0.1455836445093155, + 0.12693391740322113, + 0.12305508553981781, + 0.060897767543792725 + ], + "credit_map": [ + 0.02408515475690365, + 0.014067652635276318, + 0.012330717407166958, + 0.008604152128100395, + 0.01828157529234886, + 0.020466085523366928, + 0.019361721351742744, + 0.015914741903543472, + 0.011112871579825878, + 0.007430077996104956, + 0.006966484244912863, + 0.006788109429180622, + 0.011265024542808533, + 0.01049482449889183, + 0.009253935888409615, + 0.0066443937830626965, + 0.00686543807387352, + 0.0103797921910882, + 0.010202642530202866, + 0.005908425897359848, + 0.006051272619515657, + 0.007818189449608326, + 0.011065846309065819, + 0.006394921336323023, + 0.00920672994107008, + 0.009745226241648197, + 0.012720184400677681, + 0.007766334805637598, + 0.004967592656612396, + 0.004481776617467403, + 0.005137264728546143, + 0.009777767583727837, + 0.008608890697360039, + 0.038568783551454544, + 0.008582484908401966, + 0.008264048956334591, + 0.010869164951145649, + 0.006374887190759182, + 0.005801429506391287, + 0.03421565517783165, + 0.00638977624475956, + 0.025778381153941154, + 0.006559162400662899, + 0.0035259004216641188, + 0.03643202781677246, + 0.005753614939749241, + 0.0024522021412849426, + 0.003796289674937725, + 0.011708812788128853, + 0.04015207290649414, + 0.00822338368743658, + 0.0060411482118070126, + 0.052285779267549515, + 0.013486518524587154, + 0.004312026314437389, + 0.009373895823955536, + 0.006487313192337751, + 0.04473991319537163, + 0.005288461688905954, + 0.006947810295969248, + 0.05137290433049202, + 0.005857204552739859, + 0.003402166999876499, + 0.002838129410520196, + 0.010187364183366299, + 0.047876205295324326, + 0.006984327919781208, + 0.003373293671756983, + 0.007059885188937187, + 0.04099477827548981, + 0.0029683082830160856, + 0.003610924119129777, + 0.011345522478222847, + 0.012573519721627235, + 0.004983088467270136, + 0.0049564712680876255, + 0.004164210986346006, + 0.011512723751366138, + 0.002999821212142706, + 0.008362412452697754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.00465986505150795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046440837904810905 + ], + "opposite_logprob_tokens": [ + -5.37964391708374 + ], + "logprob_reward": -0.0046440837904810905, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046440837904810905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14863009750843048, + 0.062163226306438446, + 0.045800402760505676, + 0.0562719851732254, + 0.11146129667758942, + 0.18836377561092377, + 0.08541324734687805, + 0.10282136499881744, + 0.13482020795345306, + 0.06425435096025467 + ], + "credit_map": [ + 0.027528826147317886, + 0.0199102982878685, + 0.012596389278769493, + 0.011869994923472404, + 0.022804586216807365, + 0.022481761872768402, + 0.016512224450707436, + 0.01492602750658989, + 0.006809853482991457, + 0.006522388197481632, + 0.004566977731883526, + 0.009500371292233467, + 0.010215452872216702, + 0.01028531976044178, + 0.006837916560471058, + 0.007424944546073675, + 0.004763235338032246, + 0.00626984890550375, + 0.00919867493212223, + 0.003188295755535364, + 0.005985391326248646, + 0.005348585546016693, + 0.005431512836366892, + 0.00561485905200243, + 0.006690580863505602, + 0.009687483310699463, + 0.009789708070456982, + 0.005481554660946131, + 0.0052633523009717464, + 0.004675969015806913, + 0.005387452431023121, + 0.009295881725847721, + 0.0072432588785886765, + 0.03544364869594574, + 0.005648014601320028, + 0.007141739130020142, + 0.012315049767494202, + 0.010549189522862434, + 0.0057228184305131435, + 0.027397584170103073, + 0.008400787599384785, + 0.0713382214307785, + 0.00865253061056137, + 0.0065890029072761536, + 0.0751602053642273, + 0.008763360790908337, + 0.003946375101804733, + 0.005513266194611788, + 0.005184997338801622, + 0.03054439276456833, + 0.004354329314082861, + 0.002414105227217078, + 0.032866381108760834, + 0.005227892193943262, + 0.001987781375646591, + 0.0028333759400993586, + 0.006996197625994682, + 0.03480268642306328, + 0.0049018459394574165, + 0.002978681353852153, + 0.04370585456490517, + 0.005406400188803673, + 0.002251978497952223, + 0.0017777190078049898, + 0.011390136554837227, + 0.052000097930431366, + 0.007380770985037088, + 0.0036787379067391157, + 0.00800972618162632, + 0.044431764632463455, + 0.0032196450047194958, + 0.004709326196461916, + 0.007417895831167698, + 0.012965256348252296, + 0.00773970689624548, + 0.00458612572401762, + 0.003656214801594615, + 0.011339914985001087, + 0.0038999118842184544, + 0.012649321928620338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.0046440837904810905 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028194934129714966 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.0028194934129714966, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028194934129714966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09064378589391708, + 0.026478491723537445, + 0.022523794323205948, + 0.030724551528692245, + 0.09776376187801361, + 0.09862799197435379, + 0.12372446060180664, + 0.39447587728500366, + 0.04377017542719841, + 0.07126709818840027 + ], + "credit_map": [ + 0.009365417994558811, + 0.01139675173908472, + 0.007274249568581581, + 0.006333213299512863, + 0.015165278688073158, + 0.015362785197794437, + 0.012050681747496128, + 0.013695413246750832, + 0.0032653938978910446, + 0.002195867244154215, + 0.0025344642344862223, + 0.0033671646378934383, + 0.004856023006141186, + 0.0034628910943865776, + 0.0038578840903937817, + 0.002938802121207118, + 0.0030793785117566586, + 0.0026141300331801176, + 0.0030954228714108467, + 0.0025406579952687025, + 0.002954937983304262, + 0.0032036344055086374, + 0.0024222666397690773, + 0.002613365650177002, + 0.0034469752572476864, + 0.005581617821007967, + 0.005121259950101376, + 0.002289655152708292, + 0.003462783293798566, + 0.0034613371826708317, + 0.0028922706842422485, + 0.00446865102276206, + 0.006977573968470097, + 0.029328787699341774, + 0.0073466734029352665, + 0.0076400404796004295, + 0.008147618733346462, + 0.004817786160856485, + 0.004623223561793566, + 0.028882062062621117, + 0.006024383474141359, + 0.03236299380660057, + 0.0049135517328977585, + 0.0027532826643437147, + 0.041688621044158936, + 0.004912536591291428, + 0.002428026869893074, + 0.0035445981193333864, + 0.00883392058312893, + 0.043028831481933594, + 0.007242127321660519, + 0.004549282602965832, + 0.045567020773887634, + 0.006904930807650089, + 0.003518323414027691, + 0.0040800259448587894, + 0.021066773682832718, + 0.14486539363861084, + 0.012870305217802525, + 0.0083374734967947, + 0.174617737531662, + 0.0196432676166296, + 0.0069794850423932076, + 0.00609543826431036, + 0.003450841875746846, + 0.016729678958654404, + 0.0028672993648797274, + 0.0017787128454074264, + 0.002681792713701725, + 0.013715727254748344, + 0.0010636134538799524, + 0.0014825079124420881, + 0.012863744050264359, + 0.01858905516564846, + 0.009240374900400639, + 0.0036830592434853315, + 0.0036642749328166246, + 0.012930100783705711, + 0.004637652076780796, + 0.0056588416919112206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.0028194934129714966 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high angle with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sheet of fine, slightly yellowed rice paper with visible fibers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wolf-hair brush with a bamboo barrel and a tip saturated with jet-black ink visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the brush move across the paper to create a continuous, undulating black line depicting mountain contours?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft and even, resembling light from an ancient window lattice?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496275797486305 + ], + "opposite_logprob_tokens": [ + -2.9299628734588623 + ], + "logprob_reward": -0.05496275797486305, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496275797486305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08580143004655838, + 0.05017136037349701, + 0.050807174295186996, + 0.10167855769395828, + 0.12587957084178925, + 0.049951426684856415, + 0.0567953959107399, + 0.1076057031750679, + 0.22941642999649048, + 0.1418929398059845 + ], + "credit_map": [ + 0.010449182242155075, + 0.007351030129939318, + 0.01119526382535696, + 0.005319700576364994, + 0.008620519191026688, + 0.01635175384581089, + 0.01383237075060606, + 0.012681606225669384, + 0.006952713709324598, + 0.006840888410806656, + 0.004467744380235672, + 0.004950417205691338, + 0.007163801696151495, + 0.007426518015563488, + 0.007375330198556185, + 0.00499394815415144, + 0.008760832250118256, + 0.008600389584898949, + 0.0038883001543581486, + 0.0036472908686846495, + 0.00778112793341279, + 0.007881451398134232, + 0.005486329551786184, + 0.004761451855301857, + 0.012096085585653782, + 0.01756742037832737, + 0.02078108675777912, + 0.009028245694935322, + 0.006489092949777842, + 0.008494625799357891, + 0.009579205885529518, + 0.017642799764871597, + 0.008470933884382248, + 0.040466032922267914, + 0.007813864387571812, + 0.009833145886659622, + 0.012814020738005638, + 0.007973119616508484, + 0.006434515584260225, + 0.03207392990589142, + 0.002963334321975708, + 0.01686740666627884, + 0.0023574770893901587, + 0.0017888968577608466, + 0.020407533273100853, + 0.0024808156304061413, + 0.0014031471218913794, + 0.0016828140942379832, + 0.005129764322191477, + 0.014583983458578587, + 0.004038399551063776, + 0.002316202037036419, + 0.01761789433658123, + 0.005916077643632889, + 0.0019040299812331796, + 0.005289044231176376, + 0.005953775718808174, + 0.0388270728290081, + 0.003649227088317275, + 0.0024958294816315174, + 0.04718363657593727, + 0.005377435125410557, + 0.002411413937807083, + 0.0017073159106075764, + 0.01955748349428177, + 0.09032060205936432, + 0.012780296616256237, + 0.005796787794679403, + 0.013543409295380116, + 0.07568013668060303, + 0.005391538608819246, + 0.006346176844090223, + 0.023225858807563782, + 0.036353856325149536, + 0.020699452608823776, + 0.008896374143660069, + 0.005964489188045263, + 0.025170493870973587, + 0.009094520471990108, + 0.012487883679568768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.05496275797486305 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_14/index_tensor([342, 342])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_14/index_tensor([342, 342])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4ed41394e6c40fe53c3d7e34a19be62b8382ad5c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_14/index_tensor([342, 342])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "surreal macro shot, maintaining an extremely shallow depth of field, smoothly tracks a garden snail from left to right as it slowly crawls across the textured surface of damp stone. The camera moves in a steady, fluid motion, keeping the snail's head in sharp focus. The snail's shell is a delicate spiral of earthy brown and pale yellow, its surface brimming with moisture and displaying fine, concentric growth lines. Its soft, pale beige body is fully extended, its upper pair of optical tentacles, tipped with tiny black eye-pouches, raised and gently wiggling, while shorter, lower tentacles delicately probe the path ahead. The snail advances with a deliberate, mesmerizing rhythm, moving through a vibrant emerald green moss cushioned by dark slate-gray stone. Its trail leaves a sticky, transparent trail that catches the soft morning light, shimmering across the damp, dark surface with a spectacular display of iridescent colors. Soft, unfocused sunlight filters through the invisible canopy, creating gentle highlights on snail shells and glittering on tiny water droplets clinging to moss, evoking a tranquil and contemplative atmosphere of intimacy, natural wonder.", + "seed": 31963, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_14/index_342_seed_31963_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06081602098420262, + "vlm_logprob": -0.06081602098420262, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.06081602098420262, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004908535163849592 + ], + "opposite_logprob_tokens": [ + -5.754908561706543 + ], + "logprob_reward": -0.004908535163849592, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004908535163849592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09380003064870834, + 0.05941176787018776, + 0.09361685812473297, + 0.05355825275182724, + 0.1122821643948555, + 0.1252807378768921, + 0.10719788819551468, + 0.06043033301830292, + 0.24675889313220978, + 0.04766303673386574 + ], + "credit_map": [ + 0.006254382431507111, + 0.0072755650617182255, + 0.012516329064965248, + 0.028355857357382774, + 0.010944192297756672, + 0.010124020278453827, + 0.011216666549444199, + 0.007113010622560978, + 0.0039649587124586105, + 0.0031170849688351154, + 0.009432141669094563, + 0.014004257507622242, + 0.0065337494015693665, + 0.005980375688523054, + 0.009856150485575199, + 0.0065230452455580235, + 0.0074754562228918076, + 0.006548399571329355, + 0.01395624503493309, + 0.006139396224170923, + 0.01017822977155447, + 0.0318543016910553, + 0.007463559508323669, + 0.01000126264989376, + 0.007376570720225573, + 0.005752496886998415, + 0.006658805068582296, + 0.00557198328897357, + 0.007907894439995289, + 0.0071806637570261955, + 0.006792241241782904, + 0.006317595951259136, + 0.014777923002839088, + 0.007971015758812428, + 0.0146569749340415, + 0.008138587698340416, + 0.011896045878529549, + 0.02643393725156784, + 0.01004189345985651, + 0.01836577244102955, + 0.017298899590969086, + 0.013780106790363789, + 0.02478395402431488, + 0.0146620599552989, + 0.013096031732857227, + 0.01910373568534851, + 0.011560726910829544, + 0.010995217598974705, + 0.00975678488612175, + 0.008797695860266685, + 0.01500412542372942, + 0.012363327667117119, + 0.009546375833451748, + 0.029049430042505264, + 0.007733715232461691, + 0.014946420677006245, + 0.009282147511839867, + 0.006288481876254082, + 0.008738468401134014, + 0.0078049772419035435, + 0.007707915268838406, + 0.005276392679661512, + 0.00953675527125597, + 0.005795187782496214, + 0.07081785798072815, + 0.02301483228802681, + 0.023407381027936935, + 0.017474349588155746, + 0.018248334527015686, + 0.01661289855837822, + 0.016585826873779297, + 0.06059738248586655, + 0.005593596026301384, + 0.0035969458986073732, + 0.013107670471072197, + 0.004589869640767574, + 0.006138443015515804, + 0.00423617335036397, + 0.004926586523652077, + 0.0054737417958676815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.004908535163849592 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078222543001175 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.016078222543001175, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078222543001175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13540399074554443, + 0.11434752494096756, + 0.09320161491632462, + 0.0966620072722435, + 0.054139573127031326, + 0.09349633008241653, + 0.23003388941287994, + 0.04548051580786705, + 0.041598714888095856, + 0.09563586860895157 + ], + "credit_map": [ + 0.010876350104808807, + 0.011146148666739464, + 0.013785090297460556, + 0.02768096514046192, + 0.023044809699058533, + 0.01658078096807003, + 0.017043275758624077, + 0.015246570110321045, + 0.007319436874240637, + 0.007085413672029972, + 0.018541736528277397, + 0.03506910800933838, + 0.012089905329048634, + 0.009105756878852844, + 0.014413502998650074, + 0.010722660459578037, + 0.007422447670251131, + 0.00711328350007534, + 0.009593838825821877, + 0.021560030058026314, + 0.007924654521048069, + 0.020313085988163948, + 0.008089381270110607, + 0.011184895411133766, + 0.010199296288192272, + 0.00975855067372322, + 0.011511337012052536, + 0.007635373622179031, + 0.014515703544020653, + 0.0110581424087286, + 0.009938365779817104, + 0.022045237943530083, + 0.005921694450080395, + 0.003937691915780306, + 0.007613879628479481, + 0.004670380614697933, + 0.004261848516762257, + 0.012627807445824146, + 0.00647770194336772, + 0.008628569543361664, + 0.012390421703457832, + 0.009432456456124783, + 0.01687016524374485, + 0.010555910877883434, + 0.011038756929337978, + 0.010113933123648167, + 0.012498867698013783, + 0.010595825500786304, + 0.01832096464931965, + 0.010923757217824459, + 0.032193347811698914, + 0.018376469612121582, + 0.024113979190587997, + 0.0660511702299118, + 0.024899287149310112, + 0.03515491262078285, + 0.007012519054114819, + 0.004955838434398174, + 0.006093332078307867, + 0.004898933693766594, + 0.005714626982808113, + 0.0034620179794728756, + 0.008702525869011879, + 0.004640718922019005, + 0.005306948907673359, + 0.004960100166499615, + 0.0061043440364301205, + 0.006889214273542166, + 0.003090632613748312, + 0.006239472888410091, + 0.006222609896212816, + 0.0027853907085955143, + 0.00907254870980978, + 0.007051210384815931, + 0.02376563660800457, + 0.00859176553785801, + 0.01254834420979023, + 0.009989648126065731, + 0.013882633298635483, + 0.010734084993600845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.016078222543001175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419165637344122 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.01419165637344122, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419165637344122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12712843716144562, + 0.09449005872011185, + 0.09528982639312744, + 0.09369951486587524, + 0.0742446631193161, + 0.09552609175443649, + 0.14398391544818878, + 0.11655837297439575, + 0.05402114614844322, + 0.10505793243646622 + ], + "credit_map": [ + 0.006968407891690731, + 0.009760206565260887, + 0.01640244573354721, + 0.030324243009090424, + 0.020480584353208542, + 0.015225755050778389, + 0.016957107931375504, + 0.011009680107235909, + 0.005151048302650452, + 0.004398162476718426, + 0.013262061402201653, + 0.030752407386898994, + 0.010117526166141033, + 0.008914130739867687, + 0.012348526157438755, + 0.009546198882162571, + 0.010460183955729008, + 0.008359335362911224, + 0.011294309981167316, + 0.00796441175043583, + 0.011862048879265785, + 0.027035484090447426, + 0.007148468401283026, + 0.011165586300194263, + 0.009409311227500439, + 0.008571868762373924, + 0.011329221539199352, + 0.007431648671627045, + 0.014226148836314678, + 0.012269978411495686, + 0.009447895921766758, + 0.021013442426919937, + 0.006123106461018324, + 0.004265207331627607, + 0.010361287742853165, + 0.005986731965094805, + 0.007861148566007614, + 0.021557696163654327, + 0.0073273819871246815, + 0.010762102901935577, + 0.009835729375481606, + 0.00606126245111227, + 0.024323178455233574, + 0.008546164259314537, + 0.009517351165413857, + 0.01429540291428566, + 0.011563604697585106, + 0.011383402161300182, + 0.013687875121831894, + 0.01284700445830822, + 0.02008151076734066, + 0.015682579949498177, + 0.021172765642404556, + 0.028485795482993126, + 0.01702449843287468, + 0.015001893974840641, + 0.008129166439175606, + 0.010846585966646671, + 0.01822366565465927, + 0.012675944715738297, + 0.016527239233255386, + 0.017559830099344254, + 0.02274775505065918, + 0.009848185814917088, + 0.014739308506250381, + 0.004827491939067841, + 0.0055364822037518024, + 0.0045203519985079765, + 0.0038435617461800575, + 0.003455054946243763, + 0.004007246810942888, + 0.013091649860143661, + 0.01216280646622181, + 0.009764627553522587, + 0.030125312507152557, + 0.010111468844115734, + 0.013303550891578197, + 0.009832490235567093, + 0.010011289268732071, + 0.009746386669576168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.01419165637344122 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989134788513184 + ], + "opposite_logprob_tokens": [ + -1.8048913478851318 + ], + "logprob_reward": -0.17989134788513184, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989134788513184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14704446494579315, + 0.27056261897087097, + 0.08049198985099792, + 0.07418811321258545, + 0.08196451514959335, + 0.06846065074205399, + 0.10884224623441696, + 0.07178831845521927, + 0.04905334860086441, + 0.04760374128818512 + ], + "credit_map": [ + 0.00848010927438736, + 0.010191468521952629, + 0.02188137359917164, + 0.029289238154888153, + 0.02284439653158188, + 0.01871107518672943, + 0.02207971177995205, + 0.013567091897130013, + 0.014959831722080708, + 0.012968458235263824, + 0.04459911212325096, + 0.09240898489952087, + 0.028300851583480835, + 0.019965169951319695, + 0.03131493180990219, + 0.0260452963411808, + 0.007800533901900053, + 0.007392533589154482, + 0.010055418126285076, + 0.006793648470193148, + 0.009313582442700863, + 0.02373739890754223, + 0.005537886638194323, + 0.009860983118414879, + 0.008788414299488068, + 0.007419149857014418, + 0.007569499313831329, + 0.007785734720528126, + 0.01139388419687748, + 0.008233786560595036, + 0.00758385332301259, + 0.015413791872560978, + 0.007558654528111219, + 0.00490386039018631, + 0.012139931321144104, + 0.004134959075599909, + 0.008570067584514618, + 0.023876696825027466, + 0.005185580346733332, + 0.01559476274996996, + 0.008361220359802246, + 0.005599370691925287, + 0.016546331346035004, + 0.006355546414852142, + 0.007432159502059221, + 0.008037656545639038, + 0.007531312759965658, + 0.008597052656114101, + 0.00973676797002554, + 0.0080415615811944, + 0.014877140522003174, + 0.01233989093452692, + 0.012872119434177876, + 0.019459303468465805, + 0.018592776730656624, + 0.012922680005431175, + 0.010851147584617138, + 0.0074753230437636375, + 0.009350179694592953, + 0.008985755033791065, + 0.008179634809494019, + 0.004647061228752136, + 0.016212599352002144, + 0.006086616311222315, + 0.012768613174557686, + 0.004176550079137087, + 0.005364728160202503, + 0.0032653564121574163, + 0.004251419100910425, + 0.005315152462571859, + 0.003985326737165451, + 0.009926199913024902, + 0.0055111912079155445, + 0.00634854007512331, + 0.007797050755470991, + 0.006684667896479368, + 0.004044869914650917, + 0.006149573251605034, + 0.005920917261391878, + 0.005146931391209364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.17989134788513184 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890103429555893 + ], + "opposite_logprob_tokens": [ + -2.464010238647461 + ], + "logprob_reward": -0.0890103429555893, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890103429555893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06883294880390167, + 0.13748198747634888, + 0.06164883077144623, + 0.05817783251404762, + 0.03906708210706711, + 0.09143926948308945, + 0.08208684623241425, + 0.023273227736353874, + 0.39547640085220337, + 0.04251556470990181 + ], + "credit_map": [ + 0.0064137219451367855, + 0.0064498125575482845, + 0.009774117730557919, + 0.012089419178664684, + 0.012752917595207691, + 0.006399237085133791, + 0.00879676267504692, + 0.00615696469321847, + 0.008986375294625759, + 0.008988016285002232, + 0.019203167408704758, + 0.03833344206213951, + 0.01603592187166214, + 0.013163303025066853, + 0.017924970015883446, + 0.014846812933683395, + 0.005154498387128115, + 0.0035718728322535753, + 0.0065153068862855434, + 0.013912489637732506, + 0.0050834352150559425, + 0.008783865720033646, + 0.009143218398094177, + 0.009484152309596539, + 0.00709923543035984, + 0.004216607194393873, + 0.008802596479654312, + 0.0064363982528448105, + 0.005789431277662516, + 0.005160405300557613, + 0.007023773156106472, + 0.013649392873048782, + 0.0035189955960959196, + 0.0027653006836771965, + 0.005739521700888872, + 0.00289138057269156, + 0.0037089402321726084, + 0.010403048247098923, + 0.0033998216968029737, + 0.006640075705945492, + 0.009537018835544586, + 0.005263189785182476, + 0.028120556846261024, + 0.007985798642039299, + 0.007475666236132383, + 0.014242870733141899, + 0.006621527951210737, + 0.012192646972835064, + 0.0068926578387618065, + 0.006281990557909012, + 0.010745076462626457, + 0.00739599484950304, + 0.010365531779825687, + 0.020351912826299667, + 0.009392590261995792, + 0.010661100968718529, + 0.002808895893394947, + 0.0027564733754843473, + 0.003582902019843459, + 0.0025314386002719402, + 0.002744466532021761, + 0.002366228261962533, + 0.004248877987265587, + 0.0022339471615850925, + 0.11157061159610748, + 0.03694134205579758, + 0.03726248815655708, + 0.02901664935052395, + 0.030312558636069298, + 0.028278609737753868, + 0.029824888333678246, + 0.09226927906274796, + 0.0047930083237588406, + 0.003974127117544413, + 0.009909006766974926, + 0.004611670505255461, + 0.00516744377091527, + 0.004031446296721697, + 0.005032648798078299, + 0.0049962131306529045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.0890103429555893 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "surreal macro shot, maintaining an extremely shallow depth of field, smoothly tracks a garden snail from left to right as it slowly crawls across the textured surface of damp stone. The camera moves in a steady, fluid motion, keeping the snail's head in sharp focus. The snail's shell is a delicate spiral of earthy brown and pale yellow, its surface brimming with moisture and displaying fine, concentric growth lines. Its soft, pale beige body is fully extended, its upper pair of optical tentacles, tipped with tiny black eye-pouches, raised and gently wiggling, while shorter, lower tentacles delicately probe the path ahead. The snail advances with a deliberate, mesmerizing rhythm, moving through a vibrant emerald green moss cushioned by dark slate-gray stone. Its trail leaves a sticky, transparent trail that catches the soft morning light, shimmering across the damp, dark surface with a spectacular display of iridescent colors. Soft, unfocused sunlight filters through the invisible canopy, creating gentle highlights on snail shells and glittering on tiny water droplets clinging to moss, evoking a tranquil and contemplative atmosphere of intimacy, natural wonder.", + "seed": 31964, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_14/index_342_seed_31964_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05774301174096763, + "vlm_logprob": -0.05774301174096763, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.05774301174096763, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031155638862401247 + ], + "opposite_logprob_tokens": [ + -6.128115653991699 + ], + "logprob_reward": -0.0031155638862401247, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0031155638862401247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3288455605506897, + 0.08913735300302505, + 0.05303831771016121, + 0.03155215084552765, + 0.03382733091711998, + 0.07389909029006958, + 0.046013426035642624, + 0.1280931681394577, + 0.08185514062643051, + 0.1337384283542633 + ], + "credit_map": [ + 0.03573733568191528, + 0.01899322308599949, + 0.054549895226955414, + 0.02288178913295269, + 0.04165634512901306, + 0.09281426668167114, + 0.030453599989414215, + 0.031759146600961685, + 0.007060318253934383, + 0.00893024355173111, + 0.01385626569390297, + 0.01928524300456047, + 0.011814163066446781, + 0.008617118000984192, + 0.007508160080760717, + 0.012065844610333443, + 0.00978958047926426, + 0.006438874173909426, + 0.009556218050420284, + 0.0044216495007276535, + 0.008407500572502613, + 0.0056367553770542145, + 0.004180393647402525, + 0.00460734823718667, + 0.004155324772000313, + 0.0030007059685885906, + 0.0029354458674788475, + 0.004028879106044769, + 0.0031732171773910522, + 0.005177644081413746, + 0.005063094198703766, + 0.0040178424678742886, + 0.003948632627725601, + 0.002804861404001713, + 0.0036941131111234426, + 0.006129208952188492, + 0.003203734289854765, + 0.004357672296464443, + 0.0046696774661540985, + 0.005019431468099356, + 0.008479163981974125, + 0.007561420556157827, + 0.00847829319536686, + 0.010209267027676105, + 0.014504265040159225, + 0.008032958954572678, + 0.008156423456966877, + 0.008477303199470043, + 0.00546861719340086, + 0.008889971300959587, + 0.004775760229676962, + 0.00648621516302228, + 0.004501336254179478, + 0.004409190267324448, + 0.0072967433370649815, + 0.004185597412288189, + 0.016719875857234, + 0.023389212787151337, + 0.014212723821401596, + 0.02288864739239216, + 0.010660132393240929, + 0.010045239701867104, + 0.018001189455389977, + 0.012176151387393475, + 0.005219284910708666, + 0.00632589403539896, + 0.011318040080368519, + 0.009366411715745926, + 0.004614992532879114, + 0.003770962357521057, + 0.028378093615174294, + 0.012861471623182297, + 0.013277839869260788, + 0.008898419328033924, + 0.047272708266973495, + 0.009782396256923676, + 0.006046121008694172, + 0.015263956040143967, + 0.010725334286689758, + 0.02247166633605957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0031155638862401247 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071652993559837 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.016071652993559837, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071652993559837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26021143794059753, + 0.11698133498430252, + 0.13016173243522644, + 0.050333961844444275, + 0.04844121262431145, + 0.0623895563185215, + 0.05437314510345459, + 0.05344276502728462, + 0.12914720177650452, + 0.09451758861541748 + ], + "credit_map": [ + 0.016806719824671745, + 0.014176623895764351, + 0.06926439702510834, + 0.01834939792752266, + 0.025319427251815796, + 0.06555577367544174, + 0.03395061939954758, + 0.016788512468338013, + 0.009870979934930801, + 0.007073018234223127, + 0.025402303785085678, + 0.008642868138849735, + 0.01436520740389824, + 0.03812040761113167, + 0.005234759300947189, + 0.008271795697510242, + 0.010796314105391502, + 0.008870587684214115, + 0.02553810551762581, + 0.008227601647377014, + 0.029315268620848656, + 0.019640043377876282, + 0.009739808738231659, + 0.018034018576145172, + 0.004156538285315037, + 0.003350131446495652, + 0.0048734997399151325, + 0.006633122451603413, + 0.004202236421406269, + 0.007889881730079651, + 0.009556952863931656, + 0.00967160239815712, + 0.004417830612510443, + 0.00640134047716856, + 0.005022907163947821, + 0.006081565283238888, + 0.0069105979055166245, + 0.005895616021007299, + 0.007263537496328354, + 0.006447821389883757, + 0.006648121401667595, + 0.004772051703184843, + 0.004993711598217487, + 0.008267226628959179, + 0.010344862006604671, + 0.010387836024165154, + 0.007551718968898058, + 0.00942403357475996, + 0.008181893266737461, + 0.009526783600449562, + 0.00830770842730999, + 0.008466293103992939, + 0.0037528090178966522, + 0.004623075481504202, + 0.007007218897342682, + 0.004507365636527538, + 0.007870669476687908, + 0.010483510792255402, + 0.004842009395360947, + 0.00799743551760912, + 0.00427706865593791, + 0.0038303371984511614, + 0.007631771732121706, + 0.006509966682642698, + 0.010239135473966599, + 0.01647995412349701, + 0.01396520808339119, + 0.01508321426808834, + 0.009523202665150166, + 0.007885311730206013, + 0.036683693528175354, + 0.019287491217255592, + 0.007540055550634861, + 0.007232044357806444, + 0.03265094757080078, + 0.009988317266106606, + 0.0036942530423402786, + 0.009147158823907375, + 0.00832273717969656, + 0.015942079946398735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.016071652993559837 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944904405623674 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944904405623674, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944904405623674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3580774664878845, + 0.11535875499248505, + 0.09447337687015533, + 0.12277799099683762, + 0.027995245531201363, + 0.04876663535833359, + 0.04932994022965431, + 0.03415803238749504, + 0.07579223066568375, + 0.07327032089233398 + ], + "credit_map": [ + 0.022364262491464615, + 0.019025521352887154, + 0.0914670079946518, + 0.022635983303189278, + 0.030749239027500153, + 0.11746692657470703, + 0.03518583998084068, + 0.01918267272412777, + 0.009952201507985592, + 0.011254548095166683, + 0.0184191782027483, + 0.020441871136426926, + 0.012058066204190254, + 0.017665406689047813, + 0.008426787331700325, + 0.017140699550509453, + 0.004699891433119774, + 0.004514879081398249, + 0.01897011324763298, + 0.0067869010381400585, + 0.022993501275777817, + 0.01267978735268116, + 0.006173692177981138, + 0.017654607072472572, + 0.006487320642918348, + 0.008456433191895485, + 0.01676543615758419, + 0.016854479908943176, + 0.013299861922860146, + 0.02086958847939968, + 0.025523580610752106, + 0.014521279372274876, + 0.0022566677071154118, + 0.0023225608747452497, + 0.0031336434185504913, + 0.003321645315736532, + 0.0031891220714896917, + 0.006774094421416521, + 0.003462473163381219, + 0.0035350399557501078, + 0.004554040264338255, + 0.0037221049424260855, + 0.004461189731955528, + 0.008118703030049801, + 0.008658383972942829, + 0.004614061675965786, + 0.005723841488361359, + 0.008914312347769737, + 0.0077433474361896515, + 0.00814339891076088, + 0.00650863628834486, + 0.0072897495701909065, + 0.0034162644296884537, + 0.00563958752900362, + 0.006330483593046665, + 0.004258474800735712, + 0.004045296460390091, + 0.006323081906884909, + 0.00410285871475935, + 0.005091245751827955, + 0.0029289976228028536, + 0.003464323002845049, + 0.004003155045211315, + 0.004199074115604162, + 0.005814130883663893, + 0.006712908856570721, + 0.01050940528512001, + 0.007735597901046276, + 0.00546472892165184, + 0.005464971996843815, + 0.023866022005677223, + 0.010224461555480957, + 0.0057906643487513065, + 0.006991333793848753, + 0.025700422003865242, + 0.007162019144743681, + 0.002844899659976363, + 0.006745930761098862, + 0.006339769344776869, + 0.01169528253376484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.005944904405623674 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156389474868774 + ], + "opposite_logprob_tokens": [ + -1.701563835144043 + ], + "logprob_reward": -0.20156389474868774, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156389474868774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.497090220451355, + 0.12918032705783844, + 0.09256460517644882, + 0.06584366410970688, + 0.014821898192167282, + 0.059325337409973145, + 0.030472740530967712, + 0.04106559231877327, + 0.01903383433818817, + 0.0506017804145813 + ], + "credit_map": [ + 0.03120570257306099, + 0.029405023902654648, + 0.11589844524860382, + 0.023158352822065353, + 0.06779100745916367, + 0.1907671093940735, + 0.01668141968548298, + 0.022183135151863098, + 0.012228826060891151, + 0.013720519840717316, + 0.019365258514881134, + 0.028609978035092354, + 0.011848293244838715, + 0.013479764573276043, + 0.011550051160156727, + 0.01837764121592045, + 0.00560637004673481, + 0.004209564533084631, + 0.017228255048394203, + 0.008875098079442978, + 0.01916470192372799, + 0.012752801179885864, + 0.009494753554463387, + 0.015233058482408524, + 0.007002153433859348, + 0.00435492442920804, + 0.008703618310391903, + 0.007034606300294399, + 0.004591990727931261, + 0.014494438655674458, + 0.01057391706854105, + 0.009088017977774143, + 0.0013316409895196557, + 0.0016347605269402266, + 0.00133429653942585, + 0.0018538954900577664, + 0.001801669248379767, + 0.0022209619637578726, + 0.0018091876991093159, + 0.0028354860842227936, + 0.004172910936176777, + 0.004391063936054707, + 0.004306561313569546, + 0.009311855770647526, + 0.005446172785013914, + 0.006781908217817545, + 0.007492259610444307, + 0.01742260716855526, + 0.003952618222683668, + 0.004676035605370998, + 0.00370183028280735, + 0.0035727117210626602, + 0.0029681511223316193, + 0.004444127902388573, + 0.0038467294070869684, + 0.0033105378970503807, + 0.0043661342933773994, + 0.00641808845102787, + 0.004601332359015942, + 0.0055460622534155846, + 0.0042160493321716785, + 0.003891418455168605, + 0.00677006971091032, + 0.005256440956145525, + 0.0019065127708017826, + 0.002950883936136961, + 0.001547158812172711, + 0.0031107314862310886, + 0.001744549022987485, + 0.0018971354002133012, + 0.0035056995693594217, + 0.0023711631074547768, + 0.0035850259009748697, + 0.004177962429821491, + 0.016261985525488853, + 0.004289995413273573, + 0.0033777414355427027, + 0.004631138406693935, + 0.005020839162170887, + 0.009257090277969837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.20156389474868774 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062019042670726776 + ], + "opposite_logprob_tokens": [ + -2.812019109725952 + ], + "logprob_reward": -0.062019042670726776, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062019042670726776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.374443382024765, + 0.21232998371124268, + 0.04067925363779068, + 0.030252821743488312, + 0.02240987867116928, + 0.044169068336486816, + 0.051151081919670105, + 0.045790936797857285, + 0.12559880316257477, + 0.05317475274205208 + ], + "credit_map": [ + 0.02318332903087139, + 0.021591458469629288, + 0.09244449436664581, + 0.02057868242263794, + 0.04112793132662773, + 0.13734202086925507, + 0.022855637595057487, + 0.015319861471652985, + 0.017413722351193428, + 0.01359550654888153, + 0.048724062740802765, + 0.011002888903021812, + 0.0288749560713768, + 0.07259302586317062, + 0.008310205303132534, + 0.011815635487437248, + 0.005844568368047476, + 0.003247614949941635, + 0.005476395133882761, + 0.0031208666041493416, + 0.009876181371510029, + 0.006716806907206774, + 0.0031690963078290224, + 0.0032277260906994343, + 0.0022784664761275053, + 0.002617273712530732, + 0.0037408005446195602, + 0.004606487229466438, + 0.00299835205078125, + 0.004114236682653427, + 0.006229023914784193, + 0.003668184159323573, + 0.002252243459224701, + 0.002763960277661681, + 0.0019718639086931944, + 0.0027020645793527365, + 0.0035324341151863337, + 0.003517777193337679, + 0.0029808799736201763, + 0.002688656561076641, + 0.003388115204870701, + 0.002638692269101739, + 0.0034956960007548332, + 0.007481817156076431, + 0.005956138018518686, + 0.005172386299818754, + 0.005005097482353449, + 0.011031129397451878, + 0.006377874407917261, + 0.009515010751783848, + 0.004787932150065899, + 0.006979914382100105, + 0.005259035155177116, + 0.006730576045811176, + 0.006164481397718191, + 0.005336262751370668, + 0.0031058252789080143, + 0.0057510812766849995, + 0.003609137609601021, + 0.011564834043383598, + 0.0030834791250526905, + 0.0038899839855730534, + 0.006466627586632967, + 0.008319973014295101, + 0.008523949421942234, + 0.010016259737312794, + 0.018142934888601303, + 0.012734804302453995, + 0.008171288296580315, + 0.00819643959403038, + 0.042403530329465866, + 0.017409609630703926, + 0.004310429096221924, + 0.005521224811673164, + 0.014092162251472473, + 0.00807712972164154, + 0.0028187409043312073, + 0.005879602860659361, + 0.005688690114766359, + 0.006786782760173082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.062019042670726776 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "surreal macro shot, maintaining an extremely shallow depth of field, smoothly tracks a garden snail from left to right as it slowly crawls across the textured surface of damp stone. The camera moves in a steady, fluid motion, keeping the snail's head in sharp focus. The snail's shell is a delicate spiral of earthy brown and pale yellow, its surface brimming with moisture and displaying fine, concentric growth lines. Its soft, pale beige body is fully extended, its upper pair of optical tentacles, tipped with tiny black eye-pouches, raised and gently wiggling, while shorter, lower tentacles delicately probe the path ahead. The snail advances with a deliberate, mesmerizing rhythm, moving through a vibrant emerald green moss cushioned by dark slate-gray stone. Its trail leaves a sticky, transparent trail that catches the soft morning light, shimmering across the damp, dark surface with a spectacular display of iridescent colors. Soft, unfocused sunlight filters through the invisible canopy, creating gentle highlights on snail shells and glittering on tiny water droplets clinging to moss, evoking a tranquil and contemplative atmosphere of intimacy, natural wonder.", + "seed": 31965, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_14/index_342_seed_31965_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21145181469619273, + "vlm_logprob": -0.21145181469619273, + "vlm_yes_no_margin": 2.300000011920929, + "vlm_token_credit": -0.21145181469619273, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0443846695125103 + ], + "opposite_logprob_tokens": [ + -3.169384717941284 + ], + "logprob_reward": -0.0443846695125103, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0443846695125103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3936571478843689, + 0.15438368916511536, + 0.07114685326814651, + 0.03349198028445244, + 0.05832034349441528, + 0.061967089772224426, + 0.051456842571496964, + 0.01572590135037899, + 0.09855850785970688, + 0.06129167601466179 + ], + "credit_map": [ + 0.0480240061879158, + 0.06782679259777069, + 0.04512462764978409, + 0.07114549726247787, + 0.037686847150325775, + 0.02823736146092415, + 0.06781575083732605, + 0.02779626101255417, + 0.0280571561306715, + 0.01389883179217577, + 0.02212325669825077, + 0.019116129726171494, + 0.012544083409011364, + 0.01689002849161625, + 0.023568009957671165, + 0.018186192959547043, + 0.0074184853583574295, + 0.011520424857735634, + 0.008119001053273678, + 0.011480738408863544, + 0.006772531196475029, + 0.0064384425058960915, + 0.00696859136223793, + 0.012428637593984604, + 0.0031189422588795424, + 0.003630666062235832, + 0.005727797280997038, + 0.005254060961306095, + 0.0030087551567703485, + 0.003266156418249011, + 0.004786332603543997, + 0.00469926605001092, + 0.0057303812354803085, + 0.0047188024036586285, + 0.00816367007791996, + 0.0070089539512991905, + 0.005145680624991655, + 0.010213370434939861, + 0.010904938913881779, + 0.006434550043195486, + 0.00588209368288517, + 0.00515763321891427, + 0.01216508261859417, + 0.006992231123149395, + 0.0050273132510483265, + 0.010700885206460953, + 0.010146637447178364, + 0.005895212758332491, + 0.005359190981835127, + 0.007421082817018032, + 0.006980330217629671, + 0.0034419286530464888, + 0.011771168559789658, + 0.006115727126598358, + 0.006272030062973499, + 0.0040953801944851875, + 0.0014501793775707483, + 0.002306212903931737, + 0.0019382210448384285, + 0.0013866652734577656, + 0.0019561578519642353, + 0.002057708101347089, + 0.0022144292015582323, + 0.0024163275957107544, + 0.008617156185209751, + 0.011068633757531643, + 0.01100081205368042, + 0.014197515323758125, + 0.0153254559263587, + 0.008102836087346077, + 0.01923254504799843, + 0.011013551615178585, + 0.005813791416585445, + 0.006696147844195366, + 0.010234071873128414, + 0.006873581558465958, + 0.004664945881813765, + 0.005767383147031069, + 0.009497715160250664, + 0.011744040995836258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0443846695125103 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023296110332012177 + ], + "opposite_logprob_tokens": [ + -3.7732961177825928 + ], + "logprob_reward": -0.023296110332012177, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023296110332012177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34214380383491516, + 0.17073450982570648, + 0.08130072057247162, + 0.10973476618528366, + 0.033557385206222534, + 0.04800432175397873, + 0.024404054507613182, + 0.048822738230228424, + 0.08289013057947159, + 0.05840754136443138 + ], + "credit_map": [ + 0.04404196888208389, + 0.05771093815565109, + 0.037961751222610474, + 0.06366026401519775, + 0.034977588802576065, + 0.022341448813676834, + 0.06026657670736313, + 0.02118329145014286, + 0.02159213088452816, + 0.015720149502158165, + 0.02639562077820301, + 0.024684416130185127, + 0.015152310952544212, + 0.016005387529730797, + 0.028262123465538025, + 0.022922372445464134, + 0.008775100111961365, + 0.0133309131488204, + 0.010643891990184784, + 0.008359041064977646, + 0.00870940089225769, + 0.007656197529286146, + 0.009222764521837234, + 0.014603412710130215, + 0.0100636575371027, + 0.01618773862719536, + 0.016419347375631332, + 0.011686045676469803, + 0.012500345706939697, + 0.01068330928683281, + 0.017276158556342125, + 0.014918164350092411, + 0.0035828007385134697, + 0.003920774906873703, + 0.005230110604315996, + 0.005042685195803642, + 0.003892225679010153, + 0.003243994200602174, + 0.005590285640209913, + 0.003054506378248334, + 0.005025921389460564, + 0.005932918284088373, + 0.008269020356237888, + 0.004693083930760622, + 0.003911097068339586, + 0.005880358628928661, + 0.007507441099733114, + 0.006784482859075069, + 0.00347623392008245, + 0.004106395877897739, + 0.002970590488985181, + 0.003349857870489359, + 0.002300737891346216, + 0.002705187303945422, + 0.003668553661555052, + 0.0018264998216181993, + 0.0048168497160077095, + 0.006441292352974415, + 0.006025844719260931, + 0.006246959324926138, + 0.005176478996872902, + 0.005150507669895887, + 0.007063473109155893, + 0.007901332341134548, + 0.007419027388095856, + 0.009655337780714035, + 0.00869610533118248, + 0.01291951909661293, + 0.01216250378638506, + 0.007019442040473223, + 0.015891363844275475, + 0.009126833640038967, + 0.009150258265435696, + 0.006992613431066275, + 0.007620146032422781, + 0.004617844242602587, + 0.006031875964254141, + 0.0062285237945616245, + 0.008378775790333748, + 0.009387500584125519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.023296110332012177 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634178288280964 + ], + "opposite_logprob_tokens": [ + -3.651341676712036 + ], + "logprob_reward": -0.02634178288280964, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634178288280964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2193364053964615, + 0.284244567155838, + 0.09964317828416824, + 0.040846552699804306, + 0.039771031588315964, + 0.058222342282533646, + 0.056544866412878036, + 0.07021131366491318, + 0.05043358355760574, + 0.08074620366096497 + ], + "credit_map": [ + 0.027005115523934364, + 0.03927955403923988, + 0.022011270746588707, + 0.03723197430372238, + 0.021489756181836128, + 0.01495039090514183, + 0.04525711387395859, + 0.012111230753362179, + 0.06720176339149475, + 0.028823407366871834, + 0.04478462412953377, + 0.02753790281713009, + 0.021495351567864418, + 0.023729052394628525, + 0.037928394973278046, + 0.03274407610297203, + 0.010901890695095062, + 0.014180103316903114, + 0.012609481811523438, + 0.014242441393435001, + 0.010793917812407017, + 0.008166622370481491, + 0.015918949618935585, + 0.012829764746129513, + 0.005007675848901272, + 0.0043684206902980804, + 0.005170641001313925, + 0.004565437324345112, + 0.003346585901454091, + 0.003938425797969103, + 0.007871105335652828, + 0.0065782638266682625, + 0.004160686396062374, + 0.004581931047141552, + 0.0064921616576612, + 0.0047455704770982265, + 0.003679150715470314, + 0.004500680603086948, + 0.007215934339910746, + 0.004394915420562029, + 0.007542287930846214, + 0.007109606638550758, + 0.006884963251650333, + 0.005243894178420305, + 0.006526743993163109, + 0.00968220829963684, + 0.008150601759552956, + 0.007082035299390554, + 0.006070909556001425, + 0.006104700267314911, + 0.00658365897834301, + 0.007067515980452299, + 0.0052832248620688915, + 0.008573056198656559, + 0.010576565749943256, + 0.006285232957452536, + 0.006686612032353878, + 0.01170596107840538, + 0.008401479572057724, + 0.009917003102600574, + 0.005581215023994446, + 0.008246885612607002, + 0.010342820547521114, + 0.009329336695373058, + 0.004844205919653177, + 0.005355930421501398, + 0.005930953659117222, + 0.007536652497947216, + 0.006685827858746052, + 0.0035481718368828297, + 0.010888306424021721, + 0.005643532145768404, + 0.008044596761465073, + 0.006013923790305853, + 0.013621203601360321, + 0.006587340962141752, + 0.007132317870855331, + 0.010541393421590328, + 0.011731255799531937, + 0.01707417517900467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02634178288280964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761959552764893 + ], + "opposite_logprob_tokens": [ + -0.8261959552764893 + ], + "logprob_reward": -0.5761959552764893, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761959552764893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3257211744785309, + 0.18224123120307922, + 0.060714710503816605, + 0.0601009875535965, + 0.06200851872563362, + 0.060718826949596405, + 0.05074043199419975, + 0.07745004445314407, + 0.03812086582183838, + 0.08218316733837128 + ], + "credit_map": [ + 0.04084107652306557, + 0.056328389793634415, + 0.03550148010253906, + 0.058851372450590134, + 0.032687798142433167, + 0.021474532783031464, + 0.06030033156275749, + 0.01973620243370533, + 0.029830314218997955, + 0.01506024319678545, + 0.03161007910966873, + 0.02213890664279461, + 0.01583324559032917, + 0.016916200518608093, + 0.030338523909449577, + 0.020513728260993958, + 0.0057929218746721745, + 0.0084579698741436, + 0.0073195756413042545, + 0.008634966798126698, + 0.008122310973703861, + 0.005461603868752718, + 0.01065932959318161, + 0.00626602815464139, + 0.00926102977246046, + 0.009409703314304352, + 0.007821806706488132, + 0.005388014949858189, + 0.006337258964776993, + 0.007619017269462347, + 0.007891898043453693, + 0.006372256204485893, + 0.007096204441040754, + 0.005954139865934849, + 0.010340291075408459, + 0.00779362116008997, + 0.005223303101956844, + 0.008296996355056763, + 0.010912954807281494, + 0.006391006987541914, + 0.0065590268932282925, + 0.0061113047413527966, + 0.009815667755901814, + 0.006552319973707199, + 0.006886765826493502, + 0.007527355570346117, + 0.00878442544490099, + 0.008481960743665695, + 0.00445919344201684, + 0.007140599191188812, + 0.00591094046831131, + 0.005935802590101957, + 0.00875240471214056, + 0.006497492548078299, + 0.007373230531811714, + 0.0046707685105502605, + 0.008077174425125122, + 0.012094561941921711, + 0.010024172253906727, + 0.009266434237360954, + 0.008764629252254963, + 0.008162956684827805, + 0.013012913055717945, + 0.008047204464673996, + 0.0032638635020703077, + 0.0033851508051156998, + 0.004203290678560734, + 0.0030232032295316458, + 0.004584306851029396, + 0.006850088480859995, + 0.006469924934208393, + 0.006341038271784782, + 0.009352167136967182, + 0.007145233452320099, + 0.013271782547235489, + 0.007212387397885323, + 0.007398940157145262, + 0.007465158589184284, + 0.013327410444617271, + 0.01701008901000023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.5761959552764893 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38704055547714233 + ], + "opposite_logprob_tokens": [ + -1.137040615081787 + ], + "logprob_reward": -0.38704055547714233, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38704055547714233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10813862830400467, + 0.12502902746200562, + 0.14966480433940887, + 0.04863972216844559, + 0.06869690120220184, + 0.16199231147766113, + 0.06300299614667892, + 0.07317063212394714, + 0.11561484634876251, + 0.08605015277862549 + ], + "credit_map": [ + 0.015771426260471344, + 0.010680506937205791, + 0.01807904615998268, + 0.010502812452614307, + 0.00981078576296568, + 0.01609054207801819, + 0.014981991611421108, + 0.012221509590744972, + 0.014736761339008808, + 0.014091851189732552, + 0.019552189856767654, + 0.018069345504045486, + 0.011441251263022423, + 0.01192694902420044, + 0.0204565841704607, + 0.014754078350961208, + 0.016982518136501312, + 0.021958593279123306, + 0.021228330209851265, + 0.023523565381765366, + 0.01546839065849781, + 0.010525977239012718, + 0.023733999580144882, + 0.01624341495335102, + 0.005570079665631056, + 0.005896356422454119, + 0.005571406800299883, + 0.005212694872170687, + 0.005501218605786562, + 0.005376990884542465, + 0.009871498681604862, + 0.0056394715793430805, + 0.007556497119367123, + 0.006097516510635614, + 0.01001825649291277, + 0.008884843438863754, + 0.00716007687151432, + 0.010535495355725288, + 0.011700259521603584, + 0.006743948440998793, + 0.022208819165825844, + 0.0209121722728014, + 0.01995280757546425, + 0.013795564882457256, + 0.015228798612952232, + 0.030036529526114464, + 0.022668663412332535, + 0.017188942059874535, + 0.00512709841132164, + 0.00812617503106594, + 0.006645476911216974, + 0.0058658733032643795, + 0.013997959904372692, + 0.011791749857366085, + 0.006807172205299139, + 0.004641487263143063, + 0.008390022441744804, + 0.013982631266117096, + 0.008549507707357407, + 0.007682964205741882, + 0.006383766420185566, + 0.005847741849720478, + 0.011757503263652325, + 0.01057648565620184, + 0.010367237031459808, + 0.01234697550535202, + 0.01337700430303812, + 0.01300863828510046, + 0.016420878469944, + 0.014574125409126282, + 0.02106928825378418, + 0.014450696296989918, + 0.009198402985930443, + 0.008915083482861519, + 0.010481064207851887, + 0.007356752175837755, + 0.010594003833830357, + 0.014699244871735573, + 0.012473666109144688, + 0.012331927195191383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.38704055547714233 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "surreal macro shot, maintaining an extremely shallow depth of field, smoothly tracks a garden snail from left to right as it slowly crawls across the textured surface of damp stone. The camera moves in a steady, fluid motion, keeping the snail's head in sharp focus. The snail's shell is a delicate spiral of earthy brown and pale yellow, its surface brimming with moisture and displaying fine, concentric growth lines. Its soft, pale beige body is fully extended, its upper pair of optical tentacles, tipped with tiny black eye-pouches, raised and gently wiggling, while shorter, lower tentacles delicately probe the path ahead. The snail advances with a deliberate, mesmerizing rhythm, moving through a vibrant emerald green moss cushioned by dark slate-gray stone. Its trail leaves a sticky, transparent trail that catches the soft morning light, shimmering across the damp, dark surface with a spectacular display of iridescent colors. Soft, unfocused sunlight filters through the invisible canopy, creating gentle highlights on snail shells and glittering on tiny water droplets clinging to moss, evoking a tranquil and contemplative atmosphere of intimacy, natural wonder.", + "seed": 31966, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_14/index_342_seed_31966_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09536248277872801, + "vlm_logprob": -0.09536248277872801, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.09536248277872801, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032420000061392784 + ], + "opposite_logprob_tokens": [ + -6.128242015838623 + ], + "logprob_reward": -0.0032420000061392784, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0032420000061392784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15508541464805603, + 0.05766730010509491, + 0.06270109862089157, + 0.11463506519794464, + 0.0662609189748764, + 0.1412554830312729, + 0.12102922052145004, + 0.1684754341840744, + 0.04345456138253212, + 0.06943541765213013 + ], + "credit_map": [ + 0.012544319964945316, + 0.013442860916256905, + 0.02515997551381588, + 0.01776481233537197, + 0.02502838894724846, + 0.019547410309314728, + 0.02319694496691227, + 0.018400724977254868, + 0.009049895219504833, + 0.005698859225958586, + 0.005506264511495829, + 0.0070909783244132996, + 0.007765262387692928, + 0.0084544587880373, + 0.008800714276731014, + 0.005300873890519142, + 0.006262834649533033, + 0.004798319656401873, + 0.0062154438346624374, + 0.005566793959587812, + 0.010120031423866749, + 0.012486173771321774, + 0.008290013298392296, + 0.008961490355432034, + 0.016175493597984314, + 0.010975444689393044, + 0.012352346442639828, + 0.013295992277562618, + 0.020753508433699608, + 0.01646309904754162, + 0.01339332852512598, + 0.011225856840610504, + 0.008356857113540173, + 0.005607091821730137, + 0.006755412090569735, + 0.008321627043187618, + 0.007489729207009077, + 0.011211981065571308, + 0.010280418209731579, + 0.008237812668085098, + 0.015838485211133957, + 0.010393618606030941, + 0.013208221644163132, + 0.029714837670326233, + 0.013303813524544239, + 0.02304096519947052, + 0.014791801571846008, + 0.0209637563675642, + 0.0070188925601542, + 0.00588965741917491, + 0.008628479205071926, + 0.032132070511579514, + 0.005857032723724842, + 0.030425548553466797, + 0.017156846821308136, + 0.01392071321606636, + 0.017398769035935402, + 0.012707225047051907, + 0.01310537476092577, + 0.04147970676422119, + 0.011187057010829449, + 0.04683447629213333, + 0.015517481602728367, + 0.010245359502732754, + 0.003114621387794614, + 0.0029345902148634195, + 0.00642851646989584, + 0.006852447986602783, + 0.003930571489036083, + 0.005604694597423077, + 0.0076755075715482235, + 0.006913614924997091, + 0.003611248219385743, + 0.0021642623469233513, + 0.012393467128276825, + 0.006720801815390587, + 0.00577155128121376, + 0.024096956476569176, + 0.005363850854337215, + 0.009313291870057583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.0032420000061392784 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023296110332012177 + ], + "opposite_logprob_tokens": [ + -3.7732961177825928 + ], + "logprob_reward": -0.023296110332012177, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023296110332012177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1373617798089981, + 0.05436280369758606, + 0.12930932641029358, + 0.1334206908941269, + 0.046940773725509644, + 0.1258009970188141, + 0.0675373300909996, + 0.1601388156414032, + 0.037238918244838715, + 0.10788855701684952 + ], + "credit_map": [ + 0.01955735683441162, + 0.021868012845516205, + 0.01806148886680603, + 0.014692407101392746, + 0.018509740009903908, + 0.013951213099062443, + 0.014065820723772049, + 0.016655737534165382, + 0.006507359445095062, + 0.005883026402443647, + 0.004929238930344582, + 0.00641317293047905, + 0.007900874130427837, + 0.010326464660465717, + 0.007831796072423458, + 0.004570868797600269, + 0.02063743770122528, + 0.013635463081300259, + 0.0193424541503191, + 0.013900146819651127, + 0.011755662970244884, + 0.02516842447221279, + 0.013053721748292446, + 0.011816009879112244, + 0.01822485774755478, + 0.01301579363644123, + 0.014266982674598694, + 0.015494103543460369, + 0.024757934734225273, + 0.01939702033996582, + 0.015718966722488403, + 0.012545028701424599, + 0.006132334936410189, + 0.005415599327534437, + 0.005435179453343153, + 0.00677201384678483, + 0.004880438558757305, + 0.004732621368020773, + 0.006256149150431156, + 0.007316438481211662, + 0.011712067760527134, + 0.012159408070147038, + 0.009725957177579403, + 0.03178338706493378, + 0.018062850460410118, + 0.015973074361681938, + 0.013443034142255783, + 0.012941212393343449, + 0.006710786372423172, + 0.005283679347485304, + 0.007668821606785059, + 0.010629015974700451, + 0.0039012618362903595, + 0.011644069105386734, + 0.011761783622205257, + 0.009937908500432968, + 0.013812975026667118, + 0.011426440440118313, + 0.010092957876622677, + 0.04615483805537224, + 0.009141298942267895, + 0.05190972238779068, + 0.010490085929632187, + 0.007110506761819124, + 0.0033784399274736643, + 0.003144063288345933, + 0.005041738972067833, + 0.006084497552365065, + 0.005833652336150408, + 0.004465550649911165, + 0.00456322031095624, + 0.004727753344923258, + 0.005896225571632385, + 0.004010335076600313, + 0.012309974990785122, + 0.011484459042549133, + 0.009308337233960629, + 0.038011569529771805, + 0.006336961407214403, + 0.020530687645077705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.023296110332012177 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633077144622803 + ], + "opposite_logprob_tokens": [ + -4.758633136749268 + ], + "logprob_reward": -0.008633077144622803, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633077144622803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1109398677945137, + 0.054799217730760574, + 0.10774882137775421, + 0.10463428497314453, + 0.05245813727378845, + 0.07942122220993042, + 0.09214038401842117, + 0.24337942898273468, + 0.06464284658432007, + 0.0898357629776001 + ], + "credit_map": [ + 0.011854342184960842, + 0.013963449746370316, + 0.011689679697155952, + 0.011123388074338436, + 0.017572063952684402, + 0.013057449832558632, + 0.014467372559010983, + 0.017212122678756714, + 0.0069944425486028194, + 0.00645905127748847, + 0.006239254027605057, + 0.007167032919824123, + 0.007604873273521662, + 0.006661343388259411, + 0.008713480085134506, + 0.004959738813340664, + 0.014913494698703289, + 0.010749408043920994, + 0.017029574140906334, + 0.011566465720534325, + 0.009090899489820004, + 0.024606218561530113, + 0.007930467836558819, + 0.01186229195445776, + 0.014201298356056213, + 0.009895604103803635, + 0.011207041330635548, + 0.011642837896943092, + 0.018793629482388496, + 0.01567113772034645, + 0.01217630598694086, + 0.011046428233385086, + 0.006844435352832079, + 0.005517951212823391, + 0.005409157369285822, + 0.00641630170866847, + 0.006302177906036377, + 0.0057882508262991905, + 0.008429577574133873, + 0.007750282995402813, + 0.010296453721821308, + 0.006056952755898237, + 0.008092304691672325, + 0.010321693494915962, + 0.006741918157786131, + 0.01156864408403635, + 0.014722021296620369, + 0.011621230281889439, + 0.010366910137236118, + 0.007661464158445597, + 0.010952360928058624, + 0.014774778857827187, + 0.005966269411146641, + 0.013376119546592236, + 0.015588630922138691, + 0.013453853316605091, + 0.018497571349143982, + 0.015923606231808662, + 0.014157824218273163, + 0.07303931564092636, + 0.012555484659969807, + 0.08200056850910187, + 0.016291320323944092, + 0.010913725942373276, + 0.006419598124921322, + 0.0055680000223219395, + 0.00874452106654644, + 0.010184520855545998, + 0.009568781591951847, + 0.0075108446180820465, + 0.00815478153526783, + 0.008491797372698784, + 0.004840712994337082, + 0.003813795279711485, + 0.009981644339859486, + 0.012771363370120525, + 0.007163699716329575, + 0.02708292379975319, + 0.007233434822410345, + 0.016948189586400986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.008633077144622803 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035784780979156 + ], + "opposite_logprob_tokens": [ + -1.9103578329086304 + ], + "logprob_reward": -0.16035784780979156, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035784780979156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17456547915935516, + 0.029984263703227043, + 0.10460026562213898, + 0.03198658302426338, + 0.07139219343662262, + 0.13425227999687195, + 0.05921989679336548, + 0.10425321757793427, + 0.029194794595241547, + 0.2605510354042053 + ], + "credit_map": [ + 0.013240696862339973, + 0.01753665693104267, + 0.0221644788980484, + 0.018010783940553665, + 0.024178937077522278, + 0.022918542847037315, + 0.027416054159402847, + 0.02909933216869831, + 0.0034634023904800415, + 0.0030674345325678587, + 0.00371694378554821, + 0.0033973203971982002, + 0.0027320038061589003, + 0.003830899251624942, + 0.004646327346563339, + 0.005129936151206493, + 0.009861656464636326, + 0.0073943184688687325, + 0.008754130452871323, + 0.01039263978600502, + 0.020809177309274673, + 0.0222226120531559, + 0.013046731241047382, + 0.012119005434215069, + 0.004365765489637852, + 0.0031924969516694546, + 0.0041869343258440495, + 0.004002186935395002, + 0.005072621628642082, + 0.004192995838820934, + 0.003375501837581396, + 0.003598082810640335, + 0.0060808430425822735, + 0.003972249571233988, + 0.009433493949472904, + 0.008782857097685337, + 0.008317536674439907, + 0.015575272031128407, + 0.008735890500247478, + 0.010494058020412922, + 0.012652954086661339, + 0.010026201605796814, + 0.011857259087264538, + 0.027102665975689888, + 0.009639223106205463, + 0.021244511008262634, + 0.020669279620051384, + 0.021060191094875336, + 0.0036712929140776396, + 0.002905199769884348, + 0.004975414369255304, + 0.013359492644667625, + 0.0037428378127515316, + 0.01170685701072216, + 0.009356368333101273, + 0.00950243417173624, + 0.00838179886341095, + 0.007123651448637247, + 0.006701853591948748, + 0.029336119070649147, + 0.006200758274644613, + 0.0332338847219944, + 0.007585807703435421, + 0.005689357873052359, + 0.0028805150650441647, + 0.0026064368430525064, + 0.003913396038115025, + 0.004570075776427984, + 0.003943204879760742, + 0.002901967614889145, + 0.00397089496254921, + 0.004408303182572126, + 0.013230646960437298, + 0.006496417336165905, + 0.03786608204245567, + 0.021356243640184402, + 0.02141542173922062, + 0.0954996719956398, + 0.017349909991025925, + 0.047336649149656296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.16035784780979156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a garden snail with a spiral shell of earthy brown and pale yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the snail's body pale beige, fully extended, with upper tentacles raised and wiggling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the snail crawling across a textured surface of damp stone covered in vibrant emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the snail leave a sticky, transparent trail that shimmers with iridescent colors in the light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly track the snail from left to right while maintaining a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2812833786010742 + ], + "opposite_logprob_tokens": [ + -1.4062833786010742 + ], + "logprob_reward": -0.2812833786010742, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2812833786010742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10345976799726486, + 0.06692052632570267, + 0.08720698207616806, + 0.11824923753738403, + 0.057450223714113235, + 0.08856619149446487, + 0.1282813549041748, + 0.06314495950937271, + 0.053458087146282196, + 0.23326271772384644 + ], + "credit_map": [ + 0.010753669776022434, + 0.01155068352818489, + 0.014398910105228424, + 0.009846627712249756, + 0.01304347813129425, + 0.013215833343565464, + 0.014366230927407742, + 0.016284340992569923, + 0.00973353162407875, + 0.007665696553885937, + 0.007018980570137501, + 0.008586357347667217, + 0.0075963265262544155, + 0.006759323179721832, + 0.012514820322394371, + 0.00704549066722393, + 0.00875022727996111, + 0.007023596204817295, + 0.01018628291785717, + 0.008485653437674046, + 0.012970973737537861, + 0.018483994528651237, + 0.009852944873273373, + 0.011453301645815372, + 0.01678571291267872, + 0.011565159074962139, + 0.011849328875541687, + 0.012583157047629356, + 0.021578555926680565, + 0.017717627808451653, + 0.013965733349323273, + 0.012203961610794067, + 0.007784093730151653, + 0.005605033598840237, + 0.005698671098798513, + 0.00672269519418478, + 0.007221250794827938, + 0.0062826587818562984, + 0.011040863581001759, + 0.0070949578657746315, + 0.012147237546741962, + 0.00855037197470665, + 0.00752420537173748, + 0.016514485701918602, + 0.006071043200790882, + 0.011943915858864784, + 0.010983491316437721, + 0.014831441454589367, + 0.01550016738474369, + 0.012992673553526402, + 0.018205491825938225, + 0.01940779946744442, + 0.01404732745140791, + 0.017242010682821274, + 0.01680217683315277, + 0.014083717949688435, + 0.0069116586819291115, + 0.005885728634893894, + 0.00696358410641551, + 0.011522009037435055, + 0.005496163386851549, + 0.014702717773616314, + 0.006999550852924585, + 0.00466354563832283, + 0.004309357143938541, + 0.004467912949621677, + 0.007840484380722046, + 0.009622865356504917, + 0.006459337193518877, + 0.006396085023880005, + 0.00717893335968256, + 0.007183108478784561, + 0.012051895260810852, + 0.007431538309901953, + 0.027398904785513878, + 0.022218219935894012, + 0.017426107078790665, + 0.0831097662448883, + 0.01620212011039257, + 0.047424159944057465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.2812833786010742 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_15/index_tensor([3720, 3720])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_15/index_tensor([3720, 3720])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5cfc170dd2196a9ea7db77791c78287b36175e6d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_15/index_tensor([3720, 3720])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "powerful handheld close-up, shaky and violent, closely follows a violent, downward swinging motion, focusing on the fist of an adult male. Veins bulge, knuckles white from the intense force, skin crisscrossed with tiny scars and rough texture. Illuminated by a cold, hard overhead light from directly above, this fist smashes with thunderous force into a bright red Fuji apple resting on a worn, dark-grained wooden table. The apple's smooth skin reflects a glaring flash of light just before impact. Instantly, with a crisp, popping sound, the red skin and white flesh, mixed with splattering juice, explode in slow motion, flying to every corner of the camera. The fist finally sinks deep into the apple's wreckage, trembling slightly. The entire image is filled with raw, unvarnished power and suppressed rage, imbued with a gritty, tense documentary quality.", + "seed": 30582, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_15/index_3720_seed_30582_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1714447087608278, + "vlm_logprob": -0.1714447087608278, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.1714447087608278, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10128948837518692 + ], + "opposite_logprob_tokens": [ + -2.4762895107269287 + ], + "logprob_reward": -0.10128948837518692, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10128948837518692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11769379675388336, + 0.20031847059726715, + 0.04808340594172478, + 0.05093465745449066, + 0.04398937150835991, + 0.06612132489681244, + 0.13096250593662262, + 0.07744696736335754, + 0.22642718255519867, + 0.03802226111292839 + ], + "credit_map": [ + 0.010083575733006, + 0.010865823365747929, + 0.015371110290288925, + 0.016966771334409714, + 0.013715366832911968, + 0.012336763553321362, + 0.027301473543047905, + 0.01105291023850441, + 0.012379281222820282, + 0.025655275210738182, + 0.028200900182127953, + 0.02240077406167984, + 0.02670549601316452, + 0.0196369756013155, + 0.048368118703365326, + 0.016971642151474953, + 0.006659982725977898, + 0.005373284686356783, + 0.004986980464309454, + 0.005583181045949459, + 0.0047378577291965485, + 0.007565014064311981, + 0.0061020501889288425, + 0.007075053174048662, + 0.00719452416524291, + 0.007279704324901104, + 0.005902189761400223, + 0.005139230750501156, + 0.004424168262630701, + 0.009501786902546883, + 0.005427421070635319, + 0.006065630819648504, + 0.004019573796540499, + 0.004963668528944254, + 0.0045346156693995, + 0.005544643383473158, + 0.003776007331907749, + 0.006874501705169678, + 0.006254020147025585, + 0.008022342808544636, + 0.00509173609316349, + 0.010337106883525848, + 0.006237874738872051, + 0.010961239226162434, + 0.00789424404501915, + 0.009697044268250465, + 0.007808751426637173, + 0.008093328215181828, + 0.012930233031511307, + 0.017969287931919098, + 0.01625562459230423, + 0.020949749276041985, + 0.014077521860599518, + 0.020514747127890587, + 0.011252802796661854, + 0.01701253093779087, + 0.007324676029384136, + 0.0065063838846981525, + 0.010702528990805149, + 0.010595100931823254, + 0.005053095985203981, + 0.013810955919325352, + 0.008208532817661762, + 0.01524569932371378, + 0.01906997710466385, + 0.04106728360056877, + 0.020364979282021523, + 0.03316480666399002, + 0.023782668635249138, + 0.035907164216041565, + 0.02138378657400608, + 0.031686510890722275, + 0.003490353934466839, + 0.0051832133904099464, + 0.003678099485114217, + 0.007672613486647606, + 0.0024030101485550404, + 0.006434778682887554, + 0.004934032913297415, + 0.00422615697607398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.10128948837518692 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012571258470416069 + ], + "opposite_logprob_tokens": [ + -4.387571334838867 + ], + "logprob_reward": -0.012571258470416069, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012571258470416069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23978348076343536, + 0.1434668004512787, + 0.18009938299655914, + 0.06640870124101639, + 0.049884796142578125, + 0.06227537617087364, + 0.03503753989934921, + 0.04922907054424286, + 0.10489574074745178, + 0.06891915202140808 + ], + "credit_map": [ + 0.017420323565602303, + 0.022455058991909027, + 0.02425459213554859, + 0.02737375721335411, + 0.03251182660460472, + 0.025633450597524643, + 0.0626889318227768, + 0.027445554733276367, + 0.011705337092280388, + 0.01932166889309883, + 0.013303630985319614, + 0.01445457711815834, + 0.021464960649609566, + 0.014097416773438454, + 0.030952809378504753, + 0.01816639117896557, + 0.017298731952905655, + 0.025456268340349197, + 0.020402194932103157, + 0.018593674525618553, + 0.027817901223897934, + 0.024668345227837563, + 0.023610491305589676, + 0.022251758724451065, + 0.0077010090462863445, + 0.008371278643608093, + 0.009698034264147282, + 0.005973960738629103, + 0.005943236872553825, + 0.009842617437243462, + 0.007310608867555857, + 0.011567960493266582, + 0.004345554392784834, + 0.00700117414817214, + 0.004616125486791134, + 0.007906756363809109, + 0.005295679438859224, + 0.00808737613260746, + 0.006261094007641077, + 0.006371037103235722, + 0.004213332664221525, + 0.010246987454593182, + 0.004441626835614443, + 0.012750414200127125, + 0.007222102954983711, + 0.009801926091313362, + 0.007048935629427433, + 0.006550052668899298, + 0.0023126136511564255, + 0.0026765980292111635, + 0.004548932425677776, + 0.007497897371649742, + 0.003321721451357007, + 0.006721939891576767, + 0.0032307610381394625, + 0.004727074410766363, + 0.004463201854377985, + 0.004872181918472052, + 0.005703860893845558, + 0.008885624818503857, + 0.004301415290683508, + 0.007466312497854233, + 0.0066771418787539005, + 0.0068593272008001804, + 0.007479983847588301, + 0.011509997770190239, + 0.009359823539853096, + 0.02521098591387272, + 0.007348641753196716, + 0.018549693748354912, + 0.014370064251124859, + 0.011066553182899952, + 0.00661355210468173, + 0.012605979107320309, + 0.005735170561820269, + 0.013090521097183228, + 0.007305696606636047, + 0.009892304427921772, + 0.006807994097471237, + 0.006867933552712202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.012571258470416069 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762308835983276 + ], + "opposite_logprob_tokens": [ + -0.8262308835983276 + ], + "logprob_reward": -0.5762308835983276, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762308835983276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07441224902868271, + 0.2764541208744049, + 0.2379622608423233, + 0.06694553047418594, + 0.048210084438323975, + 0.03886014223098755, + 0.03497006744146347, + 0.02363862842321396, + 0.1683744639158249, + 0.030172357335686684 + ], + "credit_map": [ + 0.010974445380270481, + 0.007247251458466053, + 0.011153203435242176, + 0.006973334588110447, + 0.0070541370660066605, + 0.00802597589790821, + 0.011890583671629429, + 0.011093328706920147, + 0.019251419231295586, + 0.03431399166584015, + 0.02919725887477398, + 0.03303203731775284, + 0.03550942987203598, + 0.02754131145775318, + 0.07196594029664993, + 0.025642771273851395, + 0.033972907811403275, + 0.022507203742861748, + 0.03452691808342934, + 0.01922478713095188, + 0.02504774183034897, + 0.042449112981557846, + 0.022174229845404625, + 0.03805937618017197, + 0.01132984273135662, + 0.009517348371446133, + 0.00870105903595686, + 0.0056788078509271145, + 0.005085610784590244, + 0.01119451504200697, + 0.005930603947490454, + 0.009507746435701847, + 0.005116738844662905, + 0.0051774680614471436, + 0.0049139088951051235, + 0.006210849620401859, + 0.0040796599350869656, + 0.00769965723156929, + 0.005739362444728613, + 0.009272441267967224, + 0.002861546352505684, + 0.0045239562168717384, + 0.0038066075649112463, + 0.006519949994981289, + 0.004632868338376284, + 0.005779056344181299, + 0.004825014621019363, + 0.005911145359277725, + 0.003101383801549673, + 0.004687842912971973, + 0.004657393787056208, + 0.004801492672413588, + 0.0031979884952306747, + 0.005253756884485483, + 0.003586849896237254, + 0.0056833624839782715, + 0.0025323349982500076, + 0.0019203549018129706, + 0.0024039207492023706, + 0.0033593028783798218, + 0.0014757659519091249, + 0.00408543786033988, + 0.002353537594899535, + 0.005507975351065397, + 0.01595601625740528, + 0.02924123965203762, + 0.013102944940328598, + 0.025012996047735214, + 0.020548509433865547, + 0.025757567957043648, + 0.013912715949118137, + 0.02484249137341976, + 0.0024619654286652803, + 0.0037490574177354574, + 0.003027088474482298, + 0.0060077449306845665, + 0.0023340911138802767, + 0.00560712069272995, + 0.00380493956618011, + 0.0031803548336029053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -0.5762308835983276 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747556384652853 + ], + "opposite_logprob_tokens": [ + -5.006747722625732 + ], + "logprob_reward": -0.006747556384652853, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747556384652853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2685970067977905, + 0.13001877069473267, + 0.05992009863257408, + 0.044042062014341354, + 0.03967607766389847, + 0.0707850232720375, + 0.03327443078160286, + 0.058031536638736725, + 0.1714438796043396, + 0.12421107292175293 + ], + "credit_map": [ + 0.013986180536448956, + 0.020490821450948715, + 0.02508714236319065, + 0.04130227491259575, + 0.030312124639749527, + 0.03240184113383293, + 0.07752367109060287, + 0.027492953464388847, + 0.012423119507730007, + 0.015807917341589928, + 0.01605881191790104, + 0.013237036764621735, + 0.016393784433603287, + 0.013206031173467636, + 0.028839509934186935, + 0.014052564278244972, + 0.005215988960117102, + 0.009525135159492493, + 0.005194345023483038, + 0.008678157813847065, + 0.00809897854924202, + 0.006930839270353317, + 0.00997473020106554, + 0.006301925517618656, + 0.0047686039470136166, + 0.0066560255363583565, + 0.005969940219074488, + 0.004613254219293594, + 0.0038591348566114902, + 0.0068158856593072414, + 0.005010601133108139, + 0.006348614580929279, + 0.003615293186157942, + 0.005076445173472166, + 0.0035451750736683607, + 0.006404371000826359, + 0.0032196149695664644, + 0.00640848046168685, + 0.006309705786406994, + 0.005096993409097195, + 0.004728048108518124, + 0.013191614300012589, + 0.004978136159479618, + 0.012959999032318592, + 0.008616491220891476, + 0.01152219157665968, + 0.007509986404329538, + 0.007278553210198879, + 0.0030429360922425985, + 0.004247024189680815, + 0.0034434893168509007, + 0.006434139329940081, + 0.0031779929995536804, + 0.005530681926757097, + 0.003184970933943987, + 0.0042131938971579075, + 0.005918245296925306, + 0.006158436182886362, + 0.006210282910615206, + 0.012522432953119278, + 0.004776317160576582, + 0.010035879909992218, + 0.005345575511455536, + 0.007064368110150099, + 0.013371935114264488, + 0.02977863699197769, + 0.015349661000072956, + 0.02857639454305172, + 0.016605766490101814, + 0.028603587299585342, + 0.021647995337843895, + 0.017509914934635162, + 0.007388357073068619, + 0.015471535734832287, + 0.00873633474111557, + 0.03619881719350815, + 0.009354307316243649, + 0.021951762959361076, + 0.012835255824029446, + 0.012274692766368389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.006747556384652853 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038435697555542 + ], + "opposite_logprob_tokens": [ + -1.9103844165802002 + ], + "logprob_reward": -0.16038435697555542, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038435697555542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1281421184539795, + 0.2082788497209549, + 0.04112841933965683, + 0.057114772498607635, + 0.0444716140627861, + 0.07819600403308868, + 0.04694177210330963, + 0.03534262627363205, + 0.2522127032279968, + 0.10817106813192368 + ], + "credit_map": [ + 0.007917121052742004, + 0.00944418366998434, + 0.012838734313845634, + 0.020811771973967552, + 0.015645233914256096, + 0.014597790315747261, + 0.035535696893930435, + 0.01135160680860281, + 0.013161237351596355, + 0.029946299269795418, + 0.021485840901732445, + 0.02412009984254837, + 0.029760707169771194, + 0.019553862512111664, + 0.052839700132608414, + 0.017411118373274803, + 0.004662045277655125, + 0.004450115375220776, + 0.005065096076577902, + 0.006402032449841499, + 0.004850615747272968, + 0.005609351675957441, + 0.0050645493902266026, + 0.0050246198661625385, + 0.00784381851553917, + 0.00877232477068901, + 0.005932309199124575, + 0.0061677489429712296, + 0.0049163480289280415, + 0.009778733365237713, + 0.005992244463413954, + 0.00771125266328454, + 0.0033317748457193375, + 0.007768808398395777, + 0.0036595005076378584, + 0.0060242824256420135, + 0.00579946581274271, + 0.006155155599117279, + 0.006156446877866983, + 0.005576186813414097, + 0.006949601694941521, + 0.01288261916488409, + 0.006152492947876453, + 0.01342257484793663, + 0.009702401235699654, + 0.012315759435296059, + 0.00821987446397543, + 0.008550680242478848, + 0.00475072069093585, + 0.005794375669211149, + 0.005076095461845398, + 0.01040053553879261, + 0.0045877667143940926, + 0.00797327421605587, + 0.004093653056770563, + 0.004265353083610535, + 0.0027911432553082705, + 0.004301012959331274, + 0.0039920499548316, + 0.006553545594215393, + 0.002513251965865493, + 0.0056126718409359455, + 0.002882949309423566, + 0.00669600535184145, + 0.019183415919542313, + 0.043660782277584076, + 0.023582691326737404, + 0.04145471379160881, + 0.023917710408568382, + 0.04219182953238487, + 0.03128451108932495, + 0.02693709172308445, + 0.005169855896383524, + 0.011304355226457119, + 0.007480211090296507, + 0.035569991916418076, + 0.007540541235357523, + 0.01928546465933323, + 0.010870423167943954, + 0.010950238443911076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.16038435697555542 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful handheld close-up, shaky and violent, closely follows a violent, downward swinging motion, focusing on the fist of an adult male. Veins bulge, knuckles white from the intense force, skin crisscrossed with tiny scars and rough texture. Illuminated by a cold, hard overhead light from directly above, this fist smashes with thunderous force into a bright red Fuji apple resting on a worn, dark-grained wooden table. The apple's smooth skin reflects a glaring flash of light just before impact. Instantly, with a crisp, popping sound, the red skin and white flesh, mixed with splattering juice, explode in slow motion, flying to every corner of the camera. The fist finally sinks deep into the apple's wreckage, trembling slightly. The entire image is filled with raw, unvarnished power and suppressed rage, imbued with a gritty, tense documentary quality.", + "seed": 30583, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_15/index_3720_seed_30583_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1836216052994132, + "vlm_logprob": -0.1836216052994132, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.1836216052994132, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1562618911266327 + ], + "opposite_logprob_tokens": [ + -2.031261920928955 + ], + "logprob_reward": -0.1562618911266327, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1562618911266327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1889808475971222, + 0.10705044120550156, + 0.07155826687812805, + 0.1279916912317276, + 0.08112278580665588, + 0.10347311943769455, + 0.06392739713191986, + 0.03784596920013428, + 0.11145774275064468, + 0.10659178346395493 + ], + "credit_map": [ + 0.023343127220869064, + 0.030704958364367485, + 0.023828929290175438, + 0.02464747428894043, + 0.02471785433590412, + 0.022464152425527573, + 0.02370494417846203, + 0.0155694130808115, + 0.010747240856289864, + 0.015642134472727776, + 0.007683543488383293, + 0.009437051601707935, + 0.017065467312932014, + 0.014297224581241608, + 0.016711832955479622, + 0.015465940348803997, + 0.01111706905066967, + 0.009723285213112831, + 0.005757672246545553, + 0.008239679038524628, + 0.006717550568282604, + 0.011638333089649677, + 0.008578386157751083, + 0.009786288253962994, + 0.014375883154571056, + 0.017629258334636688, + 0.016050955280661583, + 0.012138793244957924, + 0.01962212286889553, + 0.01731131039559841, + 0.015051636844873428, + 0.015811733901500702, + 0.010653479024767876, + 0.010973628610372543, + 0.008128536865115166, + 0.009256315417587757, + 0.005041260272264481, + 0.014092051424086094, + 0.009866740554571152, + 0.013110769912600517, + 0.011276045814156532, + 0.013661791570484638, + 0.01021951250731945, + 0.008862173184752464, + 0.01230523269623518, + 0.013998105190694332, + 0.01130397617816925, + 0.021846279501914978, + 0.008477984927594662, + 0.0061660814099013805, + 0.0061896187253296375, + 0.006103000603616238, + 0.006856283638626337, + 0.010480445809662342, + 0.010983270592987537, + 0.008670714683830738, + 0.004764288663864136, + 0.004482955206185579, + 0.0034817904233932495, + 0.004005168564617634, + 0.0036569705698639154, + 0.0056584542617201805, + 0.004919643513858318, + 0.006876698229461908, + 0.011971965432167053, + 0.013314928859472275, + 0.011299644596874714, + 0.008496707305312157, + 0.009210851043462753, + 0.02152496948838234, + 0.009245396591722965, + 0.02639327198266983, + 0.010712900198996067, + 0.01109716109931469, + 0.010009913705289364, + 0.011265791021287441, + 0.017632510513067245, + 0.01910082809627056, + 0.01270932424813509, + 0.014063348062336445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.1562618911266327 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011098084971308708 + ], + "opposite_logprob_tokens": [ + -4.5110979080200195 + ], + "logprob_reward": -0.011098084971308708, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011098084971308708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.402523934841156, + 0.03977172449231148, + 0.10622268170118332, + 0.08435200154781342, + 0.04098515585064888, + 0.08468974381685257, + 0.04221491888165474, + 0.028122859075665474, + 0.07464618980884552, + 0.09647077322006226 + ], + "credit_map": [ + 0.06599237024784088, + 0.04128601402044296, + 0.11514461040496826, + 0.044700637459754944, + 0.03676793724298477, + 0.027340533211827278, + 0.04206633195281029, + 0.02922550030052662, + 0.005121150519698858, + 0.0038955535273998976, + 0.0042410860769450665, + 0.0034330147318542004, + 0.004488789010792971, + 0.006437008269131184, + 0.003602290991693735, + 0.00855283159762621, + 0.014221053570508957, + 0.013708924874663353, + 0.007385085802525282, + 0.015295892022550106, + 0.012322950176894665, + 0.01657048612833023, + 0.01514015719294548, + 0.011578132398426533, + 0.00917038694024086, + 0.014937251806259155, + 0.007934801280498505, + 0.009055795148015022, + 0.011563541367650032, + 0.010633660480380058, + 0.011278124526143074, + 0.009778442792594433, + 0.005330932326614857, + 0.004774537403136492, + 0.0035826517269015312, + 0.004419539123773575, + 0.0034591401927173138, + 0.008180728182196617, + 0.004719613119959831, + 0.006518012844026089, + 0.007650578394532204, + 0.011264149099588394, + 0.007679977919906378, + 0.006994277238845825, + 0.00856270082294941, + 0.008575160056352615, + 0.01087341271340847, + 0.023089488968253136, + 0.004931639414280653, + 0.004922440741211176, + 0.004256871994584799, + 0.004095269832760096, + 0.0036750356666743755, + 0.007419882342219353, + 0.005850881338119507, + 0.007062895223498344, + 0.003252218244597316, + 0.0034125770907849073, + 0.002375202951952815, + 0.00334643735550344, + 0.0036429802421480417, + 0.004187998827546835, + 0.002999396761879325, + 0.00490604666993022, + 0.009865155443549156, + 0.009929691441357136, + 0.007360539864748716, + 0.005612593609839678, + 0.005399915389716625, + 0.014604028314352036, + 0.00679754838347435, + 0.015076711773872375, + 0.006149073131382465, + 0.00994205754250288, + 0.007291919086128473, + 0.01417619176208973, + 0.015424991026520729, + 0.013581759296357632, + 0.015332669951021671, + 0.01457210537046194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.011098084971308708 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934375166893005 + ], + "opposite_logprob_tokens": [ + -0.6934375166893005 + ], + "logprob_reward": -0.6934375166893005, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934375166893005, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.32109248638153076, + 0.05307035893201828, + 0.08974242210388184, + 0.10563867539167404, + 0.08165080100297928, + 0.08346223086118698, + 0.10725946724414825, + 0.02403823845088482, + 0.03394857421517372, + 0.10009684413671494 + ], + "credit_map": [ + 0.035259224474430084, + 0.053881626576185226, + 0.03622409328818321, + 0.04709639027714729, + 0.04821333289146423, + 0.03491635620594025, + 0.03870232775807381, + 0.026799092069268227, + 0.006638906430453062, + 0.006030222401022911, + 0.005505429580807686, + 0.005111814476549625, + 0.00629130844026804, + 0.007416971027851105, + 0.006644398905336857, + 0.009431303478777409, + 0.01759561523795128, + 0.012136261910200119, + 0.004598170053213835, + 0.0097495811060071, + 0.009702080860733986, + 0.0148917930200696, + 0.0100138233974576, + 0.011055084876716137, + 0.01911615952849388, + 0.011121539399027824, + 0.015138324350118637, + 0.009379727765917778, + 0.01268831454217434, + 0.01673630252480507, + 0.008747834712266922, + 0.01271047629415989, + 0.015613624826073647, + 0.006770719774067402, + 0.018257644027471542, + 0.00491673406213522, + 0.004814498592168093, + 0.011602813377976418, + 0.005554485600441694, + 0.014120269566774368, + 0.011732663959264755, + 0.011315139941871166, + 0.009214900434017181, + 0.00735684996470809, + 0.009313002228736877, + 0.012138625606894493, + 0.007459545973688364, + 0.014931495301425457, + 0.016829753294587135, + 0.011249219067394733, + 0.010643667541444302, + 0.010628503747284412, + 0.007407932076603174, + 0.021302860230207443, + 0.014231489971280098, + 0.014966027811169624, + 0.003165207337588072, + 0.0032593468204140663, + 0.002285485155880451, + 0.002390501322224736, + 0.0021912853699177504, + 0.0036603733897209167, + 0.002487803576514125, + 0.004598235245794058, + 0.0030244896188378334, + 0.003995999693870544, + 0.003977099899202585, + 0.0033974333200603724, + 0.002567762741819024, + 0.005742785055190325, + 0.004388744477182627, + 0.0068542566150426865, + 0.007925055921077728, + 0.011242106556892395, + 0.006340828724205494, + 0.01694016344845295, + 0.014823907054960728, + 0.014399335719645023, + 0.017256924882531166, + 0.011168518103659153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.6934375166893005 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649622090160847 + ], + "opposite_logprob_tokens": [ + -4.758649826049805 + ], + "logprob_reward": -0.008649622090160847, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649622090160847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.359607994556427, + 0.02851863205432892, + 0.056514859199523926, + 0.029949858784675598, + 0.025602465495467186, + 0.027016496285796165, + 0.04349992051720619, + 0.03597905486822128, + 0.11964878439903259, + 0.27366194128990173 + ], + "credit_map": [ + 0.04364342242479324, + 0.06742357462644577, + 0.05081889033317566, + 0.04503912478685379, + 0.050129491835832596, + 0.03471723571419716, + 0.03507772833108902, + 0.032758526504039764, + 0.004471792373806238, + 0.003983261063694954, + 0.0024326592683792114, + 0.002717232331633568, + 0.003280090168118477, + 0.003869659500196576, + 0.0035555113572627306, + 0.004208426456898451, + 0.009101569652557373, + 0.008789634332060814, + 0.00302192778326571, + 0.005942119285464287, + 0.0068975891917943954, + 0.008465401828289032, + 0.00853758491575718, + 0.005759031977504492, + 0.0036675778683274984, + 0.0040109953843057156, + 0.0037313306238502264, + 0.002702184021472931, + 0.003169289091601968, + 0.004233784042298794, + 0.0034066480584442616, + 0.005028050392866135, + 0.003071847837418318, + 0.0032485087867826223, + 0.003158619161695242, + 0.0022547757253050804, + 0.002939472906291485, + 0.003628963604569435, + 0.0035054704640060663, + 0.00379480654373765, + 0.0021572113037109375, + 0.002998149022459984, + 0.002246930729597807, + 0.002842947607859969, + 0.0035272669047117233, + 0.0034804653842002153, + 0.004054642748087645, + 0.005708884447813034, + 0.0073915873654186726, + 0.004051891155540943, + 0.004174078814685345, + 0.003331446787342429, + 0.002870338736101985, + 0.009898406453430653, + 0.004647322464734316, + 0.00713485199958086, + 0.0038223571609705687, + 0.004327480681240559, + 0.0034302622079849243, + 0.0032326693180948496, + 0.003968243952840567, + 0.005746607203036547, + 0.004176084417849779, + 0.007275347597897053, + 0.009751854464411736, + 0.01374373584985733, + 0.01519217249006033, + 0.00814809836447239, + 0.01165282353758812, + 0.02416779100894928, + 0.008333105593919754, + 0.02865920402109623, + 0.024706820026040077, + 0.030575573444366455, + 0.020065685734152794, + 0.04839595407247543, + 0.03801645711064339, + 0.040525343269109726, + 0.04131109640002251, + 0.030065005645155907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.008649622090160847 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866091161966324 + ], + "opposite_logprob_tokens": [ + -3.04866099357605 + ], + "logprob_reward": -0.04866091161966324, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866091161966324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3112909197807312, + 0.04684853181242943, + 0.05420386791229248, + 0.048664942383766174, + 0.02332466095685959, + 0.04401034861803055, + 0.06026225537061691, + 0.06004076451063156, + 0.05790365859866142, + 0.29345008730888367 + ], + "credit_map": [ + 0.03332816809415817, + 0.04126511141657829, + 0.044725965708494186, + 0.0486607626080513, + 0.033123116940259933, + 0.041811082512140274, + 0.04070786386728287, + 0.027668839320540428, + 0.005783614236861467, + 0.006418319419026375, + 0.003926205448806286, + 0.005509800277650356, + 0.005548158194869757, + 0.006380445323884487, + 0.006987458094954491, + 0.006294528488069773, + 0.009221209213137627, + 0.007768543437123299, + 0.004010853357613087, + 0.0052768997848033905, + 0.004072501789778471, + 0.008842279203236103, + 0.007451245561242104, + 0.0075603327713906765, + 0.007151573896408081, + 0.006611895281821489, + 0.0042351046577095985, + 0.005661989562213421, + 0.006254030857235193, + 0.007248272653669119, + 0.0057451664470136166, + 0.005756906699389219, + 0.003864072496071458, + 0.0026802325155586004, + 0.0029119369573891163, + 0.001890263520181179, + 0.0018642665818333626, + 0.003430884098634124, + 0.0031144709791988134, + 0.0035685335751622915, + 0.003585630562156439, + 0.0062230974435806274, + 0.005196014419198036, + 0.005524804815649986, + 0.006653911434113979, + 0.0058104731142520905, + 0.005572476889938116, + 0.005443939007818699, + 0.009540975093841553, + 0.0055699837394058704, + 0.005398263223469257, + 0.0036471192725002766, + 0.0036385103594511747, + 0.015161052346229553, + 0.004100395366549492, + 0.013205953873693943, + 0.006115963216871023, + 0.007099139038473368, + 0.005749508272856474, + 0.005434661637991667, + 0.00654417322948575, + 0.009657135233283043, + 0.006891491822898388, + 0.012548692524433136, + 0.003916599787771702, + 0.006648027803748846, + 0.006298096384853125, + 0.0042122830636799335, + 0.008176912553608418, + 0.0102952616289258, + 0.004996137227863073, + 0.013360334560275078, + 0.025837335735559464, + 0.024052903056144714, + 0.01916952058672905, + 0.06268320977687836, + 0.029290582984685898, + 0.04443725198507309, + 0.051933981478214264, + 0.036045268177986145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.04866091161966324 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful handheld close-up, shaky and violent, closely follows a violent, downward swinging motion, focusing on the fist of an adult male. Veins bulge, knuckles white from the intense force, skin crisscrossed with tiny scars and rough texture. Illuminated by a cold, hard overhead light from directly above, this fist smashes with thunderous force into a bright red Fuji apple resting on a worn, dark-grained wooden table. The apple's smooth skin reflects a glaring flash of light just before impact. Instantly, with a crisp, popping sound, the red skin and white flesh, mixed with splattering juice, explode in slow motion, flying to every corner of the camera. The fist finally sinks deep into the apple's wreckage, trembling slightly. The entire image is filled with raw, unvarnished power and suppressed rage, imbued with a gritty, tense documentary quality.", + "seed": 30584, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_15/index_3720_seed_30584_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18882707725279033, + "vlm_logprob": -0.18882707725279033, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.18882707725279033, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09258092194795609 + ], + "opposite_logprob_tokens": [ + -2.467581033706665 + ], + "logprob_reward": -0.09258092194795609, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09258092194795609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13002461194992065, + 0.04766368493437767, + 0.10389657318592072, + 0.09014416486024857, + 0.17152081429958344, + 0.16905170679092407, + 0.08978487551212311, + 0.05197273939847946, + 0.08779225498437881, + 0.0581485889852047 + ], + "credit_map": [ + 0.019431212916970253, + 0.024614963680505753, + 0.02720816060900688, + 0.01368829794228077, + 0.010909023694694042, + 0.011105152778327465, + 0.011173063889145851, + 0.011894745752215385, + 0.0057323700748384, + 0.005252311937510967, + 0.004666138906031847, + 0.005982532631605864, + 0.005852441303431988, + 0.006414391566067934, + 0.007691808044910431, + 0.006071688607335091, + 0.011199956759810448, + 0.01568315178155899, + 0.012052038684487343, + 0.021599074825644493, + 0.00999645795673132, + 0.012160382233560085, + 0.011439538560807705, + 0.00976597424596548, + 0.014325465075671673, + 0.008293740451335907, + 0.01543671265244484, + 0.010390842333436012, + 0.011350064538419247, + 0.007713655941188335, + 0.0126410610973835, + 0.009992632083594799, + 0.024144696071743965, + 0.016904717311263084, + 0.02086501009762287, + 0.023182960227131844, + 0.0205007903277874, + 0.016076454892754555, + 0.027185779064893723, + 0.022660423070192337, + 0.021257175132632256, + 0.0154075026512146, + 0.021043915301561356, + 0.027585821226239204, + 0.01735113374888897, + 0.018147757276892662, + 0.018321258947253227, + 0.029937151819467545, + 0.006441901903599501, + 0.004283216316252947, + 0.03675024211406708, + 0.008073339238762856, + 0.002671091351658106, + 0.01429254561662674, + 0.004072571638971567, + 0.013199970126152039, + 0.004209726117551327, + 0.004294320475310087, + 0.010195083916187286, + 0.005042544566094875, + 0.005713364575058222, + 0.005129138007760048, + 0.0045865923166275024, + 0.012801971286535263, + 0.005525880493223667, + 0.00983805488795042, + 0.00858794990926981, + 0.02346147410571575, + 0.008260026574134827, + 0.011649268679320812, + 0.012450100854039192, + 0.008019506931304932, + 0.012919778935611248, + 0.004947951529175043, + 0.005599063355475664, + 0.003479529172182083, + 0.006109198555350304, + 0.005588919389992952, + 0.0041220844723284245, + 0.015382065437734127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.09258092194795609 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009809975512325764 + ], + "opposite_logprob_tokens": [ + -4.634809970855713 + ], + "logprob_reward": -0.009809975512325764, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009809975512325764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11803784966468811, + 0.04807683452963829, + 0.11004119366407394, + 0.11718541383743286, + 0.1610865443944931, + 0.06834056228399277, + 0.15688756108283997, + 0.12343175709247589, + 0.05343804880976677, + 0.043474163860082626 + ], + "credit_map": [ + 0.01421404629945755, + 0.026384519413113594, + 0.021435149013996124, + 0.01225852221250534, + 0.013712702319025993, + 0.013362506404519081, + 0.0073750875890254974, + 0.0092953285202384, + 0.005201370920985937, + 0.003506665350869298, + 0.004900528583675623, + 0.006194233428686857, + 0.007689482532441616, + 0.006418972741812468, + 0.007974252104759216, + 0.006191337015479803, + 0.010611972771584988, + 0.018161317333579063, + 0.012777869589626789, + 0.016559213399887085, + 0.010315944440662861, + 0.013851311057806015, + 0.014684472233057022, + 0.013079101219773293, + 0.019783610478043556, + 0.006831613834947348, + 0.024073978886008263, + 0.014290611259639263, + 0.012286894954741001, + 0.011722068302333355, + 0.014989272691309452, + 0.013207373209297657, + 0.02327314205467701, + 0.01400679163634777, + 0.022770484909415245, + 0.02027813158929348, + 0.017293009907007217, + 0.011217898689210415, + 0.030637551099061966, + 0.02160954847931862, + 0.007575419265776873, + 0.006993959657847881, + 0.012749536894261837, + 0.011553486809134483, + 0.005913270637392998, + 0.006415645591914654, + 0.00827226135879755, + 0.008866994641721249, + 0.013017585501074791, + 0.01482460182160139, + 0.04322335496544838, + 0.021864386275410652, + 0.011734195053577423, + 0.015475722029805183, + 0.014777830801904202, + 0.021969903260469437, + 0.008699079975485802, + 0.0063044107519090176, + 0.028652114793658257, + 0.008574564941227436, + 0.006737507414072752, + 0.015317046083509922, + 0.00791835319250822, + 0.04122868925333023, + 0.005745166912674904, + 0.007591544650495052, + 0.004488806705921888, + 0.010001038201153278, + 0.005846957676112652, + 0.009383094497025013, + 0.006150045897811651, + 0.004231399390846491, + 0.006529452279210091, + 0.00466918433085084, + 0.00491259153932333, + 0.003666248405352235, + 0.004314026329666376, + 0.009079081937670708, + 0.005417317617684603, + 0.0048862663097679615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.009809975512325764 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8262020349502563 + ], + "opposite_logprob_tokens": [ + -0.5762020349502563 + ], + "logprob_reward": -0.8262020349502563, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8262020349502563, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19294874370098114, + 0.042910974472761154, + 0.1137123554944992, + 0.0904616266489029, + 0.09474797546863556, + 0.13218775391578674, + 0.06270763278007507, + 0.15570899844169617, + 0.04103100299835205, + 0.07358293235301971 + ], + "credit_map": [ + 0.02922808565199375, + 0.03800494968891144, + 0.037746578454971313, + 0.015338687226176262, + 0.01957685500383377, + 0.02041838876903057, + 0.013221925124526024, + 0.019413262605667114, + 0.00569136580452323, + 0.004489721264690161, + 0.0038311455864459276, + 0.003958620596677065, + 0.00711450120434165, + 0.0070405504666268826, + 0.0032855200115591288, + 0.007499546743929386, + 0.0199732705950737, + 0.012080333195626736, + 0.01644820161163807, + 0.016166038811206818, + 0.010223343968391418, + 0.014438290148973465, + 0.010424261912703514, + 0.013958615250885487, + 0.012099089100956917, + 0.009082343429327011, + 0.014244560152292252, + 0.012383404187858105, + 0.0126211391761899, + 0.007312234491109848, + 0.013592839241027832, + 0.00912601687014103, + 0.012686141766607761, + 0.008308276534080505, + 0.01323756668716669, + 0.012864340096712112, + 0.009946659207344055, + 0.00714101130142808, + 0.018825598061084747, + 0.011738386936485767, + 0.02375030890107155, + 0.009515034966170788, + 0.03271693363785744, + 0.015117228031158447, + 0.010581457056105137, + 0.014525583945214748, + 0.011258724145591259, + 0.014722494408488274, + 0.006023510824888945, + 0.003254845505580306, + 0.01971372216939926, + 0.005774888209998608, + 0.0027381309773772955, + 0.005959386471658945, + 0.004447008017450571, + 0.014796138741075993, + 0.012912550941109657, + 0.006463131867349148, + 0.042459726333618164, + 0.008843490853905678, + 0.00549448374658823, + 0.012507706880569458, + 0.007408319041132927, + 0.059619590640068054, + 0.0035884121898561716, + 0.006210861727595329, + 0.005987283308058977, + 0.007014099042862654, + 0.0032535234931856394, + 0.004699553828686476, + 0.0047059981152415276, + 0.005571273621171713, + 0.011786630377173424, + 0.005825545638799667, + 0.011205513961613178, + 0.005214134696871042, + 0.004420089069753885, + 0.010246398858726025, + 0.004727818537503481, + 0.020156802609562874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.8262020349502563 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013412775006145239 + ], + "opposite_logprob_tokens": [ + -6.626341342926025 + ], + "logprob_reward": -0.0013412775006145239, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013412775006145239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08715379238128662, + 0.03157106041908264, + 0.1248282641172409, + 0.16043415665626526, + 0.14962095022201538, + 0.15991538763046265, + 0.08397369086742401, + 0.05664359778165817, + 0.06803369522094727, + 0.07782542705535889 + ], + "credit_map": [ + 0.01055130548775196, + 0.01405581645667553, + 0.012351459823548794, + 0.010718482546508312, + 0.007588195614516735, + 0.01283698994666338, + 0.00826987810432911, + 0.010781659744679928, + 0.0032952027395367622, + 0.004000952001661062, + 0.003714886726811528, + 0.0032964760903269053, + 0.004489596467465162, + 0.004664959851652384, + 0.0042406790889799595, + 0.0038683076854795218, + 0.012186857871711254, + 0.017508381977677345, + 0.013320247642695904, + 0.026927659288048744, + 0.012782844714820385, + 0.016548605635762215, + 0.015097315423190594, + 0.010456347838044167, + 0.02784745953977108, + 0.012723262421786785, + 0.029983270913362503, + 0.01904793456196785, + 0.01570998877286911, + 0.01638430543243885, + 0.01971064880490303, + 0.019027281552553177, + 0.020421212539076805, + 0.011678984388709068, + 0.0226394422352314, + 0.01864209957420826, + 0.013568209484219551, + 0.0114853261038661, + 0.02819543145596981, + 0.02299022115767002, + 0.01995260827243328, + 0.014492589049041271, + 0.020003745332360268, + 0.027317343279719353, + 0.018519751727581024, + 0.014938347041606903, + 0.0188944973051548, + 0.025796495378017426, + 0.007346899248659611, + 0.0048799957148730755, + 0.029131179675459862, + 0.007856315933167934, + 0.0030604659114032984, + 0.013413075357675552, + 0.004745532758533955, + 0.013540220446884632, + 0.004590912722051144, + 0.004723814316093922, + 0.010032995603978634, + 0.005789190996438265, + 0.0042134434916079044, + 0.00530372466892004, + 0.005179653409868479, + 0.0168098583817482, + 0.005847057793289423, + 0.010781346820294857, + 0.007071260828524828, + 0.012086167931556702, + 0.007588779088109732, + 0.010174194350838661, + 0.008853144943714142, + 0.005631736945360899, + 0.00976136140525341, + 0.0068751731887459755, + 0.011254096403717995, + 0.006011963356286287, + 0.009294337593019009, + 0.010000001639127731, + 0.012202361598610878, + 0.012426127679646015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.0013412775006145239 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201176352798939 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014201176352798939, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201176352798939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09397458285093307, + 0.04161111265420914, + 0.0980602502822876, + 0.04407525435090065, + 0.10131369531154633, + 0.05319622531533241, + 0.07603186368942261, + 0.1723162978887558, + 0.23890240490436554, + 0.08051829040050507 + ], + "credit_map": [ + 0.014544490724802017, + 0.02127622812986374, + 0.01419049222022295, + 0.010222160257399082, + 0.006351425312459469, + 0.011332889087498188, + 0.006694308947771788, + 0.00936258863657713, + 0.004031222779303789, + 0.005526597145944834, + 0.004200985189527273, + 0.00544706080108881, + 0.006116018630564213, + 0.004705238621681929, + 0.006941826082766056, + 0.004642163868993521, + 0.00827750563621521, + 0.0094603281468153, + 0.009938626550137997, + 0.017272239550948143, + 0.012286767363548279, + 0.011897200718522072, + 0.02039921097457409, + 0.008528371341526508, + 0.0064009265042841434, + 0.004289174918085337, + 0.0070848725736141205, + 0.00510314479470253, + 0.0051736366003751755, + 0.00431910390034318, + 0.006667368113994598, + 0.005037026014178991, + 0.014709205366671085, + 0.00996225606650114, + 0.019414640963077545, + 0.014409683644771576, + 0.009412671439349651, + 0.008714425377547741, + 0.01527347695082426, + 0.009417332708835602, + 0.009635052643716335, + 0.004429988097399473, + 0.014427853748202324, + 0.004764039069414139, + 0.005298362113535404, + 0.005780831445008516, + 0.004066066350787878, + 0.004794029984623194, + 0.01022461149841547, + 0.007915964350104332, + 0.01904647797346115, + 0.008791184052824974, + 0.005916169378906488, + 0.006758113857358694, + 0.0074579594656825066, + 0.00992138683795929, + 0.014077413827180862, + 0.012423510663211346, + 0.04485528916120529, + 0.011131959035992622, + 0.012149382382631302, + 0.017532039433717728, + 0.013222596608102322, + 0.04692411050200462, + 0.02495981939136982, + 0.03918852284550667, + 0.024659190326929092, + 0.03480391204357147, + 0.02321704849600792, + 0.052762262523174286, + 0.0216757133603096, + 0.01763594150543213, + 0.013028486631810665, + 0.00975753739476204, + 0.008629907853901386, + 0.005662238225340843, + 0.008422121405601501, + 0.01259530894458294, + 0.010445765219628811, + 0.01197692658752203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.014201176352798939 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "powerful handheld close-up, shaky and violent, closely follows a violent, downward swinging motion, focusing on the fist of an adult male. Veins bulge, knuckles white from the intense force, skin crisscrossed with tiny scars and rough texture. Illuminated by a cold, hard overhead light from directly above, this fist smashes with thunderous force into a bright red Fuji apple resting on a worn, dark-grained wooden table. The apple's smooth skin reflects a glaring flash of light just before impact. Instantly, with a crisp, popping sound, the red skin and white flesh, mixed with splattering juice, explode in slow motion, flying to every corner of the camera. The fist finally sinks deep into the apple's wreckage, trembling slightly. The entire image is filled with raw, unvarnished power and suppressed rage, imbued with a gritty, tense documentary quality.", + "seed": 30585, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_15/index_3720_seed_30585_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3628659171052277, + "vlm_logprob": -0.3628659171052277, + "vlm_yes_no_margin": 2.799999976158142, + "vlm_token_credit": -0.3628659171052277, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05960254743695259 + ], + "opposite_logprob_tokens": [ + -3.0596024990081787 + ], + "logprob_reward": -0.05960254743695259, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05960254743695259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05330390855669975, + 0.09145446866750717, + 0.06445875018835068, + 0.06282197684049606, + 0.21835140883922577, + 0.23142540454864502, + 0.025492994114756584, + 0.1216035932302475, + 0.07453319430351257, + 0.056554291397333145 + ], + "credit_map": [ + 0.005340872332453728, + 0.0051537444815039635, + 0.006368509028106928, + 0.006310930009931326, + 0.007283863611519337, + 0.008216775953769684, + 0.007082871161401272, + 0.007546342443674803, + 0.006527971476316452, + 0.012008673511445522, + 0.012667682953178883, + 0.012809379026293755, + 0.008597264997661114, + 0.01805906929075718, + 0.011978658847510815, + 0.008805765770375729, + 0.005446658469736576, + 0.004814008716493845, + 0.011860464699566364, + 0.005314689129590988, + 0.00442844582721591, + 0.017404528334736824, + 0.0063142189756035805, + 0.008875738829374313, + 0.004262913949787617, + 0.004113912116736174, + 0.01376369595527649, + 0.004349664319306612, + 0.0060968296602368355, + 0.013010982424020767, + 0.007069157902151346, + 0.010154818184673786, + 0.015609881840646267, + 0.030039699748158455, + 0.02092226780951023, + 0.020260905846953392, + 0.04459679499268532, + 0.03770139068365097, + 0.025702429935336113, + 0.023518046364188194, + 0.016238702461123466, + 0.008165035396814346, + 0.0907176062464714, + 0.014855056069791317, + 0.016364686191082, + 0.05261984094977379, + 0.01338062435388565, + 0.019083844497799873, + 0.002584004309028387, + 0.002651605987921357, + 0.003879969008266926, + 0.00269558047875762, + 0.004799256566911936, + 0.0029005794785916805, + 0.00346953677944839, + 0.002512461505830288, + 0.01081234123557806, + 0.022872919216752052, + 0.02461804635822773, + 0.011059627868235111, + 0.014854240231215954, + 0.010175143368542194, + 0.012440052814781666, + 0.01477122213691473, + 0.005221256520599127, + 0.006742879748344421, + 0.020431388169527054, + 0.004349221475422382, + 0.010964041575789452, + 0.016093023121356964, + 0.004907755181193352, + 0.0058236317709088326, + 0.006416981108486652, + 0.004539387300610542, + 0.007287107408046722, + 0.009348518215119839, + 0.004277711734175682, + 0.004580242093652487, + 0.01125533040612936, + 0.008849012665450573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.05960254743695259 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007651663385331631 + ], + "opposite_logprob_tokens": [ + -4.8826518058776855 + ], + "logprob_reward": -0.007651663385331631, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007651663385331631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04882590472698212, + 0.056734781712293625, + 0.058451324701309204, + 0.039889805018901825, + 0.14720281958580017, + 0.407856285572052, + 0.018246576189994812, + 0.1436871886253357, + 0.024986490607261658, + 0.054118845611810684 + ], + "credit_map": [ + 0.004196952562779188, + 0.0037757698446512222, + 0.0065535991452634335, + 0.007226543966680765, + 0.006978555582463741, + 0.007702449336647987, + 0.007251325994729996, + 0.005140705034136772, + 0.006805917248129845, + 0.006692747585475445, + 0.006723185069859028, + 0.007624165154993534, + 0.006629321724176407, + 0.00895078107714653, + 0.007267922628670931, + 0.006040739361196756, + 0.004934295546263456, + 0.006208759266883135, + 0.009236998856067657, + 0.006479615345597267, + 0.005448130890727043, + 0.013105847872793674, + 0.0062333764508366585, + 0.006804297678172588, + 0.003223167033866048, + 0.0031284212600439787, + 0.0046319118700921535, + 0.004411724396049976, + 0.004813543986529112, + 0.004253079183399677, + 0.009911217726767063, + 0.005516736768186092, + 0.010012511163949966, + 0.017235126346349716, + 0.013943911530077457, + 0.014974305406212807, + 0.027191326022148132, + 0.02478080615401268, + 0.01860710233449936, + 0.02045772224664688, + 0.030654054135084152, + 0.011662069708108902, + 0.16793763637542725, + 0.026916103437542915, + 0.01730004884302616, + 0.10653211921453476, + 0.01772260293364525, + 0.029131662100553513, + 0.0013286385219544172, + 0.001525813015177846, + 0.0036754633765667677, + 0.0018093212274834514, + 0.0033740014769136906, + 0.001779661513864994, + 0.002429646672680974, + 0.0023240295704454184, + 0.013463250361382961, + 0.02612079493701458, + 0.029298720881342888, + 0.012324769049882889, + 0.01854102313518524, + 0.011944550089538097, + 0.014644354581832886, + 0.01734972931444645, + 0.0023729477543383837, + 0.0019355702679604292, + 0.006548664532601833, + 0.0017498460365459323, + 0.002744074445217848, + 0.005231560207903385, + 0.0017342072678729892, + 0.0026696212589740753, + 0.006973558571189642, + 0.006069612223654985, + 0.012651484459638596, + 0.006453349255025387, + 0.003602725686505437, + 0.004891055170446634, + 0.00807823147624731, + 0.005398828070610762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.007651663385331631 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6006044149398804 + ], + "opposite_logprob_tokens": [ + -0.22560442984104156 + ], + "logprob_reward": -1.6006044149398804, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6006044149398804, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16104720532894135, + 0.0382668599486351, + 0.2101609706878662, + 0.04390594735741615, + 0.0617486946284771, + 0.08129661530256271, + 0.05948363617062569, + 0.17689931392669678, + 0.1258225291967392, + 0.041368208825588226 + ], + "credit_map": [ + 0.023082653060555458, + 0.01489382516592741, + 0.02429238334298134, + 0.01822080835700035, + 0.019297434017062187, + 0.024802230298519135, + 0.01920776627957821, + 0.017250118777155876, + 0.004404800944030285, + 0.0035379761829972267, + 0.004831281490623951, + 0.004771224223077297, + 0.003832746995612979, + 0.0056167542934417725, + 0.006206036079674959, + 0.0050660427659749985, + 0.027929997071623802, + 0.020433202385902405, + 0.03890778124332428, + 0.021680543199181557, + 0.017978714779019356, + 0.03859567642211914, + 0.019231490790843964, + 0.025403562933206558, + 0.0036710763815790415, + 0.005089522805064917, + 0.00797101017087698, + 0.0029070680029690266, + 0.003816972952336073, + 0.0078029148280620575, + 0.004826229996979237, + 0.00782114639878273, + 0.003979015629738569, + 0.006586072500795126, + 0.010593931190669537, + 0.006886529270559549, + 0.00682600773870945, + 0.011358494870364666, + 0.007587980479001999, + 0.007930662482976913, + 0.005403605289757252, + 0.006887136958539486, + 0.023608334362506866, + 0.005926799029111862, + 0.008146638050675392, + 0.01757071726024151, + 0.007914452813565731, + 0.005838933400809765, + 0.0060395062901079655, + 0.005052854306995869, + 0.00833186972886324, + 0.011168655008077621, + 0.006720548961311579, + 0.006300502922385931, + 0.007974858395755291, + 0.007894841022789478, + 0.014983552508056164, + 0.031356487423181534, + 0.041414350271224976, + 0.014800534583628178, + 0.022100722417235374, + 0.015222704969346523, + 0.016590874642133713, + 0.02043008990585804, + 0.02308717742562294, + 0.011988024227321148, + 0.02156377211213112, + 0.014702290296554565, + 0.012785135768353939, + 0.008303295820951462, + 0.013643121346831322, + 0.019749710336327553, + 0.005824859719723463, + 0.0033384161069989204, + 0.006987713743001223, + 0.006175104528665543, + 0.0032125371508300304, + 0.004015864338725805, + 0.006881628185510635, + 0.004932087380439043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -1.6006044149398804 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036283391527831554 + ], + "opposite_logprob_tokens": [ + -5.628628253936768 + ], + "logprob_reward": -0.0036283391527831554, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036283391527831554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07272326946258545, + 0.031863823533058167, + 0.07725121080875397, + 0.08150152117013931, + 0.31380724906921387, + 0.05321655422449112, + 0.030232589691877365, + 0.21685302257537842, + 0.0581047460436821, + 0.06444601714611053 + ], + "credit_map": [ + 0.00963641982525587, + 0.005026240833103657, + 0.011109431274235249, + 0.008216663263738155, + 0.008918014355003834, + 0.011995193548500538, + 0.009946148842573166, + 0.007875164970755577, + 0.0037675544153898954, + 0.004006600007414818, + 0.0044167316518723965, + 0.003858566517010331, + 0.0028150922153145075, + 0.004271780140697956, + 0.004218563437461853, + 0.004508933052420616, + 0.010267130099236965, + 0.005732999183237553, + 0.015224958769977093, + 0.007231465075165033, + 0.006694156676530838, + 0.015701310709118843, + 0.006649277172982693, + 0.009749919176101685, + 0.004563966300338507, + 0.006345110014081001, + 0.008824186399579048, + 0.011868394911289215, + 0.008936068043112755, + 0.009380258619785309, + 0.019420132040977478, + 0.012163400650024414, + 0.02183043211698532, + 0.03586319461464882, + 0.03145064786076546, + 0.03515325114130974, + 0.05702849105000496, + 0.054319124668836594, + 0.037780847400426865, + 0.04038126394152641, + 0.003853654256090522, + 0.006195443216711283, + 0.010902661830186844, + 0.00565383629873395, + 0.006635218393057585, + 0.008613093756139278, + 0.0069717951118946075, + 0.004390852525830269, + 0.0021917768754065037, + 0.0027211506385356188, + 0.005011550150811672, + 0.004506790079176426, + 0.004971626214683056, + 0.0029220411088317633, + 0.004452362656593323, + 0.003455291735008359, + 0.019133420661091805, + 0.04008856415748596, + 0.04767846688628197, + 0.01895563118159771, + 0.026360541582107544, + 0.018780704587697983, + 0.020304080098867416, + 0.025551611557602882, + 0.007642280776053667, + 0.005841675214469433, + 0.010428206995129585, + 0.005244176834821701, + 0.008238842710852623, + 0.006473723333328962, + 0.006250329781323671, + 0.007985512726008892, + 0.007124904077500105, + 0.005584040656685829, + 0.00929717905819416, + 0.009186995215713978, + 0.005072440020740032, + 0.006089744158089161, + 0.013553445227444172, + 0.008537267334759235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.0036283391527831554 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld close-up featuring a violent, shaky camera motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an adult male's fist with visible bulging veins, white knuckles, and tiny scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a cold, hard overhead light directly above the action?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fist smash into a bright red Fuji apple on a worn, dark-grained wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the apple explode in slow motion with splattering juice and white flesh after impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428426206111908 + ], + "opposite_logprob_tokens": [ + -2.0178425312042236 + ], + "logprob_reward": -0.1428426206111908, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1428426206111908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04707242548465729, + 0.05887167900800705, + 0.07604421675205231, + 0.042756807059049606, + 0.2299942821264267, + 0.20109817385673523, + 0.020849131047725677, + 0.1518806368112564, + 0.08564367890357971, + 0.08578898012638092 + ], + "credit_map": [ + 0.0043279025703668594, + 0.00422732625156641, + 0.006014679092913866, + 0.006753319874405861, + 0.006131285801529884, + 0.006744770333170891, + 0.008288250304758549, + 0.004584889393299818, + 0.004032541066408157, + 0.007590014021843672, + 0.007183121982961893, + 0.008772543631494045, + 0.004987893160432577, + 0.011065348982810974, + 0.008839272893965244, + 0.006400939077138901, + 0.007825610227882862, + 0.007119269575923681, + 0.011929922737181187, + 0.007070268504321575, + 0.007180343847721815, + 0.017985081300139427, + 0.0066312020644545555, + 0.010302521288394928, + 0.002248377539217472, + 0.0038180178962647915, + 0.006064355839043856, + 0.0040612476877868176, + 0.005780751351267099, + 0.008114248514175415, + 0.007317650131881237, + 0.005352163687348366, + 0.01632787100970745, + 0.028997493907809258, + 0.020602501928806305, + 0.024398358538746834, + 0.04367147758603096, + 0.03846945986151695, + 0.025939064100384712, + 0.031588051468133926, + 0.015531091019511223, + 0.007262100465595722, + 0.0791928693652153, + 0.012964023277163506, + 0.011282375082373619, + 0.0502750426530838, + 0.011587643064558506, + 0.013003024272620678, + 0.0015443275915458798, + 0.0010701329447329044, + 0.00516852829605341, + 0.0023932186886668205, + 0.0028746852185577154, + 0.0020652085077017546, + 0.0027197538875043392, + 0.003013274632394314, + 0.01306353323161602, + 0.028537282720208168, + 0.03294873610138893, + 0.013242308050394058, + 0.018996505066752434, + 0.012800025753676891, + 0.014263549819588661, + 0.018028683960437775, + 0.006351330783218145, + 0.007905666716396809, + 0.01983637362718582, + 0.006444284226745367, + 0.014832776039838791, + 0.012371117249131203, + 0.008750436827540398, + 0.009151700884103775, + 0.009152578189969063, + 0.0071404860354959965, + 0.012657363899052143, + 0.012413532473146915, + 0.006697145290672779, + 0.00871550664305687, + 0.01815427467226982, + 0.010858092457056046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.1428426206111908 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_2/index_tensor([8708, 8708])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_2/index_tensor([8708, 8708])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5bef8aeea888ad89188133ddaa484ad1caa92882 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_2/index_tensor([8708, 8708])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An unsteady, handheld medium shot, its slight tremor heightening the emotional tension, captures a moment of melancholy as a young woman stands quietly at a shimmering railroad crossing amidst a swirling, heavy snowfall. Her long, dark brown hair, a few strands wet from the snow, plastered to her cheeks. A hint of confusion and sorrow lingers in her clear eyes, her cheeks slightly flushed by the chill. She wears a thick, crimson wool coat, a cream-colored wool scarf tightly wrapped around her neck. Each gentle breath condenses into a distinct white mist in the icy air, then slowly dissipates into the snow. Behind her, at the railroad crossing, the red warning lights flicker rhythmically, casting intermittent halos of red light on her and the surrounding tracks, now covered in fresh snow. The entire scene presents the delicate light and shadow and extremely high color saturation of an animated film. The deep red of the girl's coat and the pure white snow scene form a strong visual contrast. Under the soft diffuse light of the cloudy day, they together create an atmosphere that is both beautiful and lonely, full of a sense of story and a hint of imperceptible expectation.", + "seed": 14119, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_2/index_8708_seed_14119_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.26934723034501074, + "vlm_logprob": -0.26934723034501074, + "vlm_yes_no_margin": 1.649999988079071, + "vlm_token_credit": -0.26934723034501074, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3970450460910797 + ], + "opposite_logprob_tokens": [ + -1.1470450162887573 + ], + "logprob_reward": -0.3970450460910797, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3970450460910797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09969556331634521, + 0.05191270634531975, + 0.038133956491947174, + 0.16129058599472046, + 0.06162307411432266, + 0.050913695245981216, + 0.032949164509773254, + 0.028408611193299294, + 0.40433254837989807, + 0.07074006646871567 + ], + "credit_map": [ + 0.0152545515447855, + 0.01594032719731331, + 0.010916532017290592, + 0.005885193590074778, + 0.01370206382125616, + 0.02058764547109604, + 0.008362499997019768, + 0.009046748280525208, + 0.008509038016200066, + 0.00685919402167201, + 0.005377577152103186, + 0.0034810476936399937, + 0.006968526169657707, + 0.008800522424280643, + 0.006342198234051466, + 0.005574607290327549, + 0.006885575596243143, + 0.0025361001025885344, + 0.004910000134259462, + 0.002403193386271596, + 0.004310766700655222, + 0.006164418999105692, + 0.004987609107047319, + 0.005936294794082642, + 0.027110418304800987, + 0.05359421670436859, + 0.007831050083041191, + 0.00916323997080326, + 0.017471833154559135, + 0.027667922899127007, + 0.011752044782042503, + 0.006699851714074612, + 0.011797080747783184, + 0.007831831462681293, + 0.006108864210546017, + 0.0032812131103128195, + 0.009142073802649975, + 0.011844259686768055, + 0.006291314028203487, + 0.005326439626514912, + 0.008207570761442184, + 0.008105124346911907, + 0.004937135614454746, + 0.0034083460923284292, + 0.006805569399148226, + 0.0065002418123185635, + 0.005587182007730007, + 0.007362524978816509, + 0.00486691202968359, + 0.005343467462807894, + 0.004045585636049509, + 0.0020140951965004206, + 0.0031659742817282677, + 0.0052909948863089085, + 0.002764848293736577, + 0.005457286257296801, + 0.004485785961151123, + 0.0034994976595044136, + 0.0033419348765164614, + 0.0015705692349001765, + 0.0035447084810584784, + 0.006633856799453497, + 0.002173342974856496, + 0.003158918581902981, + 0.07111671566963196, + 0.028792131692171097, + 0.0668623223900795, + 0.03953196480870247, + 0.03248970955610275, + 0.04783780872821808, + 0.03467574715614319, + 0.08302613347768784, + 0.02075880579650402, + 0.00561553006991744, + 0.01417670026421547, + 0.0035383892245590687, + 0.007250043097883463, + 0.008584070950746536, + 0.004879784304648638, + 0.00593674136325717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.3970450460910797 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42889127135276794 + ], + "opposite_logprob_tokens": [ + -1.0538913011550903 + ], + "logprob_reward": -0.42889127135276794, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42889127135276794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20368549227714539, + 0.06876248866319656, + 0.05137225240468979, + 0.02816983498632908, + 0.04389680549502373, + 0.04511420056223869, + 0.043816473335027695, + 0.20527997612953186, + 0.28885748982429504, + 0.02104499563574791 + ], + "credit_map": [ + 0.03351971134543419, + 0.01563413254916668, + 0.027528220787644386, + 0.010255346074700356, + 0.020314592868089676, + 0.049459222704172134, + 0.015442310832440853, + 0.03153194487094879, + 0.01348842866718769, + 0.004667520057410002, + 0.009682761505246162, + 0.003732469631358981, + 0.007944525219500065, + 0.013535695150494576, + 0.006497447844594717, + 0.009213641285896301, + 0.012709839269518852, + 0.004075064789503813, + 0.0047657545655965805, + 0.002022393047809601, + 0.004444002639502287, + 0.011985582299530506, + 0.004171899054199457, + 0.007197716739028692, + 0.004107552580535412, + 0.007422134280204773, + 0.0017693197587504983, + 0.0014147330075502396, + 0.003444404574111104, + 0.004688918590545654, + 0.0022777491249144077, + 0.003045022487640381, + 0.007250379305332899, + 0.004002343397587538, + 0.003699467983096838, + 0.002063994761556387, + 0.004612588789314032, + 0.008646558970212936, + 0.004908240400254726, + 0.00871322862803936, + 0.007551948539912701, + 0.004537956789135933, + 0.006408047396689653, + 0.0037159377243369818, + 0.004706085193902254, + 0.007066770456731319, + 0.004274317994713783, + 0.0068531334400177, + 0.007031749002635479, + 0.014027640223503113, + 0.002471185987815261, + 0.0023759379982948303, + 0.0033843088895082474, + 0.008224042132496834, + 0.0024842610582709312, + 0.003817344782873988, + 0.030866215005517006, + 0.01550830528140068, + 0.036517564207315445, + 0.016984127461910248, + 0.013827142305672169, + 0.024899201467633247, + 0.013340979814529419, + 0.05333641916513443, + 0.05113202705979347, + 0.021136630326509476, + 0.04664616286754608, + 0.02809329889714718, + 0.02337663434445858, + 0.035564880818128586, + 0.025071771815419197, + 0.057836078107357025, + 0.0029131993651390076, + 0.0023088757880032063, + 0.002256074221804738, + 0.001298051211051643, + 0.003240562742576003, + 0.002668638015165925, + 0.0028140535578131676, + 0.0035455385223031044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.42889127135276794 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981371060013771 + ], + "opposite_logprob_tokens": [ + -3.529813766479492 + ], + "logprob_reward": -0.02981371060013771, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981371060013771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22290769219398499, + 0.05891710892319679, + 0.029774904251098633, + 0.04380440711975098, + 0.031817931681871414, + 0.06866148859262466, + 0.0651947632431984, + 0.20032747089862823, + 0.22931215167045593, + 0.04928211495280266 + ], + "credit_map": [ + 0.03301418200135231, + 0.020547911524772644, + 0.034939177334308624, + 0.01843954436480999, + 0.0201149620115757, + 0.044111575931310654, + 0.029163720086216927, + 0.022576618939638138, + 0.007476232945919037, + 0.0057394918985664845, + 0.005977831315249205, + 0.006230048835277557, + 0.009983888827264309, + 0.006555059924721718, + 0.010577796027064323, + 0.006376759149134159, + 0.004707551095634699, + 0.0024482826702296734, + 0.0024964141193777323, + 0.002950056456029415, + 0.003965649753808975, + 0.004628538154065609, + 0.005245101638138294, + 0.0033333110623061657, + 0.006960684899240732, + 0.012003202922642231, + 0.0029795316513627768, + 0.00401829881593585, + 0.004261340014636517, + 0.006992402020841837, + 0.0036313000600785017, + 0.002957647666335106, + 0.004315685480833054, + 0.00339442677795887, + 0.0035702502354979515, + 0.0030278265476226807, + 0.004384578205645084, + 0.00418752571567893, + 0.0058275857008993626, + 0.0031100495252758265, + 0.012042967602610588, + 0.010071604512631893, + 0.006895364727824926, + 0.007481238804757595, + 0.007520362734794617, + 0.008319498039782047, + 0.0074087935499846935, + 0.008921658620238304, + 0.011445646174252033, + 0.02230125479400158, + 0.0033247682731598616, + 0.003355196909978986, + 0.004981860518455505, + 0.010738303884863853, + 0.004266943316906691, + 0.0047807851806283, + 0.027920251712203026, + 0.015202641487121582, + 0.035167306661605835, + 0.01404432114213705, + 0.015227559953927994, + 0.02683434449136257, + 0.013206776231527328, + 0.05272427946329117, + 0.041443608701229095, + 0.017292216420173645, + 0.03743305429816246, + 0.0221560001373291, + 0.018873441964387894, + 0.02833901345729828, + 0.019074952229857445, + 0.04469984397292137, + 0.013402353972196579, + 0.003959653899073601, + 0.009002487175166607, + 0.004017390310764313, + 0.004034298472106457, + 0.006286890711635351, + 0.004450805950909853, + 0.004128233529627323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.02981371060013771 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206341087818146 + ], + "opposite_logprob_tokens": [ + -2.812063455581665 + ], + "logprob_reward": -0.06206341087818146, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206341087818146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10955686122179031, + 0.04835427552461624, + 0.017362285405397415, + 0.03657419607043266, + 0.02691037580370903, + 0.15695390105247498, + 0.07901851087808609, + 0.09792256355285645, + 0.2509763836860657, + 0.1763707548379898 + ], + "credit_map": [ + 0.017818080261349678, + 0.014607654884457588, + 0.01717890240252018, + 0.0070826890878379345, + 0.007213995326310396, + 0.020701592788100243, + 0.00788629986345768, + 0.017067624256014824, + 0.009843220934271812, + 0.003554530208930373, + 0.006195336114615202, + 0.002946369582787156, + 0.0050900946371257305, + 0.009716351516544819, + 0.004794645588845015, + 0.006213716231286526, + 0.0034506043884903193, + 0.0015416508540511131, + 0.0028611309826374054, + 0.0008885743445716798, + 0.0019467136589810252, + 0.002486614277586341, + 0.0014903034316375852, + 0.002696689683943987, + 0.005672914441674948, + 0.009762649424374104, + 0.002895445330068469, + 0.002099597128108144, + 0.003733540652319789, + 0.006629169452935457, + 0.002569753210991621, + 0.003211118746548891, + 0.00511960918083787, + 0.003865655977278948, + 0.0022969339042901993, + 0.0013679557014256716, + 0.003910848405212164, + 0.005190643016248941, + 0.003288998268544674, + 0.0018697282066568732, + 0.024506384506821632, + 0.02889101766049862, + 0.014591258950531483, + 0.007061496376991272, + 0.016201836988329887, + 0.02681565470993519, + 0.009830521419644356, + 0.029055705294013023, + 0.01380992028862238, + 0.02291363663971424, + 0.0064612990245223045, + 0.0044432408176362514, + 0.0068327453918755054, + 0.011145452968776226, + 0.004518921487033367, + 0.00889328308403492, + 0.014019617810845375, + 0.0074532655999064445, + 0.01504529919475317, + 0.00847501028329134, + 0.008703899569809437, + 0.012553391978144646, + 0.006928793620318174, + 0.02474326454102993, + 0.04413621872663498, + 0.017194535583257675, + 0.04177036136388779, + 0.022999808192253113, + 0.019958822056651115, + 0.030664891004562378, + 0.019867539405822754, + 0.05438415706157684, + 0.060105133801698685, + 0.011621647514402866, + 0.03756681829690933, + 0.007470159791409969, + 0.01135336421430111, + 0.025456279516220093, + 0.009552376344799995, + 0.013244951143860817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.06206341087818146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42892271280288696 + ], + "opposite_logprob_tokens": [ + -1.0539226531982422 + ], + "logprob_reward": -0.42892271280288696, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42892271280288696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02078583464026451, + 0.01835549622774124, + 0.009099761955440044, + 0.005657203029841185, + 0.01226884126663208, + 0.009875135496258736, + 0.01087117474526167, + 0.04067779704928398, + 0.8305742144584656, + 0.041834503412246704 + ], + "credit_map": [ + 0.002498435555025935, + 0.00284201349131763, + 0.0015398578252643347, + 0.002258750144392252, + 0.0028760982677340508, + 0.003279393073171377, + 0.003067797515541315, + 0.00242348900064826, + 0.003095312276855111, + 0.0016085405368357897, + 0.002700209617614746, + 0.0012403220171108842, + 0.0016881966730579734, + 0.0033251987770199776, + 0.001590587547980249, + 0.003107127733528614, + 0.0017069873865693808, + 0.000601386942435056, + 0.0010590705787763, + 0.0005357934860512614, + 0.0010642237029969692, + 0.001860937918536365, + 0.0012333965860307217, + 0.0010379654122516513, + 0.0009747460135258734, + 0.0010680065024644136, + 0.0005867593572475016, + 0.00039616733556613326, + 0.0006640606443397701, + 0.0009433141094632447, + 0.0005530770868062973, + 0.00047107230057008564, + 0.002293139696121216, + 0.0017189884092658758, + 0.001135373837314546, + 0.0006657230551354587, + 0.001795959658920765, + 0.0021286047995090485, + 0.0012247022241353989, + 0.0013063492951914668, + 0.001500523998402059, + 0.0019830730743706226, + 0.0008187229977920651, + 0.0007118419744074345, + 0.0010497788898646832, + 0.0016143085667863488, + 0.0007152357720769942, + 0.001481650280766189, + 0.001821125508286059, + 0.002796539571136236, + 0.0011446198914200068, + 0.0006773798959329724, + 0.000989019637927413, + 0.0016580922529101372, + 0.0008967253379523754, + 0.000887672882527113, + 0.005856892094016075, + 0.002951681148260832, + 0.006901584565639496, + 0.0032967161387205124, + 0.0035635381937026978, + 0.004973985254764557, + 0.0027168639935553074, + 0.01041653472930193, + 0.14830324053764343, + 0.0609770342707634, + 0.1353374570608139, + 0.08278857916593552, + 0.07156918942928314, + 0.09927111119031906, + 0.07268717885017395, + 0.15964043140411377, + 0.012691744603216648, + 0.002374197356402874, + 0.009328298270702362, + 0.0024890524800866842, + 0.003032858483493328, + 0.0052127474918961525, + 0.0027833180502057076, + 0.003922289703041315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.42892271280288696 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unsteady, handheld medium shot, its slight tremor heightening the emotional tension, captures a moment of melancholy as a young woman stands quietly at a shimmering railroad crossing amidst a swirling, heavy snowfall. Her long, dark brown hair, a few strands wet from the snow, plastered to her cheeks. A hint of confusion and sorrow lingers in her clear eyes, her cheeks slightly flushed by the chill. She wears a thick, crimson wool coat, a cream-colored wool scarf tightly wrapped around her neck. Each gentle breath condenses into a distinct white mist in the icy air, then slowly dissipates into the snow. Behind her, at the railroad crossing, the red warning lights flicker rhythmically, casting intermittent halos of red light on her and the surrounding tracks, now covered in fresh snow. The entire scene presents the delicate light and shadow and extremely high color saturation of an animated film. The deep red of the girl's coat and the pure white snow scene form a strong visual contrast. Under the soft diffuse light of the cloudy day, they together create an atmosphere that is both beautiful and lonely, full of a sense of story and a hint of imperceptible expectation.", + "seed": 14120, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_2/index_8708_seed_14120_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18649898562580347, + "vlm_logprob": -0.18649898562580347, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.18649898562580347, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4397019147872925 + ], + "opposite_logprob_tokens": [ + -1.0647019147872925 + ], + "logprob_reward": -0.4397019147872925, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4397019147872925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05388009548187256, + 0.08808579295873642, + 0.1976141631603241, + 0.04697088152170181, + 0.03295370563864708, + 0.031170332804322243, + 0.05223666876554489, + 0.4056137800216675, + 0.07482077181339264, + 0.016653841361403465 + ], + "credit_map": [ + 0.008192898705601692, + 0.006260068155825138, + 0.0034482567571103573, + 0.007587146479636431, + 0.007481070701032877, + 0.0052077872678637505, + 0.007363986689597368, + 0.008338878862559795, + 0.019963469356298447, + 0.014886169694364071, + 0.007856273092329502, + 0.0052924747578799725, + 0.005676245782524347, + 0.012456042692065239, + 0.004778328817337751, + 0.01717679388821125, + 0.07616909593343735, + 0.008584189228713512, + 0.007629842963069677, + 0.005864214152097702, + 0.007687616162002087, + 0.01123526506125927, + 0.0056902458891272545, + 0.07475369423627853, + 0.01404070109128952, + 0.005335233639925718, + 0.004671523347496986, + 0.002590528689324856, + 0.003703462425619364, + 0.0038502514362335205, + 0.002431797329336405, + 0.01034738589078188, + 0.00829642079770565, + 0.004326358903199434, + 0.0033575035631656647, + 0.002179449424147606, + 0.002408683532848954, + 0.0036995571572333574, + 0.0020494896452873945, + 0.0066362414509058, + 0.005561139900237322, + 0.002901451662182808, + 0.0023906256537884474, + 0.0047933789901435375, + 0.004380420781672001, + 0.004224598407745361, + 0.002794584957882762, + 0.004124132916331291, + 0.007106836885213852, + 0.0037495684809982777, + 0.005776066333055496, + 0.010352691635489464, + 0.006151074077934027, + 0.006797230802476406, + 0.005239568650722504, + 0.007063630968332291, + 0.13871699571609497, + 0.01779027469456196, + 0.021370049566030502, + 0.01197150256484747, + 0.023633133620023727, + 0.02529599703848362, + 0.014998863451182842, + 0.15183693170547485, + 0.022260094061493874, + 0.008463442325592041, + 0.004942053463310003, + 0.0042622508481144905, + 0.004448123276233673, + 0.005975265055894852, + 0.003423799993470311, + 0.0210457444190979, + 0.002557027153670788, + 0.0018830791814252734, + 0.0017866400303319097, + 0.0019491545390337706, + 0.0025072265416383743, + 0.0024045079480856657, + 0.0017326392699033022, + 0.0018335651839151978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.4397019147872925 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134450912475586 + ], + "opposite_logprob_tokens": [ + -1.3134450912475586 + ], + "logprob_reward": -0.3134450912475586, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134450912475586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12598782777786255, + 0.03335411101579666, + 0.19186045229434967, + 0.017432531341910362, + 0.022059021517634392, + 0.01851564832031727, + 0.1358574777841568, + 0.4047463536262512, + 0.02024860307574272, + 0.029937952756881714 + ], + "credit_map": [ + 0.02805335260927677, + 0.008429013192653656, + 0.007658079266548157, + 0.01124644000083208, + 0.010766309686005116, + 0.010789175517857075, + 0.008273047395050526, + 0.04077240452170372, + 0.007032423280179501, + 0.005299908109009266, + 0.0028741753194481134, + 0.0029447078704833984, + 0.002402488375082612, + 0.0045748548582196236, + 0.0019461301853880286, + 0.006279424298554659, + 0.07366900891065598, + 0.00869507435709238, + 0.008259630762040615, + 0.005979521200060844, + 0.007379067130386829, + 0.01158804353326559, + 0.005246356595307589, + 0.07104376703500748, + 0.0025419271551072598, + 0.0016954127931967378, + 0.002701972611248493, + 0.0018567199585959315, + 0.0026102466508746147, + 0.001849530846811831, + 0.00174477300606668, + 0.0024319468066096306, + 0.004833392798900604, + 0.0018159168539568782, + 0.003041545394808054, + 0.0011547590838745236, + 0.002003070432692766, + 0.003511792980134487, + 0.0013412297703325748, + 0.004357315134257078, + 0.004040705971419811, + 0.0016100086504593492, + 0.001462335349060595, + 0.0022822131868451834, + 0.0016935588791966438, + 0.0027064168825745583, + 0.001261814497411251, + 0.0034585949033498764, + 0.04463854804635048, + 0.007214577402919531, + 0.010744941420853138, + 0.004700111225247383, + 0.005818676203489304, + 0.009146424010396004, + 0.004717231262475252, + 0.048876967281103134, + 0.1292743682861328, + 0.019378453493118286, + 0.02588815987110138, + 0.017299365252256393, + 0.019121941179037094, + 0.030570559203624725, + 0.016588646918535233, + 0.1466248482465744, + 0.005692982580512762, + 0.0018523575272411108, + 0.0009028662461787462, + 0.0018629204714670777, + 0.0016064405208453536, + 0.001300714910030365, + 0.0010419712634757161, + 0.005988348741084337, + 0.004883531481027603, + 0.0029005976393818855, + 0.004064956214278936, + 0.003277979791164398, + 0.0029128307942301035, + 0.0046066963113844395, + 0.002607114613056183, + 0.004684245213866234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.3134450912475586 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203527010977268 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203527010977268, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203527010977268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06982075423002243, + 0.1460016518831253, + 0.29867157340049744, + 0.03427446633577347, + 0.009175139479339123, + 0.09724481403827667, + 0.049744900315999985, + 0.22587472200393677, + 0.04982829838991165, + 0.019363682717084885 + ], + "credit_map": [ + 0.013517058454453945, + 0.005126653704792261, + 0.006192618049681187, + 0.006352516822516918, + 0.006687706336379051, + 0.006916259415447712, + 0.005570457316935062, + 0.01945747807621956, + 0.04195350781083107, + 0.020854070782661438, + 0.009298279881477356, + 0.008158288896083832, + 0.0069778175093233585, + 0.014482717961072922, + 0.0059814429841935635, + 0.03829551488161087, + 0.11698009818792343, + 0.013137215748429298, + 0.010407960042357445, + 0.0088067427277565, + 0.010345318354666233, + 0.01723070815205574, + 0.007488241884857416, + 0.114275261759758, + 0.006022684276103973, + 0.008098363876342773, + 0.004860594402998686, + 0.0026087516453117132, + 0.0029339988250285387, + 0.0031084520742297173, + 0.0022351678926497698, + 0.004406447522342205, + 0.001561918412335217, + 0.0009702462702989578, + 0.0007235079538077116, + 0.0006361941923387349, + 0.000892247015144676, + 0.0016719844425097108, + 0.0007008942775428295, + 0.0020181455183774233, + 0.031411755830049515, + 0.0063053821213543415, + 0.005656805820763111, + 0.00413197698071599, + 0.004859421867877245, + 0.006656490731984377, + 0.003730661468580365, + 0.03449230641126633, + 0.015136569738388062, + 0.002657683566212654, + 0.0047256420366466045, + 0.002732305321842432, + 0.0024799846578389406, + 0.003507381770759821, + 0.0020561255514621735, + 0.016449205577373505, + 0.07405439019203186, + 0.010321210138499737, + 0.013011223636567593, + 0.00888423714786768, + 0.01133732870221138, + 0.015879018232226372, + 0.008850675076246262, + 0.08353663235902786, + 0.015165918506681919, + 0.0037734508514404297, + 0.00342820817604661, + 0.0033482802100479603, + 0.0024167709052562714, + 0.0029743225313723087, + 0.002077404409646988, + 0.016643937677145004, + 0.003354286774992943, + 0.0017094342038035393, + 0.002082123653963208, + 0.0021511963568627834, + 0.0019982056692242622, + 0.002654067939147353, + 0.0016733077354729176, + 0.0037410592194646597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.014203527010977268 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810453042387962 + ], + "opposite_logprob_tokens": [ + -3.28810453414917 + ], + "logprob_reward": -0.03810453042387962, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810453042387962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11637680232524872, + 0.11695744842290878, + 0.07796362042427063, + 0.04421104118227959, + 0.05757772549986839, + 0.024141507223248482, + 0.02143200673162937, + 0.4648240804672241, + 0.05185111239552498, + 0.02466459386050701 + ], + "credit_map": [ + 0.017420781776309013, + 0.012578189373016357, + 0.013493357226252556, + 0.010753108188509941, + 0.013139061629772186, + 0.015098730102181435, + 0.011857517994940281, + 0.022036070004105568, + 0.03282390162348747, + 0.012365695089101791, + 0.00884671974927187, + 0.005128702614456415, + 0.006654392462223768, + 0.015711648389697075, + 0.005600910633802414, + 0.02982548624277115, + 0.02059616521000862, + 0.0071973553858697414, + 0.006735232193022966, + 0.004519659094512463, + 0.0078034973703324795, + 0.006105888169258833, + 0.006876029539853334, + 0.018129805102944374, + 0.007408633828163147, + 0.006512094754725695, + 0.006265883333981037, + 0.005448623094707727, + 0.003996489569544792, + 0.00395583501085639, + 0.005062824115157127, + 0.005560664925724268, + 0.01689658872783184, + 0.003961063455790281, + 0.005146516487002373, + 0.0035701326560229063, + 0.0033097879495471716, + 0.005614453926682472, + 0.0022154750768095255, + 0.016863709315657616, + 0.004436788149178028, + 0.0017635535914450884, + 0.0020620406139642, + 0.0015088821528479457, + 0.0031986802350729704, + 0.003491607727482915, + 0.0022466641385108232, + 0.005433293525129557, + 0.0030119556467980146, + 0.003284728853031993, + 0.003148806979879737, + 0.0021330388262867928, + 0.0027933900710195303, + 0.0029277666471898556, + 0.0019562426023185253, + 0.0021760803647339344, + 0.16070005297660828, + 0.019493775442242622, + 0.021609550341963768, + 0.01350424811244011, + 0.023500513285398483, + 0.02992144413292408, + 0.01693025231361389, + 0.17916430532932281, + 0.014041232876479626, + 0.004538013134151697, + 0.0026067993603646755, + 0.003996137995272875, + 0.004102399107068777, + 0.0038423764053732157, + 0.0024763664696365595, + 0.016247788444161415, + 0.0037663846742361784, + 0.002524091862142086, + 0.0028893547132611275, + 0.002307303249835968, + 0.0025011268444359303, + 0.0047391727566719055, + 0.0021699746139347553, + 0.003767186775803566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.03810453042387962 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270398646593094 + ], + "opposite_logprob_tokens": [ + -2.127039909362793 + ], + "logprob_reward": -0.1270398646593094, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270398646593094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07623764127492905, + 0.2145807445049286, + 0.0824553444981575, + 0.027687346562743187, + 0.032372381538152695, + 0.03602795675396919, + 0.06049683317542076, + 0.3288719952106476, + 0.10669439285993576, + 0.03457533195614815 + ], + "credit_map": [ + 0.013073395937681198, + 0.007584038190543652, + 0.007420875132083893, + 0.009060412645339966, + 0.007730061188340187, + 0.009393519721925259, + 0.00642132805660367, + 0.015554009936749935, + 0.05235149711370468, + 0.036316901445388794, + 0.016389135271310806, + 0.012610778212547302, + 0.012696483172476292, + 0.02762228436768055, + 0.010906459763646126, + 0.04568719491362572, + 0.02676616609096527, + 0.0061523872427642345, + 0.00497175008058548, + 0.0034130499698221684, + 0.003842248348519206, + 0.005304936319589615, + 0.004234625492244959, + 0.027770183980464935, + 0.004345617722719908, + 0.00427246792241931, + 0.003287706058472395, + 0.0029738901648670435, + 0.0034529014956206083, + 0.0028116130270063877, + 0.002779794856905937, + 0.003763353917747736, + 0.011022050864994526, + 0.0025418337900191545, + 0.0016315765678882599, + 0.001105788047425449, + 0.00203772378154099, + 0.002296346239745617, + 0.0014238182920962572, + 0.01031324453651905, + 0.00938184093683958, + 0.00215841899625957, + 0.0024535481352359056, + 0.0026575506199151278, + 0.0025611582677811384, + 0.003935960121452808, + 0.0023890750017017126, + 0.010490401647984982, + 0.01754630357027054, + 0.003769311821088195, + 0.003932847175747156, + 0.0024069298524409533, + 0.004893757402896881, + 0.004896211437880993, + 0.00358038442209363, + 0.01947108469903469, + 0.10967186093330383, + 0.01459774561226368, + 0.01740811951458454, + 0.011834753677248955, + 0.016236916184425354, + 0.023104723542928696, + 0.012673644348978996, + 0.12334425002336502, + 0.031845491379499435, + 0.006641230545938015, + 0.007753095589578152, + 0.008129296824336052, + 0.004898523446172476, + 0.007947724312543869, + 0.005013875663280487, + 0.03446515277028084, + 0.005142563488334417, + 0.0031852361280471087, + 0.003943334333598614, + 0.0040463972836732864, + 0.0038525571580976248, + 0.003111188765615225, + 0.0037166655529290438, + 0.007577386684715748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.1270398646593094 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unsteady, handheld medium shot, its slight tremor heightening the emotional tension, captures a moment of melancholy as a young woman stands quietly at a shimmering railroad crossing amidst a swirling, heavy snowfall. Her long, dark brown hair, a few strands wet from the snow, plastered to her cheeks. A hint of confusion and sorrow lingers in her clear eyes, her cheeks slightly flushed by the chill. She wears a thick, crimson wool coat, a cream-colored wool scarf tightly wrapped around her neck. Each gentle breath condenses into a distinct white mist in the icy air, then slowly dissipates into the snow. Behind her, at the railroad crossing, the red warning lights flicker rhythmically, casting intermittent halos of red light on her and the surrounding tracks, now covered in fresh snow. The entire scene presents the delicate light and shadow and extremely high color saturation of an animated film. The deep red of the girl's coat and the pure white snow scene form a strong visual contrast. Under the soft diffuse light of the cloudy day, they together create an atmosphere that is both beautiful and lonely, full of a sense of story and a hint of imperceptible expectation.", + "seed": 14121, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_2/index_8708_seed_14121_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21532658711075783, + "vlm_logprob": -0.21532658711075783, + "vlm_yes_no_margin": 2.075, + "vlm_token_credit": -0.21532658711075783, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5395379066467285 + ], + "opposite_logprob_tokens": [ + -0.9145379066467285 + ], + "logprob_reward": -0.5395379066467285, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5395379066467285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34425240755081177, + 0.10243871808052063, + 0.09390340745449066, + 0.10023372620344162, + 0.07399243861436844, + 0.12274892628192902, + 0.0451664999127388, + 0.03122318722307682, + 0.047269273549318314, + 0.0387713760137558 + ], + "credit_map": [ + 0.044680964201688766, + 0.018575962632894516, + 0.0878119170665741, + 0.02102736383676529, + 0.020869378000497818, + 0.10252538323402405, + 0.032767001539468765, + 0.015994444489479065, + 0.020254256203770638, + 0.005962580908089876, + 0.014759406447410583, + 0.008689343929290771, + 0.011073877103626728, + 0.01385530922561884, + 0.014031094498932362, + 0.01381284836679697, + 0.008893584832549095, + 0.007878055796027184, + 0.008406469598412514, + 0.01841370016336441, + 0.01374752912670374, + 0.010296936146914959, + 0.01625383086502552, + 0.010013299994170666, + 0.015407654456794262, + 0.008183690719306469, + 0.019228126853704453, + 0.006433676928281784, + 0.010634070262312889, + 0.014184130355715752, + 0.013410909101366997, + 0.012751469388604164, + 0.010660837404429913, + 0.006804008036851883, + 0.009065438061952591, + 0.008601969107985497, + 0.008229443803429604, + 0.009626219049096107, + 0.012868218123912811, + 0.008136309683322906, + 0.009353182278573513, + 0.01186462678015232, + 0.008549175225198269, + 0.02668195404112339, + 0.018844613805413246, + 0.011892396956682205, + 0.029219312593340874, + 0.0063436636701226234, + 0.0076000262051820755, + 0.004604946356266737, + 0.005544010084122419, + 0.0074783796444535255, + 0.005148160737007856, + 0.0055075217969715595, + 0.005002210382372141, + 0.00428124750033021, + 0.004206818528473377, + 0.0031617418862879276, + 0.003187870606780052, + 0.0044929091818630695, + 0.0060671414248645306, + 0.0035725620109587908, + 0.004338477272540331, + 0.00219566747546196, + 0.007827440276741982, + 0.007310063578188419, + 0.006261611357331276, + 0.005813648924231529, + 0.005284717306494713, + 0.005919087212532759, + 0.006297548767179251, + 0.0025551547296345234, + 0.004788107704371214, + 0.006477853748947382, + 0.0048864237032830715, + 0.003613197710365057, + 0.008823595941066742, + 0.004244157578796148, + 0.0037331944331526756, + 0.0022048475220799446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.5395379066467285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28135040402412415 + ], + "opposite_logprob_tokens": [ + -1.4063503742218018 + ], + "logprob_reward": -0.28135040402412415, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28135040402412415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4899541735649109, + 0.1795438826084137, + 0.06510870158672333, + 0.062152963131666183, + 0.08405912667512894, + 0.04066194221377373, + 0.01245670486241579, + 0.02482772059738636, + 0.029555127024650574, + 0.011679653078317642 + ], + "credit_map": [ + 0.0745939314365387, + 0.01958419941365719, + 0.12281502783298492, + 0.038274385035037994, + 0.026808209717273712, + 0.12585480511188507, + 0.04230846092104912, + 0.03971515968441963, + 0.022925496101379395, + 0.009585325606167316, + 0.028353866189718246, + 0.018222751095891, + 0.016987750306725502, + 0.029150251299142838, + 0.02963748201727867, + 0.024680960923433304, + 0.010872909799218178, + 0.006047487258911133, + 0.008651775307953358, + 0.00868941843509674, + 0.007353188004344702, + 0.007993522100150585, + 0.011499017477035522, + 0.004001383204013109, + 0.01269504614174366, + 0.0062033203430473804, + 0.009881979785859585, + 0.007219893392175436, + 0.0049966019578278065, + 0.007943438366055489, + 0.007887360639870167, + 0.005325322970747948, + 0.016217073425650597, + 0.007075674831867218, + 0.012837742455303669, + 0.009752677753567696, + 0.008598467335104942, + 0.011558586731553078, + 0.009868931025266647, + 0.008149968460202217, + 0.007947081699967384, + 0.004101810045540333, + 0.0036860944237560034, + 0.005788838490843773, + 0.004745726939290762, + 0.004696500953286886, + 0.0047187902964651585, + 0.00497710146009922, + 0.001915952772833407, + 0.0015895037213340402, + 0.0013034052681177855, + 0.0017111640190705657, + 0.0021234829910099506, + 0.001363562885671854, + 0.0013632714981213212, + 0.0010863624047487974, + 0.0057510812766849995, + 0.001960654277354479, + 0.0031048667151480913, + 0.0024616222362965345, + 0.0025556799955666065, + 0.0033209749963134527, + 0.0019476000452414155, + 0.0037252414040267467, + 0.004373867996037006, + 0.002274059457704425, + 0.0051772575825452805, + 0.003058724571019411, + 0.0044214255176484585, + 0.00403321860358119, + 0.002026517875492573, + 0.004190055653452873, + 0.0019913078285753727, + 0.0015296678757295012, + 0.0016640505054965615, + 0.0009558806195855141, + 0.0013786383206024766, + 0.0014613039093092084, + 0.0011747920652851462, + 0.0015240118373185396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.28135040402412415 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016077283769845963 + ], + "opposite_logprob_tokens": [ + -4.141077518463135 + ], + "logprob_reward": -0.016077283769845963, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016077283769845963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2822308838367462, + 0.16761073470115662, + 0.1998218595981598, + 0.11516939848661423, + 0.11130478233098984, + 0.035575199872255325, + 0.0257656779140234, + 0.0221372302621603, + 0.020939260721206665, + 0.019445009529590607 + ], + "credit_map": [ + 0.029156943783164024, + 0.012095610611140728, + 0.0568239651620388, + 0.020122498273849487, + 0.022922910749912262, + 0.06650347262620926, + 0.03985067829489708, + 0.03475476801395416, + 0.025327840819954872, + 0.011577497236430645, + 0.025999734178185463, + 0.024224085733294487, + 0.013054199516773224, + 0.019509179517626762, + 0.02757137455046177, + 0.020346803590655327, + 0.021207060664892197, + 0.017977137118577957, + 0.015379623509943485, + 0.04453536868095398, + 0.024715274572372437, + 0.019206266850233078, + 0.040058545768260956, + 0.01674257032573223, + 0.012965959496796131, + 0.010256273671984673, + 0.013270321302115917, + 0.014671157114207745, + 0.015892554074525833, + 0.012477179057896137, + 0.023899389430880547, + 0.011736550368368626, + 0.01433534175157547, + 0.01026439480483532, + 0.00954827293753624, + 0.014527490362524986, + 0.020904723554849625, + 0.012957003898918629, + 0.02033703774213791, + 0.00843050517141819, + 0.005472563672810793, + 0.0031038338784128428, + 0.0031014089472591877, + 0.008321855217218399, + 0.003993102815002203, + 0.004352495539933443, + 0.005224240478128195, + 0.00200569536536932, + 0.005142182111740112, + 0.0022163446992635727, + 0.0025255922228097916, + 0.002418974880129099, + 0.0022846213541924953, + 0.003738962346687913, + 0.0026564488653093576, + 0.004782548639923334, + 0.0041101123206317425, + 0.0021824880968779325, + 0.0019647316075861454, + 0.003664507530629635, + 0.0028264482971280813, + 0.0027021842543035746, + 0.002265386050567031, + 0.00242137024179101, + 0.004942762199789286, + 0.001916617271490395, + 0.003274529241025448, + 0.0017341251950711012, + 0.002266468247398734, + 0.0030823545530438423, + 0.001815092284232378, + 0.001907307654619217, + 0.004277665633708239, + 0.002525447402149439, + 0.002343841129913926, + 0.0018602400086820126, + 0.0022286230232566595, + 0.002600169274955988, + 0.002518337219953537, + 0.0010906840907409787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.016077283769845963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038106366991996765 + ], + "opposite_logprob_tokens": [ + -3.2881064414978027 + ], + "logprob_reward": -0.038106366991996765, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038106366991996765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21823346614837646, + 0.10895056277513504, + 0.07363048195838928, + 0.07739829272031784, + 0.27209845185279846, + 0.12474903464317322, + 0.020655155181884766, + 0.034468673169612885, + 0.03673006594181061, + 0.03308583050966263 + ], + "credit_map": [ + 0.031206373125314713, + 0.014601020142436028, + 0.04075571894645691, + 0.020855167880654335, + 0.020418956875801086, + 0.028957096859812737, + 0.03944087028503418, + 0.02199825458228588, + 0.021184809505939484, + 0.006845690775662661, + 0.019744474440813065, + 0.010214058682322502, + 0.008993868716061115, + 0.014052675105631351, + 0.016859041526913643, + 0.011055940762162209, + 0.008003381080925465, + 0.006034740246832371, + 0.007731596007943153, + 0.015285760164260864, + 0.008006845600903034, + 0.008721609599888325, + 0.01150351669639349, + 0.008343029767274857, + 0.01574949361383915, + 0.006572359707206488, + 0.012721115723252296, + 0.007761185057461262, + 0.006687510292977095, + 0.009808741509914398, + 0.008424784056842327, + 0.0096731036901474, + 0.057705651968717575, + 0.0235566608607769, + 0.04284754395484924, + 0.022390585392713547, + 0.024291284382343292, + 0.03611336275935173, + 0.024356093257665634, + 0.04083726555109024, + 0.02092321217060089, + 0.01007838360965252, + 0.008509845472872257, + 0.018541298806667328, + 0.014832166954874992, + 0.014109214767813683, + 0.01875458098948002, + 0.019000330939888954, + 0.004412941634654999, + 0.0019565653055906296, + 0.0024639826733618975, + 0.0019410598324611783, + 0.002590096788480878, + 0.002803338458761573, + 0.002083285478875041, + 0.0024038860574364662, + 0.00448411051183939, + 0.0032820042688399553, + 0.003432241501286626, + 0.00684297364205122, + 0.004924594424664974, + 0.0034930051770061255, + 0.004332986660301685, + 0.0036767562851309776, + 0.006339150480926037, + 0.005255039315670729, + 0.005852421745657921, + 0.003397776046767831, + 0.0036346898414194584, + 0.00466248020529747, + 0.003651109291240573, + 0.003937398083508015, + 0.005874880123883486, + 0.0032845069654285908, + 0.005062727723270655, + 0.0037197356577962637, + 0.0055596851743757725, + 0.004068609792739153, + 0.002844183472916484, + 0.0026714988052845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.038106366991996765 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156097412109375 + ], + "opposite_logprob_tokens": [ + -1.7015609741210938 + ], + "logprob_reward": -0.20156097412109375, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156097412109375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25141671299934387, + 0.1942506730556488, + 0.1328817456960678, + 0.10909874737262726, + 0.06440471112728119, + 0.13189151883125305, + 0.027044350281357765, + 0.029968926683068275, + 0.02557806484401226, + 0.03346461430191994 + ], + "credit_map": [ + 0.03403951972723007, + 0.016281938180327415, + 0.03836480900645256, + 0.027044594287872314, + 0.018695730715990067, + 0.02956249564886093, + 0.05995280668139458, + 0.027474839240312576, + 0.030935019254684448, + 0.010167120024561882, + 0.03735821694135666, + 0.019003862515091896, + 0.012032625265419483, + 0.03482648730278015, + 0.02406631037592888, + 0.025861036032438278, + 0.01874220371246338, + 0.010963402688503265, + 0.014940476976335049, + 0.02575770765542984, + 0.01441988255828619, + 0.017592446878552437, + 0.0170606579631567, + 0.013404951430857182, + 0.018154440447688103, + 0.008874423801898956, + 0.016489680856466293, + 0.011715760454535484, + 0.009278465062379837, + 0.012906603515148163, + 0.012361222878098488, + 0.019318144768476486, + 0.012127699330449104, + 0.004991439171135426, + 0.00841339398175478, + 0.007958262227475643, + 0.0079570347443223, + 0.0068874419666826725, + 0.0081364456564188, + 0.007932991720736027, + 0.020434245467185974, + 0.01026490330696106, + 0.012396900914609432, + 0.024413075298070908, + 0.01242893747985363, + 0.01452571153640747, + 0.019914820790290833, + 0.017512915655970573, + 0.005480410531163216, + 0.003104912117123604, + 0.002458390546962619, + 0.0045088366605341434, + 0.0033839589450508356, + 0.003038414753973484, + 0.003605079371482134, + 0.0014643477043136954, + 0.0057300785556435585, + 0.002572370693087578, + 0.0037767759058624506, + 0.003249700181186199, + 0.0037051797844469547, + 0.003878817893564701, + 0.0030637341551482677, + 0.003992270678281784, + 0.004302738234400749, + 0.0021964511834084988, + 0.004454458132386208, + 0.0026260276790708303, + 0.0026689565274864435, + 0.0034434024710208178, + 0.0023808542173355818, + 0.0035051763989031315, + 0.007187576498836279, + 0.0038173485081642866, + 0.004363085608929396, + 0.0026443067472428083, + 0.0030722422525286674, + 0.005195921752601862, + 0.002084062434732914, + 0.0051000723615288734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.20156097412109375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unsteady, handheld medium shot, its slight tremor heightening the emotional tension, captures a moment of melancholy as a young woman stands quietly at a shimmering railroad crossing amidst a swirling, heavy snowfall. Her long, dark brown hair, a few strands wet from the snow, plastered to her cheeks. A hint of confusion and sorrow lingers in her clear eyes, her cheeks slightly flushed by the chill. She wears a thick, crimson wool coat, a cream-colored wool scarf tightly wrapped around her neck. Each gentle breath condenses into a distinct white mist in the icy air, then slowly dissipates into the snow. Behind her, at the railroad crossing, the red warning lights flicker rhythmically, casting intermittent halos of red light on her and the surrounding tracks, now covered in fresh snow. The entire scene presents the delicate light and shadow and extremely high color saturation of an animated film. The deep red of the girl's coat and the pure white snow scene form a strong visual contrast. Under the soft diffuse light of the cloudy day, they together create an atmosphere that is both beautiful and lonely, full of a sense of story and a hint of imperceptible expectation.", + "seed": 14122, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_2/index_8708_seed_14122_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.126284534111619, + "vlm_logprob": -0.126284534111619, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.126284534111619, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4826090931892395 + ], + "opposite_logprob_tokens": [ + -0.9826090931892395 + ], + "logprob_reward": -0.4826090931892395, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4826090931892395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18667805194854736, + 0.14916780591011047, + 0.12788733839988708, + 0.16363665461540222, + 0.16024212539196014, + 0.06339574605226517, + 0.06611897051334381, + 0.029707135632634163, + 0.03667283430695534, + 0.016493340954184532 + ], + "credit_map": [ + 0.02559845708310604, + 0.02157389000058174, + 0.027931170538067818, + 0.01768975891172886, + 0.030605286359786987, + 0.0248482134193182, + 0.017285998910665512, + 0.021145284175872803, + 0.019195934757590294, + 0.024199379608035088, + 0.01093962136656046, + 0.0290807094424963, + 0.02209322713315487, + 0.015955591574311256, + 0.019047128036618233, + 0.008656212128698826, + 0.027001112699508667, + 0.019421247765421867, + 0.017998576164245605, + 0.012629921548068523, + 0.014562929980456829, + 0.012937516905367374, + 0.011281901970505714, + 0.012054127641022205, + 0.014225730672478676, + 0.03224600851535797, + 0.018810337409377098, + 0.022173913195729256, + 0.024057775735855103, + 0.017034195363521576, + 0.016337629407644272, + 0.018751056864857674, + 0.018224895000457764, + 0.023160135373473167, + 0.014156512916088104, + 0.032520513981580734, + 0.016453687101602554, + 0.01725730299949646, + 0.024812115356326103, + 0.013656970113515854, + 0.012672358192503452, + 0.0057156323455274105, + 0.010436264798045158, + 0.004386889282613993, + 0.0047514233738183975, + 0.006199653726071119, + 0.0069547793827950954, + 0.012278747744858265, + 0.008670874871313572, + 0.012857806868851185, + 0.007577959448099136, + 0.006589028052985668, + 0.010584479197859764, + 0.007621311116963625, + 0.006342378444969654, + 0.005875132512301207, + 0.0057010771706700325, + 0.00407181354239583, + 0.0027610159013420343, + 0.0028851956594735384, + 0.0036396337673068047, + 0.0031164910178631544, + 0.0026231228839606047, + 0.004908784758299589, + 0.006147181149572134, + 0.006357989739626646, + 0.0035486984997987747, + 0.004973115865141153, + 0.004321357235312462, + 0.0043653161264956, + 0.003977295942604542, + 0.0029818781185895205, + 0.002260397421196103, + 0.0022767947521060705, + 0.0026535827200859785, + 0.0014053555205464363, + 0.002418907592073083, + 0.002249994548037648, + 0.0015471248188987374, + 0.0016811820678412914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.4826090931892395 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358619332313538 + ], + "opposite_logprob_tokens": [ + -3.6513586044311523 + ], + "logprob_reward": -0.026358619332313538, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358619332313538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35685983300209045, + 0.06433700770139694, + 0.10653158277273178, + 0.08165650069713593, + 0.080265112221241, + 0.05051339790225029, + 0.1652250438928604, + 0.015564779751002789, + 0.038600362837314606, + 0.04044632986187935 + ], + "credit_map": [ + 0.06336106359958649, + 0.02728518284857273, + 0.11352661997079849, + 0.02048998884856701, + 0.029262926429510117, + 0.04256689175963402, + 0.02439349703490734, + 0.035973697900772095, + 0.011241462081670761, + 0.010595858097076416, + 0.009379723109304905, + 0.008238543756306171, + 0.006732953246682882, + 0.006737223826348782, + 0.00527412723749876, + 0.0061371237970888615, + 0.02535291202366352, + 0.01328000146895647, + 0.014645557850599289, + 0.01103855762630701, + 0.009456243366003036, + 0.01229977235198021, + 0.009480121545493603, + 0.01097842212766409, + 0.01410458143800497, + 0.006423701997846365, + 0.021506546065211296, + 0.007188571151345968, + 0.005723460111767054, + 0.0077657876536250114, + 0.007766161113977432, + 0.011177699081599712, + 0.008302861824631691, + 0.013655153103172779, + 0.0069752829149365425, + 0.01661103032529354, + 0.009218554943799973, + 0.0077421762980520725, + 0.011676869355142117, + 0.0060831885784864426, + 0.009260588325560093, + 0.004829084500670433, + 0.006470272783190012, + 0.0038474544417113066, + 0.004323754925280809, + 0.004791968036442995, + 0.00519416993483901, + 0.01179611124098301, + 0.03047957643866539, + 0.016867689788341522, + 0.03374819830060005, + 0.012254835106432438, + 0.016861658543348312, + 0.01806323602795601, + 0.012191161513328552, + 0.024758705869317055, + 0.002100854180753231, + 0.0030234733130782843, + 0.0019283693982288241, + 0.0013359925942495465, + 0.0015653008595108986, + 0.001636054366827011, + 0.0018366971053183079, + 0.002138038631528616, + 0.00924172904342413, + 0.003214119467884302, + 0.005691674072295427, + 0.0022219722159206867, + 0.002427441766485572, + 0.00481068529188633, + 0.0027980839367955923, + 0.008194660767912865, + 0.008896037004888058, + 0.003226004308089614, + 0.010232068598270416, + 0.0025253870990127325, + 0.0034793135710060596, + 0.0044766198843717575, + 0.002751322230324149, + 0.004859582521021366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.026358619332313538 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097875464707613 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.0097875464707613, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097875464707613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13086725771427155, + 0.1535748839378357, + 0.07203713804483414, + 0.04227651655673981, + 0.25952741503715515, + 0.03548983857035637, + 0.09718748927116394, + 0.038920365273952484, + 0.06060391664505005, + 0.1095152199268341 + ], + "credit_map": [ + 0.020401302725076675, + 0.01643489859998226, + 0.02438330091536045, + 0.014135582372546196, + 0.017650531604886055, + 0.012634975835680962, + 0.012173060327768326, + 0.013053597882390022, + 0.02663303166627884, + 0.022940365597605705, + 0.014658845961093903, + 0.026157883927226067, + 0.013547907583415508, + 0.014787363819777966, + 0.0205417238175869, + 0.01430774386972189, + 0.015546787530183792, + 0.006855516228824854, + 0.009019511751830578, + 0.01011833269149065, + 0.006438788026571274, + 0.006977966520935297, + 0.011788954958319664, + 0.00529127474874258, + 0.006921226624399424, + 0.007901746779680252, + 0.006323151756078005, + 0.003878115676343441, + 0.005676629021763802, + 0.004484910052269697, + 0.003383058588951826, + 0.003707671072334051, + 0.037557248026132584, + 0.028210721909999847, + 0.02815556339919567, + 0.04533369094133377, + 0.022759554907679558, + 0.02360389567911625, + 0.03924626111984253, + 0.0346604660153389, + 0.005538663361221552, + 0.004603567533195019, + 0.0034641395322978497, + 0.005613410379737616, + 0.0036438582465052605, + 0.0033949408680200577, + 0.004485745448619127, + 0.004745506681501865, + 0.01655442640185356, + 0.01627988927066326, + 0.014549956656992435, + 0.009413366205990314, + 0.011849699541926384, + 0.011763963848352432, + 0.008769350126385689, + 0.008006823249161243, + 0.005675645545125008, + 0.006037631072103977, + 0.0051452345214784145, + 0.004041251726448536, + 0.004558441694825888, + 0.004098786041140556, + 0.003994889557361603, + 0.005368479061871767, + 0.01005138736218214, + 0.009687265381217003, + 0.005875662434846163, + 0.006696848198771477, + 0.007271722424775362, + 0.0066360286436975, + 0.006521372124552727, + 0.007863625884056091, + 0.014731045812368393, + 0.021098429337143898, + 0.015141509473323822, + 0.012665239162743092, + 0.013423476368188858, + 0.011302931234240532, + 0.010592550970613956, + 0.010560022667050362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0097875464707613 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336962565779686 + ], + "opposite_logprob_tokens": [ + -3.408696174621582 + ], + "logprob_reward": -0.0336962565779686, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336962565779686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.50779128074646, + 0.062272559851408005, + 0.04841354116797447, + 0.07924645394086838, + 0.07027073949575424, + 0.02150742895901203, + 0.09106943756341934, + 0.03869963809847832, + 0.028151683509349823, + 0.05257721617817879 + ], + "credit_map": [ + 0.0797479972243309, + 0.031614452600479126, + 0.08022788166999817, + 0.02984553948044777, + 0.03959525749087334, + 0.06796930730342865, + 0.047715481370687485, + 0.13107538223266602, + 0.011409226804971695, + 0.007932747714221478, + 0.007489676587283611, + 0.008234330452978611, + 0.005667048040777445, + 0.006112491711974144, + 0.007211764343082905, + 0.008215275593101978, + 0.009518932551145554, + 0.007005068939179182, + 0.008579058572649956, + 0.004126499406993389, + 0.004231573548167944, + 0.0046781087294220924, + 0.004546490032225847, + 0.005727811250835657, + 0.010453416034579277, + 0.006402290426194668, + 0.011629834771156311, + 0.006863303016871214, + 0.007824555970728397, + 0.010648873634636402, + 0.007870717905461788, + 0.017553463578224182, + 0.013854902237653732, + 0.007602343335747719, + 0.0075786178931593895, + 0.010090280324220657, + 0.006141578312963247, + 0.006760127376765013, + 0.008260590955615044, + 0.009982307441532612, + 0.00423352001234889, + 0.0021346183493733406, + 0.00312699843198061, + 0.0018980372697114944, + 0.0024492989759892225, + 0.002318062586709857, + 0.0022043657954782248, + 0.0031425287015736103, + 0.015072043985128403, + 0.008580856956541538, + 0.014491657726466656, + 0.0075891753658652306, + 0.007525489665567875, + 0.012334738858044147, + 0.007645425386726856, + 0.017830053344368935, + 0.006522142328321934, + 0.0035684339236468077, + 0.005241608712822199, + 0.0030992499087005854, + 0.004330300725996494, + 0.0037647634744644165, + 0.0037776196841150522, + 0.008395521901547909, + 0.004294716753065586, + 0.003923255950212479, + 0.004985195584595203, + 0.0029977457597851753, + 0.0026806413661688566, + 0.002999457297846675, + 0.0021566448267549276, + 0.004114026203751564, + 0.00833663996309042, + 0.004700542893260717, + 0.010796576738357544, + 0.004152929875999689, + 0.004338896833360195, + 0.006075834855437279, + 0.004836706444621086, + 0.00933909323066473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.0336962565779686 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot with an unsteady, handheld camera movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at a railroad crossing during heavy snowfall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character a young woman wearing a thick crimson wool coat and a cream-colored wool scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are red warning lights visible flickering rhythmically behind the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of an animated film with extremely high color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897115498781204 + ], + "opposite_logprob_tokens": [ + -2.5789711475372314 + ], + "logprob_reward": -0.07897115498781204, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897115498781204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20554828643798828, + 0.2620640993118286, + 0.05446169897913933, + 0.03896172344684601, + 0.041591305285692215, + 0.04741397500038147, + 0.13449127972126007, + 0.0772557258605957, + 0.05713885650038719, + 0.081073097884655 + ], + "credit_map": [ + 0.033602241426706314, + 0.02765456959605217, + 0.035540107637643814, + 0.019471513107419014, + 0.01903129182755947, + 0.02203367091715336, + 0.01997440680861473, + 0.028240466490387917, + 0.0328739657998085, + 0.04611854627728462, + 0.02262480929493904, + 0.048114191740751266, + 0.026902060955762863, + 0.026653127744793892, + 0.03691078722476959, + 0.021866606548428535, + 0.007756727281957865, + 0.009270291775465012, + 0.006909665651619434, + 0.005495051853358746, + 0.007851243019104004, + 0.005760255269706249, + 0.006488917861133814, + 0.004929542075842619, + 0.004862028174102306, + 0.0065588862635195255, + 0.005762961693108082, + 0.004599934909492731, + 0.004356265999376774, + 0.004391390830278397, + 0.0034082485362887383, + 0.005022004712373018, + 0.0063303085044026375, + 0.008122711442410946, + 0.0035587986931204796, + 0.005860848352313042, + 0.006671235430985689, + 0.00365385296754539, + 0.004594855010509491, + 0.002798694185912609, + 0.0104445219039917, + 0.004791195038706064, + 0.004842257127165794, + 0.0046771615743637085, + 0.005973737686872482, + 0.004929551854729652, + 0.0038481131196022034, + 0.007907437160611153, + 0.020541440695524216, + 0.019815025851130486, + 0.023127585649490356, + 0.010233672335743904, + 0.017529867589473724, + 0.015001746825873852, + 0.008586524054408073, + 0.019655415788292885, + 0.012958782725036144, + 0.0075202942825853825, + 0.010111930780112743, + 0.005976548884063959, + 0.009408303536474705, + 0.0077544827945530415, + 0.009596078656613827, + 0.013929303735494614, + 0.01137352455407381, + 0.006684230174869299, + 0.010045363567769527, + 0.0036623873747885227, + 0.004864174872636795, + 0.007144996430724859, + 0.003954366780817509, + 0.009409811347723007, + 0.00993417203426361, + 0.016680782660841942, + 0.010064245201647282, + 0.00939169805496931, + 0.010856140404939651, + 0.008871158584952354, + 0.008026433177292347, + 0.007248471025377512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.07897115498781204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_3/index_tensor([706, 706])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_3/index_tensor([706, 706])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6a8aa285a97d65a0d81f8a120746ab486678879a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_3/index_tensor([706, 706])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot, framed from a fixed low-side angle, captures a scene of profound tranquility that embodies traditional Japanese aesthetics. In the background, water cascades in a gentle, continuous stream, flowing down a dark, damp rock wall. The rock surface is almost completely covered by a thick, velvety carpet of deep green moss that appears saturated with moisture. In the foreground, a sturdy fern serves as the central subject, its dark, gnarled base firmly planted. From its core, a magnificent crown of large, feathery fronds unfolds, each boasting a striking copper-red hue that contrasts strikingly with the earthy tones of the scene. A fine, ethereal mist, created by the gentle rush of the waterfall, drifts through the air and delicately settles on the fern's fronds, causing countless tiny droplets to cling to the edges of each leaf, shimmering like miniature jewels. The entire composition is bathed in a soft, diffused, cool-toned light that enhances the rich textures of the wet stone and delicate foliage, while the minimalist composition creates a serene, contemplative atmosphere, finding a quiet, fleeting beauty in this perfectly balanced natural moment.", + "seed": 28510, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_3/index_706_seed_28510_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.007401557289995253, + "vlm_logprob": -0.007401557289995253, + "vlm_yes_no_margin": 5.425, + "vlm_token_credit": -0.007401557289995253, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014431505464017391 + ], + "opposite_logprob_tokens": [ + -4.639431476593018 + ], + "logprob_reward": -0.014431505464017391, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.014431505464017391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10065135359764099, + 0.1982303112745285, + 0.31913551688194275, + 0.04552465304732323, + 0.04431859776377678, + 0.03058895282447338, + 0.052496276795864105, + 0.07909572869539261, + 0.06441671401262283, + 0.06554188579320908 + ], + "credit_map": [ + 0.008210272528231144, + 0.012042987160384655, + 0.014385050162672997, + 0.01896039769053459, + 0.00706287520006299, + 0.01100632268935442, + 0.016863208264112473, + 0.012120245024561882, + 0.006378220394253731, + 0.014253570698201656, + 0.01117327157407999, + 0.009170119650661945, + 0.017932727932929993, + 0.022118043154478073, + 0.05761931091547012, + 0.05958506464958191, + 0.006947134621441364, + 0.006297505926340818, + 0.005351726431399584, + 0.00904653687030077, + 0.008789760060608387, + 0.014232178218662739, + 0.13456828892230988, + 0.1339024156332016, + 0.003999422304332256, + 0.0049473876133561134, + 0.008175131864845753, + 0.006714735180139542, + 0.0037946784868836403, + 0.004604957066476345, + 0.006204917561262846, + 0.007083428092300892, + 0.003191961906850338, + 0.003312681568786502, + 0.005879608914256096, + 0.008197339251637459, + 0.0040844157338142395, + 0.004451228305697441, + 0.007917284034192562, + 0.007284080609679222, + 0.0029673203825950623, + 0.0036830632016062737, + 0.003218527417629957, + 0.0030097675044089556, + 0.0042757922783494, + 0.00522985402494669, + 0.004198010545223951, + 0.004006620496511459, + 0.0035405682865530252, + 0.006411720532923937, + 0.007574805524200201, + 0.004736724309623241, + 0.005869431421160698, + 0.008300596848130226, + 0.007679592352360487, + 0.008382846601307392, + 0.008444493636488914, + 0.00602177157998085, + 0.009075191803276539, + 0.013229395262897015, + 0.00583891524001956, + 0.008436019532382488, + 0.01391545683145523, + 0.014134488068521023, + 0.0025625370908528566, + 0.0029630744829773903, + 0.0046483674086630344, + 0.004182299133390188, + 0.004113304428756237, + 0.005975340958684683, + 0.019553739577531815, + 0.020418059080839157, + 0.007179782260209322, + 0.006810815539211035, + 0.008559666574001312, + 0.007704867050051689, + 0.008063599467277527, + 0.010758927091956139, + 0.008106028661131859, + 0.008358211256563663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.014431505464017391 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008186326595023274 + ], + "opposite_logprob_tokens": [ + -7.125818729400635 + ], + "logprob_reward": -0.0008186326595023274, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008186326595023274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05913601443171501, + 0.06697497516870499, + 0.6643725037574768, + 0.00953615177422762, + 0.033283524215221405, + 0.003913314081728458, + 0.011233638040721416, + 0.016489803791046143, + 0.12021130323410034, + 0.014848760329186916 + ], + "credit_map": [ + 0.005404969211667776, + 0.005258046556264162, + 0.009435637854039669, + 0.011690294370055199, + 0.00565094081684947, + 0.0059263575822114944, + 0.00883224606513977, + 0.0069375233724713326, + 0.002603257307782769, + 0.0024704174138605595, + 0.00332346698269248, + 0.004842059686779976, + 0.004468236118555069, + 0.004194770008325577, + 0.022520745173096657, + 0.02255202643573284, + 0.01393825002014637, + 0.011733653955161572, + 0.009607772342860699, + 0.018029242753982544, + 0.017162874341011047, + 0.029086945578455925, + 0.2821347117424011, + 0.28267908096313477, + 0.0009170564590021968, + 0.0008884540875442326, + 0.0018936744891107082, + 0.001604428282007575, + 0.0006272316095419228, + 0.0008905320428311825, + 0.001263029989786446, + 0.0014517451636493206, + 0.0025340321008116007, + 0.0028201439417898655, + 0.003738142317160964, + 0.004933695308864117, + 0.00391044095158577, + 0.005592011846601963, + 0.005282572470605373, + 0.004472485277801752, + 0.00039766865666024387, + 0.00030994333792477846, + 0.0007225965964607894, + 0.000673041446134448, + 0.0003582575882319361, + 0.000414624490076676, + 0.00045102735748514533, + 0.0005861546960659325, + 0.0010009357938542962, + 0.0011909918393939734, + 0.002012509386986494, + 0.0015271292068064213, + 0.000798717956058681, + 0.0008509582839906216, + 0.0018408092437312007, + 0.00201158644631505, + 0.0014727793168276548, + 0.0012223615776747465, + 0.0018310487503185868, + 0.0026065013371407986, + 0.001439219107851386, + 0.0018488058121874928, + 0.0031311579514294863, + 0.0029379294719547033, + 0.0030095218680799007, + 0.0026710263919085264, + 0.002681097714230418, + 0.004144612234085798, + 0.003743459703400731, + 0.005716932937502861, + 0.04827949032187462, + 0.04996516555547714, + 0.0010348072974011302, + 0.0009636451140977442, + 0.0012473397655412555, + 0.0021114093251526356, + 0.0021204629447311163, + 0.0021833034697920084, + 0.0028038150630891323, + 0.002383976709097624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.0008186326595023274 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537590228021145 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537590228021145, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537590228021145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021530989557504654, + 0.1058327853679657, + 0.8057209253311157, + 0.00466712424531579, + 0.0048240600153803825, + 0.0020304007921367884, + 0.005662280134856701, + 0.007771695498377085, + 0.040248800069093704, + 0.0017109657637774944 + ], + "credit_map": [ + 0.001758136902935803, + 0.001965276664122939, + 0.002590777352452278, + 0.0030492611695080996, + 0.0018152249976992607, + 0.003355162451043725, + 0.0034457421861588955, + 0.0035514070186764, + 0.00292330514639616, + 0.0040216906927526, + 0.0041282749734818935, + 0.005304405465722084, + 0.008140300400555134, + 0.008266693912446499, + 0.03629760444164276, + 0.03675049915909767, + 0.017163671553134918, + 0.014642146416008472, + 0.011710349470376968, + 0.022347120568156242, + 0.02105022221803665, + 0.03551168367266655, + 0.3414367139339447, + 0.34185898303985596, + 0.00033640608307905495, + 0.0006449129432439804, + 0.00046628364361822605, + 0.0004864325746893883, + 0.0006374214426614344, + 0.0008505639852955937, + 0.0006570337573066354, + 0.0005880694952793419, + 0.0004055771860294044, + 0.0009770282777026296, + 0.0007534843753091991, + 0.000529810378793627, + 0.0003924582852050662, + 0.0005739877233281732, + 0.0005552227958105505, + 0.0006364908185787499, + 0.00018230357090942562, + 0.00034956299350596964, + 0.00031640141969546676, + 0.00032278316211886704, + 0.00017079559620469809, + 0.000194505337276496, + 0.00023669075744692236, + 0.0002573578094597906, + 0.0004893678124062717, + 0.0004043117805849761, + 0.001327641773968935, + 0.0006573050050064921, + 0.0002779632923193276, + 0.00041017154580913484, + 0.0009748683078214526, + 0.0011206499766558409, + 0.0008756545139476657, + 0.0008199598523788154, + 0.0011285669170320034, + 0.0013011739356443286, + 0.0004932942101731896, + 0.0005602900055237114, + 0.001473058364354074, + 0.0011196971172466874, + 0.0012374675134196877, + 0.0009064540499821305, + 0.0009789462201297283, + 0.001486012595705688, + 0.0014286823570728302, + 0.0024066234473139048, + 0.015474787913262844, + 0.016329823061823845, + 0.00012820094707421958, + 0.00013342378952074796, + 0.00019625751883722842, + 0.00023500120732933283, + 0.00016720172425266355, + 0.00020480790408328176, + 0.0003265773120801896, + 0.0003194952732883394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.012537590228021145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024867109023034573 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.0024867109023034573, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024867109023034573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08846091479063034, + 0.30775660276412964, + 0.1410938799381256, + 0.0461936853826046, + 0.019897088408470154, + 0.01164545863866806, + 0.016102207824587822, + 0.107172891497612, + 0.2499455064535141, + 0.01173181738704443 + ], + "credit_map": [ + 0.006301260087639093, + 0.011630943045020103, + 0.012471327558159828, + 0.014386944472789764, + 0.009596374817192554, + 0.01154578197747469, + 0.011787150986492634, + 0.010741136036813259, + 0.011387255042791367, + 0.027188193053007126, + 0.022594986483454704, + 0.018627990037202835, + 0.031281255185604095, + 0.039961304515600204, + 0.07631940394639969, + 0.08039619773626328, + 0.0031167450360953808, + 0.002849510870873928, + 0.0027887742035090923, + 0.004372316412627697, + 0.00456230528652668, + 0.007137475069612265, + 0.058034125715494156, + 0.0582326278090477, + 0.0031934569124132395, + 0.007308971136808395, + 0.006091070827096701, + 0.005368489306420088, + 0.0032119974493980408, + 0.004310242366045713, + 0.007870031520724297, + 0.008839420974254608, + 0.0019255938241258264, + 0.001826845807954669, + 0.0028176680207252502, + 0.0026336917653679848, + 0.0020042648538947105, + 0.003350937971845269, + 0.0028373205568641424, + 0.002500764327123761, + 0.0008493513450957835, + 0.0014275979483500123, + 0.0014952635392546654, + 0.0011764891678467393, + 0.001223956816829741, + 0.001769870868884027, + 0.0017731937114149332, + 0.0019297351827844977, + 0.0011283401399850845, + 0.0021963289473205805, + 0.0019101443467661738, + 0.0015010667266324162, + 0.001433568075299263, + 0.00169238424859941, + 0.003138137049973011, + 0.003102237591519952, + 0.01240074634552002, + 0.013366231694817543, + 0.01465116161853075, + 0.013887656852602959, + 0.007635494694113731, + 0.008904930204153061, + 0.016976725310087204, + 0.019349947571754456, + 0.005696062929928303, + 0.005868936888873577, + 0.00714926328510046, + 0.00938617903739214, + 0.007823845371603966, + 0.012270165607333183, + 0.0988859087228775, + 0.10286515206098557, + 0.0011445041745901108, + 0.0008787905680947006, + 0.0015699538635089993, + 0.001956653082743287, + 0.0011426964774727821, + 0.0016625906573608518, + 0.0017856310587376356, + 0.0015909973299130797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.0024867109023034573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733347196131945 + ], + "opposite_logprob_tokens": [ + -5.006733417510986 + ], + "logprob_reward": -0.006733347196131945, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733347196131945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10569792240858078, + 0.07435078173875809, + 0.355533242225647, + 0.030454561114311218, + 0.029833799228072166, + 0.01189764216542244, + 0.02136717177927494, + 0.11210045963525772, + 0.24378062784671783, + 0.014983809553086758 + ], + "credit_map": [ + 0.008897755295038223, + 0.010378474369645119, + 0.017610689625144005, + 0.014383381232619286, + 0.008606035262346268, + 0.013633053749799728, + 0.016681794077157974, + 0.015506738796830177, + 0.0049383267760276794, + 0.008683539927005768, + 0.009062264114618301, + 0.006573859602212906, + 0.0065009803511202335, + 0.008889795280992985, + 0.014692778699100018, + 0.015009243972599506, + 0.007454726379364729, + 0.00685007544234395, + 0.00611857557669282, + 0.010012208484113216, + 0.009159247390925884, + 0.015665413811802864, + 0.15009239315986633, + 0.15018059313297272, + 0.003046896541491151, + 0.004545288626104593, + 0.005724935792386532, + 0.003927291836589575, + 0.0023167654871940613, + 0.0025289072655141354, + 0.003976147621870041, + 0.004388328641653061, + 0.003335012588649988, + 0.002985288156196475, + 0.0056019918993115425, + 0.003719095606356859, + 0.0031185136176645756, + 0.00351890386082232, + 0.003461573040112853, + 0.004093419294804335, + 0.0012507427018135786, + 0.0012132440460845828, + 0.001919255591928959, + 0.001332107000052929, + 0.001310843857936561, + 0.001814092742279172, + 0.0015884687891229987, + 0.001468887086957693, + 0.0016634612111374736, + 0.0026396596804261208, + 0.0019668478053063154, + 0.0016658575041219592, + 0.00297130411490798, + 0.0032850687857717276, + 0.0035682981833815575, + 0.003606675425544381, + 0.011600763536989689, + 0.019601084291934967, + 0.014428224414587021, + 0.012329649180173874, + 0.009240052662789822, + 0.011652182787656784, + 0.014776776544749737, + 0.0184717308729887, + 0.005601434502750635, + 0.00593833951279521, + 0.007912290282547474, + 0.00877598486840725, + 0.007453334983438253, + 0.012093428522348404, + 0.09613648802042007, + 0.09986932575702667, + 0.0012876963010057807, + 0.0016083433292806149, + 0.0015326915308833122, + 0.002027429174631834, + 0.001237348304130137, + 0.0022708033211529255, + 0.0026016829069703817, + 0.0024178153835237026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.006733347196131945 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, framed from a fixed low-side angle, captures a scene of profound tranquility that embodies traditional Japanese aesthetics. In the background, water cascades in a gentle, continuous stream, flowing down a dark, damp rock wall. The rock surface is almost completely covered by a thick, velvety carpet of deep green moss that appears saturated with moisture. In the foreground, a sturdy fern serves as the central subject, its dark, gnarled base firmly planted. From its core, a magnificent crown of large, feathery fronds unfolds, each boasting a striking copper-red hue that contrasts strikingly with the earthy tones of the scene. A fine, ethereal mist, created by the gentle rush of the waterfall, drifts through the air and delicately settles on the fern's fronds, causing countless tiny droplets to cling to the edges of each leaf, shimmering like miniature jewels. The entire composition is bathed in a soft, diffused, cool-toned light that enhances the rich textures of the wet stone and delicate foliage, while the minimalist composition creates a serene, contemplative atmosphere, finding a quiet, fleeting beauty in this perfectly balanced natural moment.", + "seed": 28511, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_3/index_706_seed_28511_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008321292395703495, + "vlm_logprob": -0.008321292395703495, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.008321292395703495, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024721547961235046 + ], + "opposite_logprob_tokens": [ + -4.149721622467041 + ], + "logprob_reward": -0.024721547961235046, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.024721547961235046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14164642989635468, + 0.18900658190250397, + 0.09560174494981766, + 0.1228644847869873, + 0.11428651213645935, + 0.09728585928678513, + 0.060351770371198654, + 0.06326998025178909, + 0.08737047016620636, + 0.028316203504800797 + ], + "credit_map": [ + 0.015890298411250114, + 0.016799703240394592, + 0.022842124104499817, + 0.015570137649774551, + 0.011230623349547386, + 0.011238356120884418, + 0.02248493768274784, + 0.02559024840593338, + 0.020437069237232208, + 0.032316774129867554, + 0.028950924053788185, + 0.029964428395032883, + 0.011064678430557251, + 0.010217089205980301, + 0.02626270055770874, + 0.029792921617627144, + 0.007790062576532364, + 0.012866801582276821, + 0.010811532847583294, + 0.01449198741465807, + 0.010072189383208752, + 0.00942664872854948, + 0.015591926872730255, + 0.014550586231052876, + 0.017253436148166656, + 0.014018241316080093, + 0.018574276939034462, + 0.015651965513825417, + 0.008653363212943077, + 0.00910690426826477, + 0.021068274974822998, + 0.01853802241384983, + 0.011733464896678925, + 0.010391387157142162, + 0.018544375896453857, + 0.01622791774570942, + 0.009643729776144028, + 0.01216133777052164, + 0.015240898355841637, + 0.020343385636806488, + 0.009639227762818336, + 0.010085529647767544, + 0.015831464901566505, + 0.01270235888659954, + 0.0071559567004442215, + 0.007476283702999353, + 0.017402542755007744, + 0.016992487013339996, + 0.007405645679682493, + 0.005056473426520824, + 0.014692751690745354, + 0.0056260074488818645, + 0.0030132948886603117, + 0.00486957048997283, + 0.010742764919996262, + 0.008945252746343613, + 0.005316164810210466, + 0.0030441114213317633, + 0.004913961980491877, + 0.004355226643383503, + 0.005216976162046194, + 0.005198400933295488, + 0.017476050183176994, + 0.017749078571796417, + 0.01471053995192051, + 0.007607499603182077, + 0.012251253239810467, + 0.011571521870791912, + 0.007941467687487602, + 0.008514150977134705, + 0.01364416629076004, + 0.011129859834909439, + 0.004269447643309832, + 0.001860931166447699, + 0.005013544578105211, + 0.004290398675948381, + 0.0020413692109286785, + 0.002331256400793791, + 0.00471434136852622, + 0.0037949136458337307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.024721547961235046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258894059807062 + ], + "opposite_logprob_tokens": [ + -5.255259037017822 + ], + "logprob_reward": -0.005258894059807062, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258894059807062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1800168752670288, + 0.12305829674005508, + 0.03724398463964462, + 0.06483094394207001, + 0.03718065842986107, + 0.14747601747512817, + 0.03012884221971035, + 0.21037794649600983, + 0.098890520632267, + 0.07079587876796722 + ], + "credit_map": [ + 0.017470702528953552, + 0.02107284776866436, + 0.027323395013809204, + 0.019436292350292206, + 0.01789459027349949, + 0.013726003468036652, + 0.028772378340363503, + 0.03432068973779678, + 0.010892353951931, + 0.018671900033950806, + 0.01779239997267723, + 0.017254140228033066, + 0.01270356122404337, + 0.013928718864917755, + 0.013708733022212982, + 0.018106499686837196, + 0.003493562573567033, + 0.0049582310020923615, + 0.005925567355006933, + 0.00520451320335269, + 0.004050269722938538, + 0.003103514201939106, + 0.0051050325855612755, + 0.005403297021985054, + 0.007170894183218479, + 0.008633984252810478, + 0.008740300312638283, + 0.007081765215843916, + 0.004812199622392654, + 0.006174611859023571, + 0.01188320480287075, + 0.010333982296288013, + 0.0034124322701245546, + 0.0030097845010459423, + 0.0069929310120642185, + 0.004470158368349075, + 0.003924892749637365, + 0.004272813443094492, + 0.004587362986057997, + 0.006510287988930941, + 0.02013818547129631, + 0.016066066920757294, + 0.018360329791903496, + 0.011218137107789516, + 0.010939781554043293, + 0.008916649036109447, + 0.03527883067727089, + 0.026558049023151398, + 0.00333077902905643, + 0.002390514127910137, + 0.0043425122275948524, + 0.0028430435340851545, + 0.0028589058201760054, + 0.0032423832453787327, + 0.006180175114423037, + 0.004940533544868231, + 0.011776464991271496, + 0.009462638758122921, + 0.007084631361067295, + 0.007450055796653032, + 0.0073743341490626335, + 0.013001227751374245, + 0.08540861308574677, + 0.06881999969482422, + 0.01302495226264, + 0.008146539330482483, + 0.01867849938571453, + 0.010686208494007587, + 0.010024413466453552, + 0.010219116695225239, + 0.014253838919103146, + 0.013856956735253334, + 0.009632056578993797, + 0.005045978352427483, + 0.00972555112093687, + 0.007038906216621399, + 0.008033601567149162, + 0.006857094820588827, + 0.014268430881202221, + 0.010194266214966774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.005258894059807062 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820087829604745 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.002820087829604745, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820087829604745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22868354618549347, + 0.09988375753164291, + 0.10750354826450348, + 0.05283956602215767, + 0.04822373762726784, + 0.1465848833322525, + 0.09912828356027603, + 0.0821356251835823, + 0.07995820790529251, + 0.055058803409338 + ], + "credit_map": [ + 0.024632807821035385, + 0.023410364985466003, + 0.0464291013777256, + 0.025279689580202103, + 0.023245053365826607, + 0.022551044821739197, + 0.027847323566675186, + 0.03528815507888794, + 0.008461469784379005, + 0.011550860479474068, + 0.012795886024832726, + 0.011550496332347393, + 0.012911549769341946, + 0.014260894618928432, + 0.01376653928309679, + 0.01458605658262968, + 0.010562070645391941, + 0.010222322307527065, + 0.014224695041775703, + 0.011203989386558533, + 0.014033547602593899, + 0.016985604539513588, + 0.01492479257285595, + 0.0153465261682868, + 0.006050179246813059, + 0.00519887451082468, + 0.011056315153837204, + 0.005442796275019646, + 0.00480605335906148, + 0.006081425119191408, + 0.005927105899900198, + 0.008276817388832569, + 0.005458547733724117, + 0.004861455876380205, + 0.009395016357302666, + 0.005646317265927792, + 0.004634773824363947, + 0.0058004120364785194, + 0.004938037134706974, + 0.0074891760013997555, + 0.013319894671440125, + 0.011815031990408897, + 0.018935998901724815, + 0.012183260172605515, + 0.019291363656520844, + 0.02304619364440441, + 0.022520774975419044, + 0.02547236531972885, + 0.008458517491817474, + 0.009748798795044422, + 0.012903408147394657, + 0.0066282604821026325, + 0.012455897405743599, + 0.013243692927062511, + 0.020679857581853867, + 0.01500985398888588, + 0.0108413090929389, + 0.007108673918992281, + 0.00950234942138195, + 0.008761036209762096, + 0.009587186388671398, + 0.009982317686080933, + 0.013826059177517891, + 0.012526686303317547, + 0.009145869873464108, + 0.008639106526970863, + 0.008861669339239597, + 0.010817297734320164, + 0.008449174463748932, + 0.009462140500545502, + 0.015111512504518032, + 0.00947143230587244, + 0.009680218063294888, + 0.004069904331117868, + 0.008934149518609047, + 0.005841147620230913, + 0.005363150965422392, + 0.0055673872120678425, + 0.007850914262235165, + 0.00775193003937602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.002820087829604745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178047270514071 + ], + "opposite_logprob_tokens": [ + -6.75117826461792 + ], + "logprob_reward": -0.001178047270514071, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178047270514071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18779493868350983, + 0.11119271069765091, + 0.08823816478252411, + 0.05913932994008064, + 0.12626416981220245, + 0.13138681650161743, + 0.04845490679144859, + 0.07222530245780945, + 0.1328667402267456, + 0.04243691265583038 + ], + "credit_map": [ + 0.019930390641093254, + 0.019898438826203346, + 0.02916751801967621, + 0.01925312541425228, + 0.021368516609072685, + 0.011501604691147804, + 0.029259871691465378, + 0.037415456026792526, + 0.009488841518759727, + 0.01386557798832655, + 0.01866140030324459, + 0.0171795766800642, + 0.010236694477498531, + 0.009058930911123753, + 0.015746045857667923, + 0.01695564202964306, + 0.009452429600059986, + 0.009485793299973011, + 0.013876003213226795, + 0.010843922384083271, + 0.009751868434250355, + 0.008169883862137794, + 0.01316160336136818, + 0.013496661558747292, + 0.00742762815207243, + 0.00801033340394497, + 0.011401867493987083, + 0.0064284005202353, + 0.004405922722071409, + 0.004924368113279343, + 0.006353910081088543, + 0.010186899453401566, + 0.014292431063950062, + 0.014150621369481087, + 0.02485254406929016, + 0.02001281827688217, + 0.007386333774775267, + 0.008449776098132133, + 0.01707659289240837, + 0.02004305087029934, + 0.01986416056752205, + 0.015978069975972176, + 0.020804792642593384, + 0.01641070656478405, + 0.00620294502004981, + 0.005443235859274864, + 0.026133297011256218, + 0.020549608394503593, + 0.006204310804605484, + 0.0031471415422856808, + 0.008320685476064682, + 0.004671473056077957, + 0.004157106392085552, + 0.004756057169288397, + 0.009903154335916042, + 0.007294976152479649, + 0.009224240668118, + 0.005789064336568117, + 0.009295807220041752, + 0.00678670359775424, + 0.005365491379052401, + 0.005599044729024172, + 0.014748293906450272, + 0.015416659414768219, + 0.03162446990609169, + 0.007698675151914358, + 0.029883746057748795, + 0.0150193702429533, + 0.005178641062229872, + 0.006716722156852484, + 0.016734067350625992, + 0.02001105435192585, + 0.0065903631038963795, + 0.003055212087929249, + 0.00764081533998251, + 0.006591256707906723, + 0.0029124142602086067, + 0.0038189697079360485, + 0.006665337830781937, + 0.0051625450141727924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.001178047270514071 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627884857356548 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007627884857356548, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627884857356548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23867353796958923, + 0.0905146598815918, + 0.06778421252965927, + 0.07869835197925568, + 0.06839507818222046, + 0.04735986143350601, + 0.07550277560949326, + 0.13248886168003082, + 0.14970256388187408, + 0.050880078226327896 + ], + "credit_map": [ + 0.023437434807419777, + 0.028091508895158768, + 0.04302886128425598, + 0.031352948397397995, + 0.023241354152560234, + 0.014640958979725838, + 0.0340898372232914, + 0.04079065099358559, + 0.009810218587517738, + 0.00759790139272809, + 0.011402074247598648, + 0.010637150146067142, + 0.01162035670131445, + 0.01025457214564085, + 0.01340417843312025, + 0.015788208693265915, + 0.006628178525716066, + 0.007813071832060814, + 0.010377181693911552, + 0.00961965974420309, + 0.007795429322868586, + 0.0054472642950713634, + 0.010217416100203991, + 0.009886007755994797, + 0.008288085460662842, + 0.009407907724380493, + 0.010469465516507626, + 0.009614056907594204, + 0.006991672795265913, + 0.007895592600107193, + 0.013837900944054127, + 0.012193674221634865, + 0.007914392277598381, + 0.007434690836817026, + 0.013881037943065166, + 0.00861146580427885, + 0.004813340492546558, + 0.005008267238736153, + 0.00892309658229351, + 0.0118087874725461, + 0.006851670797914267, + 0.0069306716322898865, + 0.00603462615981698, + 0.005259667057543993, + 0.004852034617215395, + 0.005301500204950571, + 0.006096496246755123, + 0.006033194251358509, + 0.009622793644666672, + 0.0044791204854846, + 0.01275133341550827, + 0.00511031411588192, + 0.006304789800196886, + 0.006711764261126518, + 0.016677124425768852, + 0.013845536857843399, + 0.031437672674655914, + 0.008870108053088188, + 0.024933738633990288, + 0.014562057331204414, + 0.00725152250379324, + 0.008414818905293941, + 0.01804352179169655, + 0.018975429236888885, + 0.03502301499247551, + 0.009284019470214844, + 0.033207688480615616, + 0.017300738021731377, + 0.005976441316306591, + 0.008980609476566315, + 0.018401246517896652, + 0.02152879163622856, + 0.00793858990073204, + 0.0034908598754554987, + 0.009122608229517937, + 0.006751000881195068, + 0.003610291751101613, + 0.005171719007194042, + 0.0077483439818024635, + 0.007046665996313095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.007627884857356548 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, framed from a fixed low-side angle, captures a scene of profound tranquility that embodies traditional Japanese aesthetics. In the background, water cascades in a gentle, continuous stream, flowing down a dark, damp rock wall. The rock surface is almost completely covered by a thick, velvety carpet of deep green moss that appears saturated with moisture. In the foreground, a sturdy fern serves as the central subject, its dark, gnarled base firmly planted. From its core, a magnificent crown of large, feathery fronds unfolds, each boasting a striking copper-red hue that contrasts strikingly with the earthy tones of the scene. A fine, ethereal mist, created by the gentle rush of the waterfall, drifts through the air and delicately settles on the fern's fronds, causing countless tiny droplets to cling to the edges of each leaf, shimmering like miniature jewels. The entire composition is bathed in a soft, diffused, cool-toned light that enhances the rich textures of the wet stone and delicate foliage, while the minimalist composition creates a serene, contemplative atmosphere, finding a quiet, fleeting beauty in this perfectly balanced natural moment.", + "seed": 28512, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_3/index_706_seed_28512_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016290637478232384, + "vlm_logprob": -0.016290637478232384, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.016290637478232384, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036111246794462204 + ], + "opposite_logprob_tokens": [ + -3.661111354827881 + ], + "logprob_reward": -0.036111246794462204, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.036111246794462204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24752971529960632, + 0.13433942198753357, + 0.05737939476966858, + 0.13272559642791748, + 0.056649692356586456, + 0.09443111717700958, + 0.07667267322540283, + 0.058906275779008865, + 0.0665360689163208, + 0.07483001798391342 + ], + "credit_map": [ + 0.030506866052746773, + 0.0440017394721508, + 0.02652852050960064, + 0.02878379449248314, + 0.02979547157883644, + 0.02527967095375061, + 0.028108812868595123, + 0.034524835646152496, + 0.013892920687794685, + 0.024215856567025185, + 0.015504235401749611, + 0.02182921953499317, + 0.012984402477741241, + 0.01117541640996933, + 0.015981106087565422, + 0.018756264820694923, + 0.006945905275642872, + 0.0061524296179413795, + 0.010109484195709229, + 0.006194249726831913, + 0.0048305317759513855, + 0.00731167895719409, + 0.0065122852101922035, + 0.009322828613221645, + 0.013927646912634373, + 0.01429972518235445, + 0.019608192145824432, + 0.01750320754945278, + 0.020405082032084465, + 0.013347303494811058, + 0.018739141523838043, + 0.014895309694111347, + 0.005841408856213093, + 0.0059029399417340755, + 0.006914025638252497, + 0.007903315126895905, + 0.008324665017426014, + 0.006817689631134272, + 0.008076676167547703, + 0.006868972443044186, + 0.011943202465772629, + 0.011319011449813843, + 0.012493274174630642, + 0.012738535180687904, + 0.010851645842194557, + 0.010064451023936272, + 0.014498175121843815, + 0.01052282378077507, + 0.006936787161976099, + 0.009322194382548332, + 0.009025711566209793, + 0.012450885027647018, + 0.008171943947672844, + 0.009645112790167332, + 0.011630098335444927, + 0.009489939548075199, + 0.00535900192335248, + 0.00690431660041213, + 0.0073569160886108875, + 0.011231880635023117, + 0.005707853008061647, + 0.006884843576699495, + 0.0064253737218678, + 0.009036093018949032, + 0.006689003668725491, + 0.012947055511176586, + 0.008856362663209438, + 0.010535522364079952, + 0.005429133772850037, + 0.007793780881911516, + 0.007007467094808817, + 0.007277742959558964, + 0.005701021291315556, + 0.013160490430891514, + 0.0064973775297403336, + 0.010922905057668686, + 0.008800637908279896, + 0.008633585646748543, + 0.010196151211857796, + 0.010917851701378822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.036111246794462204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007641489617526531 + ], + "opposite_logprob_tokens": [ + -4.882641315460205 + ], + "logprob_reward": -0.007641489617526531, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641489617526531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10523177683353424, + 0.12014135718345642, + 0.12041220813989639, + 0.05500950291752815, + 0.11492352187633514, + 0.07043805718421936, + 0.051615990698337555, + 0.06997781991958618, + 0.08657438308000565, + 0.2056753784418106 + ], + "credit_map": [ + 0.012128546833992004, + 0.014659959822893143, + 0.012761594727635384, + 0.013319574296474457, + 0.011950149200856686, + 0.009943273849785328, + 0.017016597092151642, + 0.0134520772844553, + 0.017375919967889786, + 0.0115971639752388, + 0.02136351726949215, + 0.01694709062576294, + 0.010017829947173595, + 0.014485122635960579, + 0.012498506344854832, + 0.015856202691793442, + 0.011699545197188854, + 0.014132923446595669, + 0.020130421966314316, + 0.014723462983965874, + 0.012707463465631008, + 0.012689740397036076, + 0.01619468815624714, + 0.018133964389562607, + 0.00462896516546607, + 0.007222676184028387, + 0.0060295360162854195, + 0.00662459060549736, + 0.009220666252076626, + 0.006894286721944809, + 0.007666897028684616, + 0.006721882149577141, + 0.008546872064471245, + 0.014613443054258823, + 0.011653752066195011, + 0.01781289093196392, + 0.015782155096530914, + 0.014533701352775097, + 0.018272006884217262, + 0.013708696700632572, + 0.00551843736320734, + 0.007872658781707287, + 0.005969170946627855, + 0.013682126067578793, + 0.00880384910851717, + 0.008035225793719292, + 0.01198569219559431, + 0.0085708973929286, + 0.004480937495827675, + 0.005798483267426491, + 0.005982214119285345, + 0.009284482337534428, + 0.006151488982141018, + 0.005529716610908508, + 0.007573803421109915, + 0.006814863998442888, + 0.005153764970600605, + 0.008112121373414993, + 0.008912389166653156, + 0.01217504683881998, + 0.007954511791467667, + 0.007325614336878061, + 0.01044509932398796, + 0.00989927165210247, + 0.006131011061370373, + 0.014963350258767605, + 0.009843354113399982, + 0.015263731591403484, + 0.007847839035093784, + 0.012337598949670792, + 0.00994971301406622, + 0.010237784124910831, + 0.021552402526140213, + 0.03062058426439762, + 0.022073151543736458, + 0.029873281717300415, + 0.02212701551616192, + 0.016332555562257767, + 0.029007676988840103, + 0.03408871591091156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.007641489617526531 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812784865498543 + ], + "opposite_logprob_tokens": [ + -3.529812812805176 + ], + "logprob_reward": -0.029812784865498543, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812784865498543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16217339038848877, + 0.07511024177074432, + 0.05676392838358879, + 0.1065857782959938, + 0.08650509268045425, + 0.0823373943567276, + 0.11009034514427185, + 0.1220489889383316, + 0.11723211407661438, + 0.08115273714065552 + ], + "credit_map": [ + 0.023965513333678246, + 0.017340553924441338, + 0.023447517305612564, + 0.019957728683948517, + 0.018838513642549515, + 0.021330781280994415, + 0.015783073380589485, + 0.021509697660803795, + 0.010449389927089214, + 0.006127370521426201, + 0.012265831232070923, + 0.009455163963139057, + 0.00798861589282751, + 0.009860306046903133, + 0.010540230199694633, + 0.00842333398759365, + 0.005876812618225813, + 0.007985313422977924, + 0.006820759270340204, + 0.006994721945375204, + 0.008155185729265213, + 0.007187371142208576, + 0.006702938582748175, + 0.007040828932076693, + 0.008524742908775806, + 0.012277388945221901, + 0.010373977944254875, + 0.013405995443463326, + 0.017207041382789612, + 0.018040435388684273, + 0.01200920157134533, + 0.014746994711458683, + 0.006592729594558477, + 0.010656308382749557, + 0.007530359551310539, + 0.011169306002557278, + 0.012769056484103203, + 0.013466582633554935, + 0.012205976992845535, + 0.012114773504436016, + 0.005358440335839987, + 0.008654121309518814, + 0.007485260255634785, + 0.01475713774561882, + 0.012368195690214634, + 0.008837399072945118, + 0.015120661817491055, + 0.009756186045706272, + 0.009546206332743168, + 0.01298679318279028, + 0.013184920884668827, + 0.018996072933077812, + 0.014097308740019798, + 0.011253837496042252, + 0.016172148287296295, + 0.01385305356234312, + 0.010304315946996212, + 0.012392432428896427, + 0.017428994178771973, + 0.023638594895601273, + 0.0129005815833807, + 0.014238843694329262, + 0.015346349216997623, + 0.01579887606203556, + 0.009482164867222309, + 0.023801812902092934, + 0.01215853076428175, + 0.018698977306485176, + 0.01196456328034401, + 0.01775425858795643, + 0.01043605525046587, + 0.012935753911733627, + 0.011149104684591293, + 0.007543531246483326, + 0.008945850655436516, + 0.017071228474378586, + 0.008700278587639332, + 0.006647859234362841, + 0.010537760332226753, + 0.010557127185165882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.029812784865498543 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001940036192536354 + ], + "opposite_logprob_tokens": [ + -6.2519402503967285 + ], + "logprob_reward": -0.001940036192536354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001940036192536354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16276384890079498, + 0.4326886832714081, + 0.076542928814888, + 0.05657684803009033, + 0.028399739414453506, + 0.05171814560890198, + 0.057665176689624786, + 0.027413446456193924, + 0.03915433585643768, + 0.06707679480314255 + ], + "credit_map": [ + 0.014364921487867832, + 0.021192999556660652, + 0.019255319610238075, + 0.021240368485450745, + 0.023277992382645607, + 0.019631605595350266, + 0.022281015291810036, + 0.021519623696804047, + 0.06317438930273056, + 0.03223421797156334, + 0.07936577498912811, + 0.07213420420885086, + 0.03243521600961685, + 0.04419467970728874, + 0.0437595471739769, + 0.06539065390825272, + 0.009265996515750885, + 0.007262942846864462, + 0.013707650825381279, + 0.01164720207452774, + 0.007124980911612511, + 0.007456122897565365, + 0.009116437286138535, + 0.01096159778535366, + 0.004908858332782984, + 0.006121509242802858, + 0.006575633771717548, + 0.008471600711345673, + 0.0063054850324988365, + 0.006523423828184605, + 0.009324248880147934, + 0.008346090093255043, + 0.003161683678627014, + 0.0032586809247732162, + 0.003629374783486128, + 0.0036254164297133684, + 0.0026165656745433807, + 0.003929599653929472, + 0.0032653457019478083, + 0.004913073033094406, + 0.003591267392039299, + 0.00835241749882698, + 0.0034308775793761015, + 0.007669960614293814, + 0.007502232678234577, + 0.006706681102514267, + 0.00826648622751236, + 0.006198222748935223, + 0.005460911430418491, + 0.005779963918030262, + 0.006800386123359203, + 0.009961795061826706, + 0.006428221706300974, + 0.006453896407037973, + 0.009314922615885735, + 0.007465078961104155, + 0.0025377904530614614, + 0.0031633018516004086, + 0.0035457652993500233, + 0.004902239888906479, + 0.0027667211834341288, + 0.003361079376190901, + 0.0033145516645163298, + 0.0038219948764890432, + 0.002954502822831273, + 0.004590572789311409, + 0.003927306737750769, + 0.0062442561611533165, + 0.004813226405531168, + 0.00583733431994915, + 0.0056485868990421295, + 0.005138549953699112, + 0.006651898380368948, + 0.006164032965898514, + 0.0067995162680745125, + 0.012394058518111706, + 0.007501149084419012, + 0.005771162454038858, + 0.011210529133677483, + 0.010584449395537376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.001940036192536354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947629921138287 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947629921138287, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947629921138287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13252606987953186, + 0.1621783971786499, + 0.06485264748334885, + 0.09524258971214294, + 0.07418053597211838, + 0.06909523904323578, + 0.23337410390377045, + 0.03868841752409935, + 0.07127758860588074, + 0.05858441814780235 + ], + "credit_map": [ + 0.013477613218128681, + 0.018949264660477638, + 0.023566871881484985, + 0.01899723894894123, + 0.01372471358627081, + 0.012200010009109974, + 0.01594064198434353, + 0.01566971279680729, + 0.01640353351831436, + 0.016824930906295776, + 0.01730080507695675, + 0.02962743304669857, + 0.018932346254587173, + 0.017734086140990257, + 0.026173001155257225, + 0.01918225921690464, + 0.0080806203186512, + 0.0067402455024421215, + 0.009519499726593494, + 0.010364056564867496, + 0.007165861316025257, + 0.005771693307906389, + 0.00894829723984003, + 0.008262374438345432, + 0.009605213068425655, + 0.011987860314548016, + 0.009002738632261753, + 0.016431797295808792, + 0.010859225876629353, + 0.009373060427606106, + 0.014372887089848518, + 0.013609804213047028, + 0.007278870791196823, + 0.01025314535945654, + 0.006579046603292227, + 0.010487590916454792, + 0.00879779551178217, + 0.010015303269028664, + 0.010577651672065258, + 0.010191128589212894, + 0.0052981008775532246, + 0.01095170620828867, + 0.005790946073830128, + 0.009005334228277206, + 0.011120801791548729, + 0.00834699161350727, + 0.011216088198125362, + 0.007365273777395487, + 0.019196219742298126, + 0.01950562745332718, + 0.04385010525584221, + 0.04801615700125694, + 0.016637438908219337, + 0.019928565248847008, + 0.03663405030965805, + 0.02960592322051525, + 0.0030593271367251873, + 0.004617432132363319, + 0.006439062301069498, + 0.006405858788639307, + 0.0034849378280341625, + 0.003829684341326356, + 0.004699280485510826, + 0.00615283427760005, + 0.007852385751903057, + 0.010120775550603867, + 0.00778636708855629, + 0.00900642853230238, + 0.008730034343898296, + 0.009516800753772259, + 0.00788789987564087, + 0.010376890189945698, + 0.007771608419716358, + 0.008563114330172539, + 0.00680639548227191, + 0.008746556006371975, + 0.005557244177907705, + 0.005280188750475645, + 0.008329413831233978, + 0.007529899012297392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.005947629921138287 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, framed from a fixed low-side angle, captures a scene of profound tranquility that embodies traditional Japanese aesthetics. In the background, water cascades in a gentle, continuous stream, flowing down a dark, damp rock wall. The rock surface is almost completely covered by a thick, velvety carpet of deep green moss that appears saturated with moisture. In the foreground, a sturdy fern serves as the central subject, its dark, gnarled base firmly planted. From its core, a magnificent crown of large, feathery fronds unfolds, each boasting a striking copper-red hue that contrasts strikingly with the earthy tones of the scene. A fine, ethereal mist, created by the gentle rush of the waterfall, drifts through the air and delicately settles on the fern's fronds, causing countless tiny droplets to cling to the edges of each leaf, shimmering like miniature jewels. The entire composition is bathed in a soft, diffused, cool-toned light that enhances the rich textures of the wet stone and delicate foliage, while the minimalist composition creates a serene, contemplative atmosphere, finding a quiet, fleeting beauty in this perfectly balanced natural moment.", + "seed": 28513, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_3/index_706_seed_28513_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.011256627459079026, + "vlm_logprob": -0.011256627459079026, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.011256627459079026, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02580985799431801 + ], + "opposite_logprob_tokens": [ + -4.0258097648620605 + ], + "logprob_reward": -0.02580985799431801, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02580985799431801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.306961327791214, + 0.11761415749788284, + 0.10137134045362473, + 0.0981481745839119, + 0.03985513746738434, + 0.11828215420246124, + 0.055088091641664505, + 0.06273669749498367, + 0.06776511669158936, + 0.032177794724702835 + ], + "credit_map": [ + 0.039242397993803024, + 0.049573179334402084, + 0.04380696266889572, + 0.039676010608673096, + 0.030122991651296616, + 0.03394514322280884, + 0.035637758672237396, + 0.03495687246322632, + 0.011236930266022682, + 0.013511337339878082, + 0.013849008828401566, + 0.011718920432031155, + 0.013549872674047947, + 0.020674431696534157, + 0.017261851578950882, + 0.015811791643500328, + 0.01461299229413271, + 0.006784519646316767, + 0.02637457847595215, + 0.016136839985847473, + 0.006282043177634478, + 0.007287201005965471, + 0.011251332238316536, + 0.012641830369830132, + 0.011719465255737305, + 0.008319946005940437, + 0.01016985159367323, + 0.012849210761487484, + 0.01369666587561369, + 0.012092771008610725, + 0.017723100259900093, + 0.01157715916633606, + 0.005872043780982494, + 0.0034525964874774218, + 0.0034743668511509895, + 0.005556576885282993, + 0.004326694644987583, + 0.005119315814226866, + 0.006208187434822321, + 0.005845349747687578, + 0.015350566245615482, + 0.009599384851753712, + 0.019080646336078644, + 0.017034444957971573, + 0.009583968669176102, + 0.01293589174747467, + 0.019763262942433357, + 0.014933980070054531, + 0.006425968371331692, + 0.005644671153277159, + 0.007293825503438711, + 0.008720474317669868, + 0.006745895836502314, + 0.0057987030595541, + 0.0075078061781823635, + 0.006950743030756712, + 0.006600275635719299, + 0.010419376194477081, + 0.005676968488842249, + 0.006550267338752747, + 0.008359421975910664, + 0.009647811762988567, + 0.008519203402101994, + 0.006963368039578199, + 0.006983014754951, + 0.010199958458542824, + 0.0059602102264761925, + 0.006489661522209644, + 0.007946905680000782, + 0.014779256656765938, + 0.008376630023121834, + 0.007029476575553417, + 0.003475086996331811, + 0.0036794529296457767, + 0.0044609070755541325, + 0.00490749254822731, + 0.0033239773474633694, + 0.0037027543876320124, + 0.0038834223523736, + 0.004744695499539375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.02580985799431801 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791442193090916 + ], + "opposite_logprob_tokens": [ + -4.634791374206543 + ], + "logprob_reward": -0.009791442193090916, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791442193090916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34629562497138977, + 0.2989967167377472, + 0.06824468076229095, + 0.03808780759572983, + 0.07514301687479019, + 0.03964608907699585, + 0.011965367943048477, + 0.053337033838033676, + 0.04942174628376961, + 0.018861928954720497 + ], + "credit_map": [ + 0.03792179375886917, + 0.038648735731840134, + 0.049778666347265244, + 0.04785161837935448, + 0.036708470433950424, + 0.047915130853652954, + 0.04701996594667435, + 0.04045123606920242, + 0.03113635443150997, + 0.02657376229763031, + 0.038171637803316116, + 0.03372074291110039, + 0.03714167699217796, + 0.06181137636303902, + 0.036069706082344055, + 0.03437145799398422, + 0.007505761459469795, + 0.005382638890296221, + 0.011403363198041916, + 0.00952948722988367, + 0.005531898234039545, + 0.010900554247200489, + 0.009213955141603947, + 0.008777019567787647, + 0.003945532720535994, + 0.0026614805683493614, + 0.004616733640432358, + 0.004775729496032, + 0.005169762764126062, + 0.006376557983458042, + 0.0059633031487464905, + 0.004578706808388233, + 0.007559996098279953, + 0.005389893427491188, + 0.009043285623192787, + 0.012405328452587128, + 0.00904865749180317, + 0.009569409303367138, + 0.011780734173953533, + 0.01034571323543787, + 0.003962118644267321, + 0.003860631724819541, + 0.004368769004940987, + 0.004746532533317804, + 0.005719948559999466, + 0.004848285112529993, + 0.006917960476130247, + 0.005221844185143709, + 0.0018245098181068897, + 0.0010956537444144487, + 0.0017218163702636957, + 0.0015032190131023526, + 0.0010980244260281324, + 0.0015420044073835015, + 0.0016945102252066135, + 0.0014856301713734865, + 0.006427144166082144, + 0.006987390108406544, + 0.005603027064353228, + 0.007524043321609497, + 0.006471564993262291, + 0.006388416513800621, + 0.007850523106753826, + 0.0060849240981042385, + 0.00441643176600337, + 0.004003283102065325, + 0.005497520323842764, + 0.005820775404572487, + 0.0065790400840342045, + 0.01039926428347826, + 0.007090667728334665, + 0.005614765454083681, + 0.00233148573897779, + 0.0017658945871517062, + 0.003603132674470544, + 0.0028773078229278326, + 0.0014505964936688542, + 0.0017941523110494018, + 0.0025193598121404648, + 0.0025199991650879383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.009791442193090916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254018023610115 + ], + "opposite_logprob_tokens": [ + -4.387540340423584 + ], + "logprob_reward": -0.01254018023610115, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254018023610115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12234555184841156, + 0.3967873752117157, + 0.1330847293138504, + 0.07056287676095963, + 0.06771406531333923, + 0.058477018028497696, + 0.03723706677556038, + 0.046941377222537994, + 0.05095740407705307, + 0.01589253544807434 + ], + "credit_map": [ + 0.016568053513765335, + 0.022093992680311203, + 0.014866631478071213, + 0.012325767427682877, + 0.01243414543569088, + 0.01780623570084572, + 0.011284875683486462, + 0.014965848065912724, + 0.04528301581740379, + 0.05790995433926582, + 0.04342420771718025, + 0.03804989159107208, + 0.040317680686712265, + 0.08737337589263916, + 0.041007354855537415, + 0.04342189431190491, + 0.015114272013306618, + 0.0198264941573143, + 0.011440548114478588, + 0.014631391502916813, + 0.014652631245553493, + 0.027439137920737267, + 0.0144887063652277, + 0.015491556376218796, + 0.009006821550428867, + 0.009365452453494072, + 0.005166944582015276, + 0.006756300572305918, + 0.008543318137526512, + 0.015719179064035416, + 0.008181370794773102, + 0.007823480293154716, + 0.007867319509387016, + 0.007932053878903389, + 0.006852386053651571, + 0.009206141345202923, + 0.007849162444472313, + 0.010586242191493511, + 0.00913904793560505, + 0.00828170869499445, + 0.005638075061142445, + 0.007791387382894754, + 0.004496037494391203, + 0.0055399490520358086, + 0.009144081734120846, + 0.009064438752830029, + 0.009226663038134575, + 0.007576383650302887, + 0.004166445229202509, + 0.003372807987034321, + 0.0068388753570616245, + 0.005651763640344143, + 0.003441320499405265, + 0.0036302567459642887, + 0.005776803940534592, + 0.004358797334134579, + 0.005281821824610233, + 0.007184678688645363, + 0.004471436142921448, + 0.005082719959318638, + 0.005442622117698193, + 0.008460799232125282, + 0.005885445978492498, + 0.0051318504847586155, + 0.005224133841693401, + 0.007450379431247711, + 0.004739082884043455, + 0.0060326470993459225, + 0.006286087445914745, + 0.009331092238426208, + 0.006647107191383839, + 0.005246875341981649, + 0.0015684174140915275, + 0.0016561986412853003, + 0.0013915123417973518, + 0.0020625093020498753, + 0.0022452606353908777, + 0.0030423470307141542, + 0.0021219186019152403, + 0.0018043711315840483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.01254018023610115 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195212058722973 + ], + "opposite_logprob_tokens": [ + -6.127195358276367 + ], + "logprob_reward": -0.002195212058722973, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195212058722973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10509075969457626, + 0.12376228719949722, + 0.11477676779031754, + 0.1038215234875679, + 0.18612168729305267, + 0.10608525574207306, + 0.04471784830093384, + 0.15012039244174957, + 0.027975337579846382, + 0.037528108805418015 + ], + "credit_map": [ + 0.011523567140102386, + 0.009332641027867794, + 0.016828877851366997, + 0.015473698265850544, + 0.011576307006180286, + 0.010416772216558456, + 0.01655559241771698, + 0.013383322395384312, + 0.013375475071370602, + 0.010280686430633068, + 0.012146279215812683, + 0.014659412205219269, + 0.01754787564277649, + 0.02683793194591999, + 0.015280826948583126, + 0.013633819296956062, + 0.013663563877344131, + 0.01086621917784214, + 0.020259689539670944, + 0.015110830776393414, + 0.009388984180986881, + 0.01569388620555401, + 0.01565328985452652, + 0.01414032094180584, + 0.011593950912356377, + 0.007054970599710941, + 0.0167302917689085, + 0.013090090826153755, + 0.01355972234159708, + 0.014193445444107056, + 0.01485429797321558, + 0.01274476945400238, + 0.02220037765800953, + 0.012216168455779552, + 0.023235425353050232, + 0.03187152370810509, + 0.019221147522330284, + 0.022860629484057426, + 0.029930582270026207, + 0.02458585612475872, + 0.013036409392952919, + 0.010482460260391235, + 0.018670635297894478, + 0.01629929058253765, + 0.008173099718987942, + 0.006752348504960537, + 0.019028116017580032, + 0.013642903417348862, + 0.005110361147671938, + 0.0038148541934788227, + 0.006292932201176882, + 0.006502163130789995, + 0.004744251258671284, + 0.005814496427774429, + 0.007040467578917742, + 0.005398326553404331, + 0.017102684825658798, + 0.02067491039633751, + 0.01698356308043003, + 0.018930431455373764, + 0.017501289024949074, + 0.02351182885468006, + 0.01898767426609993, + 0.0164280254393816, + 0.0039922818541526794, + 0.0030700559727847576, + 0.0047437953762710094, + 0.002661065198481083, + 0.002944473410025239, + 0.004018994979560375, + 0.0036031347699463367, + 0.002941539976745844, + 0.0044434769079089165, + 0.0034508213866502047, + 0.007178792729973793, + 0.005703295581042767, + 0.0032201367430388927, + 0.003650616854429245, + 0.0051180096343159676, + 0.004762965254485607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.002195212058722973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static medium shot from a fixed low-side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a sturdy fern with a dark, gnarled base visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fern's fronds display a striking copper-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, damp rock wall covered in deep green moss with a cascading water stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffused, cool-toned light with fine mist settling on the fern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946444813162088 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946444813162088, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946444813162088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1484253704547882, + 0.2340395748615265, + 0.0747726634144783, + 0.05566301569342613, + 0.07895934581756592, + 0.11167311668395996, + 0.09910223633050919, + 0.062161192297935486, + 0.04544605687260628, + 0.08975742757320404 + ], + "credit_map": [ + 0.014381889253854752, + 0.013469348661601543, + 0.02494821697473526, + 0.019056133925914764, + 0.01747988350689411, + 0.017967836931347847, + 0.020286204293370247, + 0.020835863426327705, + 0.024251487106084824, + 0.017333833500742912, + 0.024943888187408447, + 0.02833864465355873, + 0.03199540451169014, + 0.05178198218345642, + 0.028765588998794556, + 0.026628749445080757, + 0.010162899270653725, + 0.008510383777320385, + 0.01346934400498867, + 0.009595217183232307, + 0.007581921294331551, + 0.006440903525799513, + 0.009439213201403618, + 0.00957278162240982, + 0.006077698897570372, + 0.004429199732840061, + 0.00747470511123538, + 0.006198369897902012, + 0.006537823937833309, + 0.008634772151708603, + 0.009731054306030273, + 0.006579394452273846, + 0.008250006474554539, + 0.005998152773827314, + 0.009055832400918007, + 0.012808398343622684, + 0.008627337403595448, + 0.013061790727078915, + 0.010686913505196571, + 0.010470923967659473, + 0.01673687994480133, + 0.008490176871418953, + 0.016980037093162537, + 0.015730632469058037, + 0.009047897532582283, + 0.007746478542685509, + 0.021772608160972595, + 0.01516840048134327, + 0.012794725596904755, + 0.007354451809078455, + 0.01789010502398014, + 0.014133120886981487, + 0.010791083797812462, + 0.008911575190722942, + 0.015471202321350574, + 0.011755968444049358, + 0.005870236549526453, + 0.006808645091950893, + 0.009470746852457523, + 0.008519991301000118, + 0.005909511353820562, + 0.00881244707852602, + 0.00844040047377348, + 0.008329212665557861, + 0.0040605030953884125, + 0.0050584073178470135, + 0.00517917238175869, + 0.005119878798723221, + 0.007002913858741522, + 0.007556464057415724, + 0.005842104088515043, + 0.005626613739877939, + 0.010302478447556496, + 0.008600293658673763, + 0.01672813482582569, + 0.009981412440538406, + 0.008336806669831276, + 0.01491598505526781, + 0.010084891691803932, + 0.010807431302964687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.005946444813162088 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_4/index_tensor([6881, 6881])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_4/index_tensor([6881, 6881])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..08de1997f6b20ab50ac97690362a4545c1863fc7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_4/index_tensor([6881, 6881])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion, low-angle tracking shot delicately captures the tragic moment a young general plunges from his panicked, neighing black warhorse. The general is clad in a finely crafted suit of gleaming silver armor, its plates intricately carved with cloud patterns. The armor, battered by fierce battle, is riddled with scratches and slight dents, yet its cool metallic sheen gleams in the dimming light. His handsome face is stained with dust and a streak of blood, and his long, raven-black hair, once tied in a bun, has unfurled in the fall, fluttering messily in the air. As his body arches backward, plummeting to the ground in an almost balletic motion, the vast scarlet cape behind him suddenly unfurls like a flower. The backdrop is a vast, muddy battlefield, strewn with broken flags and discarded weapons. The sky is consumed by a hail of arrows, dense as locusts. Countless black bolts, whistling with a sharp whistle, pierce the smoky air from all directions, plummeting alongside the general. The faint glow of distant war flames and a few burning arrows provided the only light source, casting flickering spots of light on the general's armor, freezing the violence and beauty of this moment, and creating an epic yet heartbreaking atmosphere of a hero's end.", + "seed": 27565, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_4/index_6881_seed_27565_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.042736310977488756, + "vlm_logprob": -0.042736310977488756, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.042736310977488756, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052177976816892624 + ], + "opposite_logprob_tokens": [ + -3.302177906036377 + ], + "logprob_reward": -0.052177976816892624, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.052177976816892624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32274049520492554, + 0.09833594411611557, + 0.034245628863573074, + 0.07249964028596878, + 0.1308041214942932, + 0.14718393981456757, + 0.08498550206422806, + 0.02014404535293579, + 0.034160029143095016, + 0.054900720715522766 + ], + "credit_map": [ + 0.03325888141989708, + 0.03172098472714424, + 0.04727395251393318, + 0.054947223514318466, + 0.03335045278072357, + 0.0312546044588089, + 0.03990374505519867, + 0.05103059858083725, + 0.019766297191381454, + 0.00595055241137743, + 0.006415461655706167, + 0.007955563254654408, + 0.008109478279948235, + 0.03339928761124611, + 0.006403394974768162, + 0.010335905477404594, + 0.004972059745341539, + 0.004855129402130842, + 0.0033689080737531185, + 0.0031867159996181726, + 0.006513009779155254, + 0.004150316584855318, + 0.0031102453358471394, + 0.004089241847395897, + 0.01880795508623123, + 0.004378461744636297, + 0.006198733579367399, + 0.004192688502371311, + 0.0056321765296161175, + 0.024165401235222816, + 0.0037916507571935654, + 0.0053325677290558815, + 0.023678366094827652, + 0.01460222527384758, + 0.011773715727031231, + 0.011476404964923859, + 0.02104148454964161, + 0.01572481170296669, + 0.015537383034825325, + 0.016969718039035797, + 0.02426675334572792, + 0.025802796706557274, + 0.012709149159491062, + 0.010808603838086128, + 0.023882733657956123, + 0.019191071391105652, + 0.018025634810328484, + 0.012497183866798878, + 0.01241261139512062, + 0.009481760673224926, + 0.009778815321624279, + 0.004854684229940176, + 0.011270344257354736, + 0.016856851056218147, + 0.011941667646169662, + 0.008388768881559372, + 0.0030820597894489765, + 0.002242565853521228, + 0.0018606161465868354, + 0.001620409544557333, + 0.0033027473837137222, + 0.0036181106697767973, + 0.0026728769298642874, + 0.0017446576384827495, + 0.005740612745285034, + 0.0038791976403445005, + 0.002958765486255288, + 0.002270321361720562, + 0.006135015748441219, + 0.005872565787285566, + 0.004228631500154734, + 0.0030749139841645956, + 0.01162148267030716, + 0.0071667213924229145, + 0.0051536825485527515, + 0.003952317871153355, + 0.009812278673052788, + 0.007259102072566748, + 0.005756089463829994, + 0.004179045092314482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.052177976816892624 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038113709539175034 + ], + "opposite_logprob_tokens": [ + -3.288113594055176 + ], + "logprob_reward": -0.038113709539175034, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038113709539175034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1772775948047638, + 0.07706323266029358, + 0.08389213681221008, + 0.13895277678966522, + 0.07755153626203537, + 0.06598164141178131, + 0.11808271706104279, + 0.10849165171384811, + 0.0681162104010582, + 0.08459053933620453 + ], + "credit_map": [ + 0.026759030297398567, + 0.018657894805073738, + 0.02137961983680725, + 0.0187535360455513, + 0.027708467096090317, + 0.027058398351073265, + 0.014606322161853313, + 0.02235431969165802, + 0.0132223479449749, + 0.005646665580570698, + 0.007862108759582043, + 0.004822391550987959, + 0.008854523301124573, + 0.019091058522462845, + 0.0060622189193964005, + 0.011501927860081196, + 0.01593184657394886, + 0.011003074236214161, + 0.009686361998319626, + 0.005908544175326824, + 0.013424764387309551, + 0.01318640261888504, + 0.006483439821749926, + 0.008267701603472233, + 0.025762122124433517, + 0.01574493572115898, + 0.008115684613585472, + 0.011046899482607841, + 0.016096077859401703, + 0.03560587018728256, + 0.009874453768134117, + 0.016706734895706177, + 0.011273571290075779, + 0.006530227139592171, + 0.008609751239418983, + 0.010401030071079731, + 0.012496313080191612, + 0.011064654216170311, + 0.008619104512035847, + 0.008556884713470936, + 0.013276885263621807, + 0.009978177025914192, + 0.00884812418371439, + 0.0060043297708034515, + 0.007668965961784124, + 0.007176160346716642, + 0.008746313862502575, + 0.004282683599740267, + 0.01848449558019638, + 0.0145189817994833, + 0.016931846737861633, + 0.009414306841790676, + 0.014356371015310287, + 0.020404724404215813, + 0.013055277056992054, + 0.010916718281805515, + 0.01858784444630146, + 0.01265416294336319, + 0.008728237822651863, + 0.00881293136626482, + 0.014160391874611378, + 0.02030249498784542, + 0.014037984423339367, + 0.011207605712115765, + 0.011499394662678242, + 0.007661826442927122, + 0.005307616665959358, + 0.004329978488385677, + 0.008608795702457428, + 0.01463364064693451, + 0.007804957218468189, + 0.008270002901554108, + 0.013271056115627289, + 0.01149220485240221, + 0.0047296020202338696, + 0.004035403486341238, + 0.016060298308730125, + 0.015968022868037224, + 0.0067968652583658695, + 0.012237083166837692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.038113709539175034 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002500504720956087 + ], + "opposite_logprob_tokens": [ + -6.002500534057617 + ], + "logprob_reward": -0.002500504720956087, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002500504720956087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2808249890804291, + 0.14280518889427185, + 0.0776209756731987, + 0.08323585242033005, + 0.05594371631741524, + 0.04584397003054619, + 0.07126306742429733, + 0.07347254455089569, + 0.06653618067502975, + 0.1024535745382309 + ], + "credit_map": [ + 0.04753992706537247, + 0.02892673946917057, + 0.04075559228658676, + 0.025212716311216354, + 0.04072451591491699, + 0.04494716599583626, + 0.020640719681978226, + 0.0320776142179966, + 0.025102365761995316, + 0.010859956964850426, + 0.014643042348325253, + 0.008239775896072388, + 0.015861796215176582, + 0.04236367344856262, + 0.011249955743551254, + 0.014484614133834839, + 0.011184928007423878, + 0.008529928512871265, + 0.007492402568459511, + 0.005929307546466589, + 0.01052901428192854, + 0.01586264930665493, + 0.00814096163958311, + 0.009951782412827015, + 0.012399408966302872, + 0.008415143936872482, + 0.009101013652980328, + 0.00616608327254653, + 0.010879562236368656, + 0.018513502553105354, + 0.008362030610442162, + 0.00939911138266325, + 0.011235886253416538, + 0.009392852894961834, + 0.004834148567169905, + 0.005155571270734072, + 0.006922477390617132, + 0.006920090410858393, + 0.007036914117634296, + 0.004445776809006929, + 0.00669462326914072, + 0.010536335408687592, + 0.004907339345663786, + 0.0037213610485196114, + 0.005886262748390436, + 0.005605034995824099, + 0.005320766940712929, + 0.003172246040776372, + 0.012102131731808186, + 0.009259598329663277, + 0.00852357316762209, + 0.005870667286217213, + 0.009734494611620903, + 0.013087891042232513, + 0.005167329218238592, + 0.00751738203689456, + 0.011044505052268505, + 0.008611371740698814, + 0.005999122746288776, + 0.00546935573220253, + 0.008884342387318611, + 0.016325317323207855, + 0.008688827976584435, + 0.008449697867035866, + 0.01340240053832531, + 0.007746964693069458, + 0.005871883127838373, + 0.004393902141600847, + 0.011967241764068604, + 0.011047669686377048, + 0.005903791636228561, + 0.006202335469424725, + 0.025263069197535515, + 0.013862925581634045, + 0.00929797999560833, + 0.006914233323186636, + 0.019106552004814148, + 0.01360032707452774, + 0.008736451156437397, + 0.00567204225808382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.002500504720956087 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058476023375988 + ], + "opposite_logprob_tokens": [ + -3.89558482170105 + ], + "logprob_reward": -0.02058476023375988, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058476023375988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5270241498947144, + 0.036826398223638535, + 0.052211981266736984, + 0.11545176804065704, + 0.0628284141421318, + 0.025534484535455704, + 0.027645299211144447, + 0.0400453545153141, + 0.05431685969233513, + 0.058115292340517044 + ], + "credit_map": [ + 0.07135117799043655, + 0.053105778992176056, + 0.07865156978368759, + 0.05910424888134003, + 0.05356943979859352, + 0.0915941447019577, + 0.05148552730679512, + 0.0681622326374054, + 0.005402715876698494, + 0.002881817752495408, + 0.00623727310448885, + 0.003071116516366601, + 0.0037845075130462646, + 0.006972822826355696, + 0.004065345041453838, + 0.004410802852362394, + 0.009252536110579967, + 0.004874226171523333, + 0.0048966477625072, + 0.004766607657074928, + 0.006143154110759497, + 0.009051583707332611, + 0.005227609071880579, + 0.007999615743756294, + 0.017358750104904175, + 0.0146070821210742, + 0.013120712712407112, + 0.008676520548760891, + 0.01344138290733099, + 0.022707952186465263, + 0.010468875989317894, + 0.015070490539073944, + 0.012369475327432156, + 0.006840330548584461, + 0.006316694896668196, + 0.007075009401887655, + 0.009478742256760597, + 0.008233086206018925, + 0.006275107152760029, + 0.006239967420697212, + 0.005357736721634865, + 0.004012384917587042, + 0.002849608426913619, + 0.0026981092523783445, + 0.0033193931449204683, + 0.0027638161554932594, + 0.0029115001671016216, + 0.0016219362150877714, + 0.0036195022985339165, + 0.003606146667152643, + 0.0030552956741303205, + 0.0026650377549231052, + 0.004145530518144369, + 0.0032788750249892473, + 0.0038059300277382135, + 0.003468981711193919, + 0.007370125502347946, + 0.00535154715180397, + 0.0034916657023131847, + 0.0029846972320228815, + 0.005253889132291079, + 0.006380345206707716, + 0.005316419992595911, + 0.003896663896739483, + 0.010607992298901081, + 0.005806135479360819, + 0.0041552274487912655, + 0.0031791306100785732, + 0.008292190730571747, + 0.011061045341193676, + 0.005036369897425175, + 0.00617876835167408, + 0.007434615399688482, + 0.009001854807138443, + 0.003425524802878499, + 0.0033003161661326885, + 0.011838576756417751, + 0.010842722840607166, + 0.004765178542584181, + 0.007506502792239189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.02058476023375988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030460357666016 + ], + "opposite_logprob_tokens": [ + -2.35030460357666 + ], + "logprob_reward": -0.10030460357666016, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030460357666016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16763542592525482, + 0.11435425281524658, + 0.1019168496131897, + 0.17001576721668243, + 0.10043536871671677, + 0.05597545951604843, + 0.056699372828006744, + 0.04277871549129486, + 0.09683167934417725, + 0.09335700422525406 + ], + "credit_map": [ + 0.031426094472408295, + 0.014952135272324085, + 0.021247021853923798, + 0.01493084616959095, + 0.020118793472647667, + 0.02741917222738266, + 0.0168295968323946, + 0.020711787045001984, + 0.02006879821419716, + 0.00955211091786623, + 0.009887774474918842, + 0.008197632618248463, + 0.013090434484183788, + 0.029165981337428093, + 0.01343389693647623, + 0.01095764059573412, + 0.019979650154709816, + 0.008709713816642761, + 0.009339595213532448, + 0.006422558333724737, + 0.012332207523286343, + 0.027732159942388535, + 0.00790480151772499, + 0.009496175684034824, + 0.03771885111927986, + 0.012983807362616062, + 0.01457120105624199, + 0.00900029856711626, + 0.01976943574845791, + 0.04681132733821869, + 0.011688044294714928, + 0.01747281476855278, + 0.015226425603032112, + 0.010210231877863407, + 0.010071701370179653, + 0.010906548239290714, + 0.01795545592904091, + 0.0144724752753973, + 0.011907798238098621, + 0.00968474242836237, + 0.01231443416327238, + 0.009541257284581661, + 0.005478243343532085, + 0.0046597858890891075, + 0.005672751925885677, + 0.0067419856786727905, + 0.008210831321775913, + 0.0033561731688678265, + 0.009526502341032028, + 0.006907343864440918, + 0.008684946224093437, + 0.003435738617554307, + 0.007155039347708225, + 0.007486436516046524, + 0.006274668499827385, + 0.007228698581457138, + 0.005638439208269119, + 0.004705639556050301, + 0.004308012314140797, + 0.0037333208601921797, + 0.005836027674376965, + 0.006603762041777372, + 0.006985901854932308, + 0.0049676145426929, + 0.016405154019594193, + 0.0121240746229887, + 0.0075142537243664265, + 0.007059003692120314, + 0.01568330079317093, + 0.01756640523672104, + 0.010510365478694439, + 0.009969125501811504, + 0.01924053579568863, + 0.014526846818625927, + 0.009083686396479607, + 0.005430356599390507, + 0.013932478614151478, + 0.015508546493947506, + 0.008497439324855804, + 0.007137129083275795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.10030460357666016 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion, low-angle tracking shot delicately captures the tragic moment a young general plunges from his panicked, neighing black warhorse. The general is clad in a finely crafted suit of gleaming silver armor, its plates intricately carved with cloud patterns. The armor, battered by fierce battle, is riddled with scratches and slight dents, yet its cool metallic sheen gleams in the dimming light. His handsome face is stained with dust and a streak of blood, and his long, raven-black hair, once tied in a bun, has unfurled in the fall, fluttering messily in the air. As his body arches backward, plummeting to the ground in an almost balletic motion, the vast scarlet cape behind him suddenly unfurls like a flower. The backdrop is a vast, muddy battlefield, strewn with broken flags and discarded weapons. The sky is consumed by a hail of arrows, dense as locusts. Countless black bolts, whistling with a sharp whistle, pierce the smoky air from all directions, plummeting alongside the general. The faint glow of distant war flames and a few burning arrows provided the only light source, casting flickering spots of light on the general's armor, freezing the violence and beauty of this moment, and creating an epic yet heartbreaking atmosphere of a hero's end.", + "seed": 27566, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_4/index_6881_seed_27566_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03498019680846483, + "vlm_logprob": -0.03498019680846483, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03498019680846483, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07517952471971512 + ], + "opposite_logprob_tokens": [ + -2.7001795768737793 + ], + "logprob_reward": -0.07517952471971512, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07517952471971512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2618005871772766, + 0.07026072591543198, + 0.11062981188297272, + 0.2073957324028015, + 0.063524030148983, + 0.018472695723176003, + 0.11985037475824356, + 0.050641439855098724, + 0.028334179893136024, + 0.06909041106700897 + ], + "credit_map": [ + 0.039562907069921494, + 0.047627758234739304, + 0.052321672439575195, + 0.04269038513302803, + 0.028426703065633774, + 0.01659751497209072, + 0.017232026904821396, + 0.017341623082756996, + 0.008950653485953808, + 0.009517774917185307, + 0.010776720009744167, + 0.010330872610211372, + 0.01026071421802044, + 0.005009201355278492, + 0.00946149230003357, + 0.005953299812972546, + 0.016000496223568916, + 0.011988837271928787, + 0.012614395469427109, + 0.011774559505283833, + 0.013832672499120235, + 0.013527384959161282, + 0.020124526694417, + 0.010766944848001003, + 0.018186569213867188, + 0.028558462858200073, + 0.022399479523301125, + 0.021926648914813995, + 0.02483425848186016, + 0.022884497418999672, + 0.040655191987752914, + 0.02795061469078064, + 0.007898257113993168, + 0.005089405924081802, + 0.006289041601121426, + 0.011900645680725574, + 0.007954674772918224, + 0.00916430726647377, + 0.010082678869366646, + 0.005145024508237839, + 0.0030218809843063354, + 0.0015518643194809556, + 0.0023878931533545256, + 0.002803155453875661, + 0.0019371876260265708, + 0.0023668291978538036, + 0.0025377424899488688, + 0.0018661427311599255, + 0.01461355946958065, + 0.01622740551829338, + 0.013702010735869408, + 0.013379806652665138, + 0.012118378654122353, + 0.016351984813809395, + 0.016195636242628098, + 0.01726159080862999, + 0.0059401593171060085, + 0.007127097807824612, + 0.004498225171118975, + 0.004810780752450228, + 0.009702986106276512, + 0.004771851934492588, + 0.008229399099946022, + 0.005560941528528929, + 0.0033990214578807354, + 0.0030273697338998318, + 0.0039521474391222, + 0.004201177507638931, + 0.004016469698399305, + 0.0021165194921195507, + 0.005012274254113436, + 0.0026092007756233215, + 0.010545585304498672, + 0.006567253265529871, + 0.00999396201223135, + 0.007057430688291788, + 0.016109386458992958, + 0.00659247487783432, + 0.0076184989884495735, + 0.004605822265148163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.07517952471971512 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799826443195343 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029799826443195343, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799826443195343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24239039421081543, + 0.14392362534999847, + 0.07083183526992798, + 0.033600807189941406, + 0.07887639850378036, + 0.05447554588317871, + 0.12111622840166092, + 0.053376901894807816, + 0.07897324115037918, + 0.12243501096963882 + ], + "credit_map": [ + 0.029907293617725372, + 0.03845016285777092, + 0.049727339297533035, + 0.04234297201037407, + 0.02450055256485939, + 0.020147955045104027, + 0.016610927879810333, + 0.020703190937638283, + 0.018660100176930428, + 0.022714924067258835, + 0.015014197677373886, + 0.02083929441869259, + 0.020697256550192833, + 0.015753522515296936, + 0.01761784590780735, + 0.012626489624381065, + 0.01106873620301485, + 0.006894262973219156, + 0.015395530499517918, + 0.007847347296774387, + 0.008128432556986809, + 0.006686300039291382, + 0.008790476247668266, + 0.006020743865519762, + 0.006463427096605301, + 0.004160902928560972, + 0.004721702076494694, + 0.003025110112503171, + 0.003310448955744505, + 0.004073803778737783, + 0.004678679164499044, + 0.003166730748489499, + 0.010753634385764599, + 0.009639263153076172, + 0.012503469362854958, + 0.010941794142127037, + 0.008617757819592953, + 0.013089642859995365, + 0.008249289356172085, + 0.005081545561552048, + 0.009687365964055061, + 0.006521459203213453, + 0.007786538917571306, + 0.0064209527336061, + 0.006640348117798567, + 0.005606380756944418, + 0.006632248871028423, + 0.005180250387638807, + 0.018129859119653702, + 0.024929041042923927, + 0.01897699199616909, + 0.012929771095514297, + 0.012272578664124012, + 0.012718653306365013, + 0.011770169250667095, + 0.009389160200953484, + 0.006088905967772007, + 0.007816988043487072, + 0.005171467084437609, + 0.005990307778120041, + 0.008128102868795395, + 0.006289680954068899, + 0.006647762376815081, + 0.007243681699037552, + 0.009169861674308777, + 0.007673132698982954, + 0.012357912957668304, + 0.013241142965853214, + 0.01426014769822359, + 0.005868993233889341, + 0.009620999917387962, + 0.006781049072742462, + 0.015464154072105885, + 0.018089668825268745, + 0.022036641836166382, + 0.016923442482948303, + 0.013655345886945724, + 0.011022324673831463, + 0.015107566490769386, + 0.010135872289538383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.029799826443195343 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019459851318970323 + ], + "opposite_logprob_tokens": [ + -6.251945972442627 + ], + "logprob_reward": -0.0019459851318970323, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019459851318970323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3961522877216339, + 0.09767945855855942, + 0.11211084574460983, + 0.08275546878576279, + 0.06303472071886063, + 0.05024537071585655, + 0.06824901700019836, + 0.02872963808476925, + 0.024749180302023888, + 0.07629402726888657 + ], + "credit_map": [ + 0.052206557244062424, + 0.0650034174323082, + 0.07862809300422668, + 0.07044736295938492, + 0.04547877237200737, + 0.031825482845306396, + 0.025265315547585487, + 0.0272972472012043, + 0.01201177854090929, + 0.019265659153461456, + 0.012836403213441372, + 0.014300494454801083, + 0.014924796298146248, + 0.008637159131467342, + 0.008224090561270714, + 0.007479075808078051, + 0.01331344060599804, + 0.011196110397577286, + 0.0172683484852314, + 0.017533041536808014, + 0.011788316071033478, + 0.011957995593547821, + 0.01438174583017826, + 0.014671842567622662, + 0.009592456743121147, + 0.010240989737212658, + 0.009613662958145142, + 0.008323046378791332, + 0.011462850496172905, + 0.01109369471669197, + 0.012915967032313347, + 0.009512798860669136, + 0.006292096339166164, + 0.006425723899155855, + 0.007107810582965612, + 0.010585227981209755, + 0.007958286441862583, + 0.012221933342516422, + 0.006851262878626585, + 0.005592385306954384, + 0.010778169147670269, + 0.005135269369930029, + 0.005817752797156572, + 0.009534455835819244, + 0.00412096967920661, + 0.0058300127275288105, + 0.005073468200862408, + 0.00395527109503746, + 0.009840553626418114, + 0.013385198079049587, + 0.0067029125057160854, + 0.006202689837664366, + 0.0073198433965444565, + 0.010566860437393188, + 0.006606756243854761, + 0.007624203339219093, + 0.004229715559631586, + 0.004862533416599035, + 0.002464098623022437, + 0.002442150143906474, + 0.0036276644095778465, + 0.004369989037513733, + 0.0036752442829310894, + 0.0030582426115870476, + 0.0031766374595463276, + 0.002070213435217738, + 0.004392885137349367, + 0.004482397809624672, + 0.004128719680011272, + 0.0019047155510634184, + 0.002680297940969467, + 0.0019133134046569467, + 0.008325435221195221, + 0.008804701268672943, + 0.015539598651230335, + 0.013551020994782448, + 0.012777523137629032, + 0.004573860205709934, + 0.008532227016985416, + 0.004189661704003811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0019459851318970323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943719297647476 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943719297647476, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943719297647476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25058990716934204, + 0.22056764364242554, + 0.07414413243532181, + 0.05883580818772316, + 0.05233892425894737, + 0.13463912904262543, + 0.08598346263170242, + 0.05533919483423233, + 0.028202751651406288, + 0.03935914486646652 + ], + "credit_map": [ + 0.023588767275214195, + 0.026851730421185493, + 0.033421408385038376, + 0.03070787899196148, + 0.039543285965919495, + 0.03238995745778084, + 0.025159072130918503, + 0.038927771151065826, + 0.02241627126932144, + 0.043035831302404404, + 0.03415095433592796, + 0.033874623477458954, + 0.0300754401832819, + 0.021395711228251457, + 0.01829693280160427, + 0.017321838065981865, + 0.011393135413527489, + 0.007266698405146599, + 0.016833629459142685, + 0.00876173097640276, + 0.006663862615823746, + 0.008249986916780472, + 0.006985128857195377, + 0.007989959791302681, + 0.006830248981714249, + 0.007826297543942928, + 0.007154448423534632, + 0.005952197592705488, + 0.007357320282608271, + 0.00656169606372714, + 0.00926722027361393, + 0.0078863725066185, + 0.00665490236133337, + 0.005642484407871962, + 0.005101494491100311, + 0.007867353036999702, + 0.007150524761527777, + 0.007712909951806068, + 0.006370477378368378, + 0.005838776007294655, + 0.02710118144750595, + 0.012111248448491096, + 0.019790055230259895, + 0.032260961830616, + 0.008563443087041378, + 0.011101294308900833, + 0.014547591097652912, + 0.009163334965705872, + 0.013726633042097092, + 0.01570168137550354, + 0.015021419152617455, + 0.008496949449181557, + 0.008219877257943153, + 0.008737402968108654, + 0.008257602341473103, + 0.007821882143616676, + 0.00788026675581932, + 0.011461040936410427, + 0.005123916082084179, + 0.005673114210367203, + 0.007352776359766722, + 0.007798707578331232, + 0.006171026267111301, + 0.0038783401250839233, + 0.0036797933280467987, + 0.0023382301442325115, + 0.005732895340770483, + 0.00525331636890769, + 0.00398973748087883, + 0.0020567404571920633, + 0.0029264322947710752, + 0.0022256013471633196, + 0.00452131312340498, + 0.006173391826450825, + 0.0072282045148313046, + 0.006512816529721022, + 0.004495783243328333, + 0.003212404204532504, + 0.004204649478197098, + 0.00301057449541986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.005943719297647476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062031928449869156 + ], + "opposite_logprob_tokens": [ + -2.8120319843292236 + ], + "logprob_reward": -0.062031928449869156, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062031928449869156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20881935954093933, + 0.0822773426771164, + 0.08108235150575638, + 0.0554288811981678, + 0.17005042731761932, + 0.1088542491197586, + 0.11252665519714355, + 0.10735771059989929, + 0.01624956913292408, + 0.05735339596867561 + ], + "credit_map": [ + 0.023223578929901123, + 0.021286632865667343, + 0.03986996412277222, + 0.03013291396200657, + 0.018380673602223396, + 0.02345331571996212, + 0.020294206216931343, + 0.03217807412147522, + 0.010724197141826153, + 0.01984615996479988, + 0.009759724140167236, + 0.013431369327008724, + 0.012809821404516697, + 0.005172567442059517, + 0.005552599206566811, + 0.004980904050171375, + 0.011923609301447868, + 0.007845808751881123, + 0.017561161890625954, + 0.011674789711833, + 0.0071157123893499374, + 0.00949154607951641, + 0.009113670326769352, + 0.006356050726026297, + 0.006369828246533871, + 0.007771351374685764, + 0.008040204644203186, + 0.004101246129721403, + 0.005080845206975937, + 0.007798844017088413, + 0.010896877385675907, + 0.005369683727622032, + 0.018416408449411392, + 0.01697193831205368, + 0.02571514993906021, + 0.026997201144695282, + 0.017621418461203575, + 0.0299026258289814, + 0.02055753953754902, + 0.01386814285069704, + 0.016665352508425713, + 0.011368952691555023, + 0.014275062829256058, + 0.010827813297510147, + 0.010241040959954262, + 0.014481568709015846, + 0.01950462907552719, + 0.011489832773804665, + 0.013676184229552746, + 0.019566945731639862, + 0.020740175619721413, + 0.011833038181066513, + 0.01176253892481327, + 0.011034284718334675, + 0.013497167266905308, + 0.010416323319077492, + 0.015155874192714691, + 0.0197562538087368, + 0.010629823431372643, + 0.012273554690182209, + 0.014387490227818489, + 0.013078038580715656, + 0.012235757894814014, + 0.009840919636189938, + 0.0022762767039239407, + 0.00134804577101022, + 0.0031223571859300137, + 0.0025577049236744642, + 0.0025837444700300694, + 0.0014143807347863913, + 0.0017253459664061666, + 0.0012217136099934578, + 0.007170513737946749, + 0.008228275924921036, + 0.009909569285809994, + 0.011160646565258503, + 0.006064441055059433, + 0.004276019521057606, + 0.006613020319491625, + 0.003930903505533934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.062031928449869156 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion, low-angle tracking shot delicately captures the tragic moment a young general plunges from his panicked, neighing black warhorse. The general is clad in a finely crafted suit of gleaming silver armor, its plates intricately carved with cloud patterns. The armor, battered by fierce battle, is riddled with scratches and slight dents, yet its cool metallic sheen gleams in the dimming light. His handsome face is stained with dust and a streak of blood, and his long, raven-black hair, once tied in a bun, has unfurled in the fall, fluttering messily in the air. As his body arches backward, plummeting to the ground in an almost balletic motion, the vast scarlet cape behind him suddenly unfurls like a flower. The backdrop is a vast, muddy battlefield, strewn with broken flags and discarded weapons. The sky is consumed by a hail of arrows, dense as locusts. Countless black bolts, whistling with a sharp whistle, pierce the smoky air from all directions, plummeting alongside the general. The faint glow of distant war flames and a few burning arrows provided the only light source, casting flickering spots of light on the general's armor, freezing the violence and beauty of this moment, and creating an epic yet heartbreaking atmosphere of a hero's end.", + "seed": 27567, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_4/index_6881_seed_27567_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09068455058149993, + "vlm_logprob": -0.09068455058149993, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.09068455058149993, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3207342326641083 + ], + "opposite_logprob_tokens": [ + -1.3207342624664307 + ], + "logprob_reward": -0.3207342326641083, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3207342326641083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12814253568649292, + 0.03605910390615463, + 0.05756804719567299, + 0.10177769511938095, + 0.12862949073314667, + 0.2001112401485443, + 0.12307187169790268, + 0.07199953496456146, + 0.11828510463237762, + 0.03435534983873367 + ], + "credit_map": [ + 0.024543052539229393, + 0.016500340774655342, + 0.017283156514167786, + 0.013463225215673447, + 0.023641977459192276, + 0.01299525611102581, + 0.009179526939988136, + 0.010535994544625282, + 0.00524565577507019, + 0.003156248712912202, + 0.0037704226560890675, + 0.003341008210554719, + 0.0029102270491421223, + 0.006357762962579727, + 0.005667663179337978, + 0.005610113963484764, + 0.006957986857742071, + 0.010643751360476017, + 0.008228828199207783, + 0.006958772893995047, + 0.006046009249985218, + 0.005711530800908804, + 0.005700088106095791, + 0.007321080658584833, + 0.020595740526914597, + 0.014929931610822678, + 0.013394152745604515, + 0.01191655918955803, + 0.011348907835781574, + 0.00788190308958292, + 0.011893033981323242, + 0.009817466139793396, + 0.01901151053607464, + 0.018562451004981995, + 0.02359728328883648, + 0.014280293136835098, + 0.010290248319506645, + 0.012315265834331512, + 0.010078673250973225, + 0.020493775606155396, + 0.032115936279296875, + 0.019570235162973404, + 0.03515450283885002, + 0.01818716712296009, + 0.021763747557997704, + 0.022629400715231895, + 0.015692010521888733, + 0.03499823808670044, + 0.015867339447140694, + 0.010702730156481266, + 0.014724423177540302, + 0.015467623248696327, + 0.02390986494719982, + 0.012949330732226372, + 0.01551456842571497, + 0.013935999013483524, + 0.007126829121261835, + 0.00655136164277792, + 0.006512646563351154, + 0.004440618678927422, + 0.01145285926759243, + 0.021138697862625122, + 0.004817304201424122, + 0.009959213435649872, + 0.011409200727939606, + 0.012442000210285187, + 0.0226887296885252, + 0.014217229560017586, + 0.01120712049305439, + 0.014721722342073917, + 0.016580374911427498, + 0.015018726699054241, + 0.004679790232330561, + 0.004442855715751648, + 0.004199841525405645, + 0.004041425418108702, + 0.003792362054809928, + 0.006172391586005688, + 0.002416068222373724, + 0.004610614385455847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.3207342326641083 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496569350361824 + ], + "opposite_logprob_tokens": [ + -2.9299657344818115 + ], + "logprob_reward": -0.05496569350361824, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496569350361824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23191605508327484, + 0.11947966367006302, + 0.17937806248664856, + 0.05282659828662872, + 0.07715427130460739, + 0.10752683132886887, + 0.08296947926282883, + 0.06072600185871124, + 0.05194157734513283, + 0.03608141839504242 + ], + "credit_map": [ + 0.03034796752035618, + 0.022152390331029892, + 0.036121368408203125, + 0.03498620167374611, + 0.038547806441783905, + 0.028630880638957024, + 0.020190343260765076, + 0.02093910425901413, + 0.01993049494922161, + 0.013936340808868408, + 0.013951798900961876, + 0.013866213150322437, + 0.017857898026704788, + 0.01577572524547577, + 0.013627207837998867, + 0.010533987544476986, + 0.020608913153409958, + 0.02460879646241665, + 0.021685337647795677, + 0.02503577619791031, + 0.029008988291025162, + 0.01822836510837078, + 0.01936219073832035, + 0.02083970047533512, + 0.010279669426381588, + 0.008637174032628536, + 0.007460485212504864, + 0.0057130795903503895, + 0.00510140648111701, + 0.004696331452578306, + 0.00520136347040534, + 0.0057370844297111034, + 0.008835045620799065, + 0.008006975054740906, + 0.016810566186904907, + 0.00940290279686451, + 0.00852892268449068, + 0.005817935802042484, + 0.008056089282035828, + 0.011695833876729012, + 0.009600309655070305, + 0.01091775856912136, + 0.020771682262420654, + 0.010434199124574661, + 0.013889756985008717, + 0.01601124182343483, + 0.009768185205757618, + 0.01613369770348072, + 0.014414874836802483, + 0.007000138983130455, + 0.008745307102799416, + 0.008547267876565456, + 0.017410190775990486, + 0.008920641615986824, + 0.009152049198746681, + 0.008779008872807026, + 0.01104153972119093, + 0.005744793452322483, + 0.00654545146971941, + 0.005427283700555563, + 0.007705029100179672, + 0.011750789359211922, + 0.00474986108019948, + 0.007761254440993071, + 0.006224927958101034, + 0.004508520942181349, + 0.007867616601288319, + 0.005588226020336151, + 0.0052703674882650375, + 0.007408243138343096, + 0.008155847899615765, + 0.006917828228324652, + 0.004252418875694275, + 0.004072945099323988, + 0.004633086733520031, + 0.004860789980739355, + 0.004396900534629822, + 0.0060544791631400585, + 0.00378455244936049, + 0.004026243928819895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.05496569350361824 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036239444743841887 + ], + "opposite_logprob_tokens": [ + -5.628623962402344 + ], + "logprob_reward": -0.0036239444743841887, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036239444743841887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2668065130710602, + 0.07711490243673325, + 0.0783391147851944, + 0.06559289991855621, + 0.058651283383369446, + 0.12760639190673828, + 0.13344241678714752, + 0.05026978999376297, + 0.07548103481531143, + 0.06669560074806213 + ], + "credit_map": [ + 0.03326759859919548, + 0.031117605045437813, + 0.0426112562417984, + 0.03116590343415737, + 0.05658456310629845, + 0.031141499057412148, + 0.02030000276863575, + 0.020618092268705368, + 0.01525408960878849, + 0.007448437623679638, + 0.00837796926498413, + 0.009833447635173798, + 0.0088421031832695, + 0.008669920265674591, + 0.00999476108700037, + 0.0086941784247756, + 0.01095610298216343, + 0.012835818342864513, + 0.012226339429616928, + 0.012960988096892834, + 0.011617172509431839, + 0.005789186805486679, + 0.005165413953363895, + 0.006788091734051704, + 0.01438070647418499, + 0.010664592497050762, + 0.009825587272644043, + 0.00623665563762188, + 0.007223171181976795, + 0.0047293840907514095, + 0.005456829443573952, + 0.0070759751833975315, + 0.00866066012531519, + 0.009784622117877007, + 0.009316456504166126, + 0.007196351885795593, + 0.005627800244837999, + 0.004631807096302509, + 0.005231092218309641, + 0.00820249505341053, + 0.02560705691576004, + 0.01781744882464409, + 0.015486842952668667, + 0.010565143078565598, + 0.011383021250367165, + 0.021683627739548683, + 0.008083716966211796, + 0.016979537904262543, + 0.015080497600138187, + 0.01390580739825964, + 0.014224755577743053, + 0.019763827323913574, + 0.029962947592139244, + 0.01397704891860485, + 0.012469946406781673, + 0.014057599939405918, + 0.005965457297861576, + 0.004610490053892136, + 0.005625263787806034, + 0.0033449188340455294, + 0.005917093250900507, + 0.01256658136844635, + 0.006438723299652338, + 0.00580126466229558, + 0.008222122676670551, + 0.008157243952155113, + 0.01164519228041172, + 0.008727447129786015, + 0.009449497796595097, + 0.0120179308578372, + 0.007935304194688797, + 0.009326299652457237, + 0.008643834851682186, + 0.008603527210652828, + 0.006289240904152393, + 0.008470653556287289, + 0.008400912396609783, + 0.012490830384194851, + 0.0051050204783678055, + 0.00869158934801817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0036239444743841887 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092295654118061 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.004092295654118061, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092295654118061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17059487104415894, + 0.07252658158540726, + 0.20862849056720734, + 0.046436287462711334, + 0.0872817412018776, + 0.08832982927560806, + 0.12040617316961288, + 0.049498967826366425, + 0.09543830901384354, + 0.060858745127916336 + ], + "credit_map": [ + 0.03382202610373497, + 0.024294491857290268, + 0.03211291879415512, + 0.017280299216508865, + 0.023322729393839836, + 0.01575866900384426, + 0.012052775360643864, + 0.01195098552852869, + 0.011241386644542217, + 0.009609431959688663, + 0.008352820761501789, + 0.007603114005178213, + 0.00999229121953249, + 0.012735332362353802, + 0.00508682569488883, + 0.00790538638830185, + 0.030310506001114845, + 0.03839649632573128, + 0.021813519299030304, + 0.029906848445534706, + 0.0374879352748394, + 0.015824083238840103, + 0.01641051284968853, + 0.018478617072105408, + 0.010337989777326584, + 0.0071829236112535, + 0.006139359436929226, + 0.005306154955178499, + 0.004977069329470396, + 0.003273495938628912, + 0.004311175551265478, + 0.004908119793981314, + 0.0120902294293046, + 0.009860869497060776, + 0.018207544460892677, + 0.009363173507153988, + 0.009491592645645142, + 0.007404615171253681, + 0.007407757919281721, + 0.013455964624881744, + 0.013548283837735653, + 0.010391696356236935, + 0.014188257046043873, + 0.007557156030088663, + 0.008939292281866074, + 0.01287827081978321, + 0.006301963701844215, + 0.014524917118251324, + 0.018478741869330406, + 0.011202344670891762, + 0.012103320099413395, + 0.01594269834458828, + 0.026169121265411377, + 0.012208446860313416, + 0.012223777361214161, + 0.01207772921770811, + 0.006725619547069073, + 0.004036828875541687, + 0.0061813476495444775, + 0.003753812750801444, + 0.003603388788178563, + 0.010688788257539272, + 0.006114527117460966, + 0.008394657634198666, + 0.009908614680171013, + 0.0091432249173522, + 0.013297172263264656, + 0.009841273538768291, + 0.01052357442677021, + 0.017308823764324188, + 0.012953920289874077, + 0.012461716309189796, + 0.007349241059273481, + 0.006610063835978508, + 0.006949561648070812, + 0.007886392064392567, + 0.0068544652312994, + 0.012002437375485897, + 0.005265243351459503, + 0.00794134195894003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.004092295654118061 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700065866112709 + ], + "opposite_logprob_tokens": [ + -2.6950066089630127 + ], + "logprob_reward": -0.0700065866112709, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700065866112709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2050131857395172, + 0.11175042390823364, + 0.26099687814712524, + 0.08873496949672699, + 0.06421786546707153, + 0.05102583020925522, + 0.03424820303916931, + 0.0481700636446476, + 0.10504791885614395, + 0.030794689431786537 + ], + "credit_map": [ + 0.0224457960575819, + 0.03299944847822189, + 0.032890286296606064, + 0.02633790113031864, + 0.03369041159749031, + 0.022926215082406998, + 0.01811269298195839, + 0.01561042945832014, + 0.021561939269304276, + 0.012428488582372665, + 0.013305220752954483, + 0.007934745401144028, + 0.016159920021891594, + 0.017923666164278984, + 0.010173074901103973, + 0.012263369746506214, + 0.03226431459188461, + 0.03788420930504799, + 0.03109605982899666, + 0.03643240034580231, + 0.03959271311759949, + 0.023871950805187225, + 0.028040897101163864, + 0.03181435167789459, + 0.014185807667672634, + 0.013938546180725098, + 0.01077681127935648, + 0.01131479348987341, + 0.010692211799323559, + 0.008721190504729748, + 0.009768899530172348, + 0.00933672022074461, + 0.007358159404247999, + 0.006551784463226795, + 0.013926209881901741, + 0.008727412670850754, + 0.005885195918381214, + 0.005205652676522732, + 0.007553966715931892, + 0.009009486064314842, + 0.008714818395674229, + 0.007768620736896992, + 0.008047517389059067, + 0.0037241047248244286, + 0.0052106440998613834, + 0.010083110071718693, + 0.0026129966136068106, + 0.004864020273089409, + 0.004535664338618517, + 0.0032405678648501635, + 0.004623441956937313, + 0.0045046755112707615, + 0.0052200560458004475, + 0.0030953348614275455, + 0.004790156614035368, + 0.0042383079417049885, + 0.005104013253003359, + 0.0037648025900125504, + 0.005954550579190254, + 0.004390453454107046, + 0.006517862901091576, + 0.008414935320615768, + 0.004507806152105331, + 0.009515642188489437, + 0.010737120173871517, + 0.009194828569889069, + 0.017437709495425224, + 0.010972745716571808, + 0.011775891296565533, + 0.017191052436828613, + 0.014132537879049778, + 0.01360604353249073, + 0.0033447211608290672, + 0.002936979290097952, + 0.006422231439501047, + 0.003908757586032152, + 0.00297541543841362, + 0.005038299132138491, + 0.0023472437169402838, + 0.0038210428319871426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0700065866112709 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion, low-angle tracking shot delicately captures the tragic moment a young general plunges from his panicked, neighing black warhorse. The general is clad in a finely crafted suit of gleaming silver armor, its plates intricately carved with cloud patterns. The armor, battered by fierce battle, is riddled with scratches and slight dents, yet its cool metallic sheen gleams in the dimming light. His handsome face is stained with dust and a streak of blood, and his long, raven-black hair, once tied in a bun, has unfurled in the fall, fluttering messily in the air. As his body arches backward, plummeting to the ground in an almost balletic motion, the vast scarlet cape behind him suddenly unfurls like a flower. The backdrop is a vast, muddy battlefield, strewn with broken flags and discarded weapons. The sky is consumed by a hail of arrows, dense as locusts. Countless black bolts, whistling with a sharp whistle, pierce the smoky air from all directions, plummeting alongside the general. The faint glow of distant war flames and a few burning arrows provided the only light source, casting flickering spots of light on the general's armor, freezing the violence and beauty of this moment, and creating an epic yet heartbreaking atmosphere of a hero's end.", + "seed": 27568, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_4/index_6881_seed_27568_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05654266611672938, + "vlm_logprob": -0.05654266611672938, + "vlm_yes_no_margin": 3.774999976158142, + "vlm_token_credit": -0.05654266611672938, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11858009546995163 + ], + "opposite_logprob_tokens": [ + -2.2435801029205322 + ], + "logprob_reward": -0.11858009546995163, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11858009546995163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23708485066890717, + 0.055579282343387604, + 0.05647556856274605, + 0.05722799152135849, + 0.05013526603579521, + 0.0388522669672966, + 0.29236042499542236, + 0.09226249903440475, + 0.036186713725328445, + 0.08383515477180481 + ], + "credit_map": [ + 0.03374409303069115, + 0.04246196895837784, + 0.03890688717365265, + 0.028562676161527634, + 0.02810213342308998, + 0.02698131464421749, + 0.022490214556455612, + 0.015835555270314217, + 0.007413955871015787, + 0.007826950401067734, + 0.007323546335101128, + 0.008182831108570099, + 0.007522877771407366, + 0.005720095243304968, + 0.005819999147206545, + 0.0057690260000526905, + 0.007687872741371393, + 0.009620487689971924, + 0.006366932298988104, + 0.008890162222087383, + 0.008245806209743023, + 0.006388742011040449, + 0.004900410771369934, + 0.004375155549496412, + 0.006899602245539427, + 0.0061913239769637585, + 0.009031124413013458, + 0.0065008229576051235, + 0.007199534680694342, + 0.0072872815653681755, + 0.007751475088298321, + 0.006366828456521034, + 0.004067668225616217, + 0.007673037238419056, + 0.008292910642921925, + 0.012457719072699547, + 0.00366222090087831, + 0.004424427170306444, + 0.004248329903930426, + 0.005308954045176506, + 0.0053747426718473434, + 0.008302087895572186, + 0.00490971514955163, + 0.0030341092497110367, + 0.005515235476195812, + 0.0036720565985888243, + 0.004167886916548014, + 0.0038764330092817545, + 0.022880515083670616, + 0.044978536665439606, + 0.03696439042687416, + 0.032978206872940063, + 0.03314245492219925, + 0.030215376988053322, + 0.057490911334753036, + 0.03371003642678261, + 0.010081044398248196, + 0.009581252001225948, + 0.017537228763103485, + 0.008260301314294338, + 0.01268580462783575, + 0.010520175099372864, + 0.012739907018840313, + 0.010856793262064457, + 0.0030900887213647366, + 0.003339464543387294, + 0.009066257625818253, + 0.0031899940222501755, + 0.004877293482422829, + 0.0065705496817827225, + 0.0024746155831962824, + 0.0035784475039690733, + 0.009104665368795395, + 0.008323706686496735, + 0.012347308918833733, + 0.00911713857203722, + 0.012958959676325321, + 0.01444661058485508, + 0.009822226129472256, + 0.007714536041021347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.11858009546995163 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799826443195343 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029799826443195343, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799826443195343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19473257660865784, + 0.0599011555314064, + 0.04098692536354065, + 0.029890786856412888, + 0.10701482743024826, + 0.028133461251854897, + 0.16531813144683838, + 0.19548532366752625, + 0.10547180473804474, + 0.07306496798992157 + ], + "credit_map": [ + 0.026480374857783318, + 0.03563031181693077, + 0.02328447811305523, + 0.023530203849077225, + 0.025360044091939926, + 0.023326152935624123, + 0.01938292384147644, + 0.01773807965219021, + 0.01003708690404892, + 0.007635863032191992, + 0.0081325089558959, + 0.00599804800003767, + 0.005154828075319529, + 0.006731295492500067, + 0.0053319246508181095, + 0.01087960135191679, + 0.004991253837943077, + 0.006335430312901735, + 0.005489920731633902, + 0.006160945165902376, + 0.004351544659584761, + 0.004823446739464998, + 0.005364628043025732, + 0.0034697572700679302, + 0.0038129098247736692, + 0.004534244071692228, + 0.004448332823812962, + 0.003463204950094223, + 0.0029486261773854494, + 0.0034264291170984507, + 0.003317189170047641, + 0.003939851187169552, + 0.008240610361099243, + 0.013867918401956558, + 0.016899414360523224, + 0.026571404188871384, + 0.008777283132076263, + 0.009896501898765564, + 0.010109679773449898, + 0.012652011588215828, + 0.003703904105350375, + 0.004155502188950777, + 0.002609175629913807, + 0.002426777733489871, + 0.0051522753201425076, + 0.0032548250164836645, + 0.004362626001238823, + 0.0024683757219463587, + 0.013630129396915436, + 0.023715557530522346, + 0.0229656845331192, + 0.02024855464696884, + 0.021800480782985687, + 0.019518472254276276, + 0.027008328586816788, + 0.01643092930316925, + 0.014187674038112164, + 0.05106757581233978, + 0.03181937336921692, + 0.019142484292387962, + 0.0211124736815691, + 0.019705576822161674, + 0.018479198217391968, + 0.019970975816249847, + 0.012636073864996433, + 0.011985084041953087, + 0.025733936578035355, + 0.010144224390387535, + 0.009729628451168537, + 0.01864626072347164, + 0.006238315720111132, + 0.010358289815485477, + 0.006800958886742592, + 0.0097648398950696, + 0.009233430027961731, + 0.007184530608355999, + 0.011560651473701, + 0.014602106995880604, + 0.007546770386397839, + 0.006371686700731516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.029799826443195343 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031993414741009474 + ], + "opposite_logprob_tokens": [ + -5.753199577331543 + ], + "logprob_reward": -0.0031993414741009474, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031993414741009474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2911391854286194, + 0.0648801252245903, + 0.07912231981754303, + 0.04227684810757637, + 0.11855094134807587, + 0.04741683974862099, + 0.12008985877037048, + 0.10348370671272278, + 0.046471647918224335, + 0.08656850457191467 + ], + "credit_map": [ + 0.041649214923381805, + 0.054171036928892136, + 0.03522442653775215, + 0.04097236692905426, + 0.03653307259082794, + 0.033424779772758484, + 0.022564472630620003, + 0.026599839329719543, + 0.010853475891053677, + 0.007861326448619366, + 0.008716108277440071, + 0.010829845443367958, + 0.00743675883859396, + 0.005546425469219685, + 0.0074136522598564625, + 0.0062225316651165485, + 0.011503099463880062, + 0.010858269408345222, + 0.009222433902323246, + 0.011515502817928791, + 0.011373011395335197, + 0.010661013424396515, + 0.006771279498934746, + 0.007217709440737963, + 0.007408303674310446, + 0.0057542771100997925, + 0.005567863583564758, + 0.005151214078068733, + 0.005484743043780327, + 0.004249536897987127, + 0.004585322458297014, + 0.004075587261468172, + 0.00961565226316452, + 0.016962667927145958, + 0.018712876364588737, + 0.030555810779333115, + 0.009091923013329506, + 0.011255361139774323, + 0.009939393028616905, + 0.012417256832122803, + 0.00676231924444437, + 0.006788593251258135, + 0.005055240821093321, + 0.004167578648775816, + 0.01123275887221098, + 0.005399999208748341, + 0.004265945870429277, + 0.003744404297322035, + 0.010509814135730267, + 0.013352124951779842, + 0.01802828535437584, + 0.010520190931856632, + 0.015263918787240982, + 0.014314297586679459, + 0.024050723761320114, + 0.014050502330064774, + 0.008486242964863777, + 0.02381419762969017, + 0.015582851134240627, + 0.005642462056130171, + 0.011130375787615776, + 0.014841760508716106, + 0.014899791218340397, + 0.00908602587878704, + 0.007928028702735901, + 0.0058621917851269245, + 0.007194379344582558, + 0.0049943928606808186, + 0.00544534157961607, + 0.006918503902852535, + 0.003577328287065029, + 0.0045514837838709354, + 0.0072990660555660725, + 0.01266736164689064, + 0.011892091482877731, + 0.01037018932402134, + 0.01224153395742178, + 0.012702463194727898, + 0.012045999988913536, + 0.007349799387156963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0031993414741009474 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040924143977463245 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.0040924143977463245, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040924143977463245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16434694826602936, + 0.08791186660528183, + 0.054960720241069794, + 0.11989466100931168, + 0.06570892781019211, + 0.032200757414102554, + 0.08781874179840088, + 0.11608271300792694, + 0.17503871023654938, + 0.09603597968816757 + ], + "credit_map": [ + 0.024435682222247124, + 0.02615184523165226, + 0.02132136933505535, + 0.023322001099586487, + 0.02073737047612667, + 0.019666917622089386, + 0.01473232265561819, + 0.013979430310428143, + 0.009977077133953571, + 0.01572336256504059, + 0.010679914616048336, + 0.014391789212822914, + 0.012008375488221645, + 0.008416129276156425, + 0.008440242148935795, + 0.008274964988231659, + 0.00568645354360342, + 0.007514247205108404, + 0.00622576754540205, + 0.007546157576143742, + 0.008039891719818115, + 0.00693484116345644, + 0.0059018987230956554, + 0.0071114529855549335, + 0.022004075348377228, + 0.019005363807082176, + 0.02013111487030983, + 0.018359920009970665, + 0.013077891431748867, + 0.00883899349719286, + 0.01012349221855402, + 0.008353805169463158, + 0.005270110908895731, + 0.006666290108114481, + 0.011537086218595505, + 0.017980212345719337, + 0.003852115012705326, + 0.005345643498003483, + 0.00906088761985302, + 0.005996582563966513, + 0.002987185027450323, + 0.003343746764585376, + 0.0034060326870530844, + 0.003781235311180353, + 0.0070695471949875355, + 0.00431998074054718, + 0.004332083743065596, + 0.0029609405901283026, + 0.00838417373597622, + 0.008285810239613056, + 0.009220127947628498, + 0.010346491821110249, + 0.01069747656583786, + 0.01191522553563118, + 0.018401406705379486, + 0.010568017140030861, + 0.011450260877609253, + 0.013247969560325146, + 0.017949294298887253, + 0.00932989176362753, + 0.015681710094213486, + 0.016065914183855057, + 0.020824335515499115, + 0.01153331995010376, + 0.01757231540977955, + 0.009229473769664764, + 0.058911554515361786, + 0.010554283857345581, + 0.013646022416651249, + 0.04335494339466095, + 0.0073716058395802975, + 0.014398500323295593, + 0.007936595007777214, + 0.012810343876481056, + 0.01477727759629488, + 0.012067553587257862, + 0.013739265501499176, + 0.01400167029350996, + 0.012947756797075272, + 0.007755505386739969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.0040924143977463245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young general clad in silver armor with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the general falling backward from a panicked black warhorse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the general's face stained with dust and a streak of blood while his hair unfurls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast muddy battlefield with broken flags and discarded weapons under a hail of arrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot capturing the scene in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704165279865265 + ], + "opposite_logprob_tokens": [ + -2.1270415782928467 + ], + "logprob_reward": -0.12704165279865265, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12704165279865265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06551452726125717, + 0.06519950926303864, + 0.05921096354722977, + 0.050428666174411774, + 0.03609306737780571, + 0.04482749104499817, + 0.13351617753505707, + 0.12680715322494507, + 0.23082612454891205, + 0.1875763237476349 + ], + "credit_map": [ + 0.009153183549642563, + 0.00893336907029152, + 0.00789730530232191, + 0.010846760123968124, + 0.008018489927053452, + 0.008092635311186314, + 0.006812223233282566, + 0.005760564003139734, + 0.007921664975583553, + 0.010456179268658161, + 0.0087627824395895, + 0.010211453773081303, + 0.007783901412039995, + 0.004612782970070839, + 0.007259948179125786, + 0.008190800435841084, + 0.010559022426605225, + 0.008443756029009819, + 0.007822214625775814, + 0.007613624911755323, + 0.009347968734800816, + 0.005837169475853443, + 0.004907211288809776, + 0.004679999314248562, + 0.009976676665246487, + 0.010226337239146233, + 0.0052008549682796, + 0.005290407221764326, + 0.007986507378518581, + 0.003039638977497816, + 0.004267743322998285, + 0.004440500400960445, + 0.004559500142931938, + 0.0040930453687906265, + 0.004968465771526098, + 0.0060558137483894825, + 0.003366198157891631, + 0.004816393833607435, + 0.004751839675009251, + 0.0034818097483366728, + 0.005827174987643957, + 0.00558999041095376, + 0.0060470225289464, + 0.006087560672312975, + 0.006834082771092653, + 0.005523881874978542, + 0.0044727809727191925, + 0.004444997292011976, + 0.0130483228713274, + 0.015693197026848793, + 0.017880354076623917, + 0.01668204739689827, + 0.019468624144792557, + 0.019619597122073174, + 0.018755923956632614, + 0.012368111871182919, + 0.010696487501263618, + 0.03187352791428566, + 0.01930169388651848, + 0.010115491226315498, + 0.012768905609846115, + 0.014189663343131542, + 0.014875483699142933, + 0.0129858972504735, + 0.02231902815401554, + 0.01580367423593998, + 0.07121706008911133, + 0.0166474636644125, + 0.020630573853850365, + 0.04942738637328148, + 0.011124351061880589, + 0.02365659363567829, + 0.018886109814047813, + 0.018385782837867737, + 0.029973939061164856, + 0.02392594702541828, + 0.030662085860967636, + 0.023327436298131943, + 0.02668311819434166, + 0.01573190838098526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.12704165279865265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_5/index_tensor([1152, 1152])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_5/index_tensor([1152, 1152])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c3f7e640d118903bf0e1074387c4ef2aa5402de0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_5/index_tensor([1152, 1152])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long, uninterrupted, dynamic tracking shot locks onto and circles a weathered wooden easel spinning rapidly in mid-air. The old easel's surface is covered in scratches, dents, and the fine cracks of age. Layers of thick, dried-up paint, from deep Prussian blue and fiery cadmium red to brilliant lemon yellow and pure titanium white, some of which have cracked, creating a mottled texture. Amidst a pure black, boundless void, the easel undergoes a wild, erratic rotation about its central axis. With each violent tumble and twist, the intense centrifugal force hurls the still-dry, viscous paint, creating dazzling streaks of color and splashes of liquid against the inky black background. These color trails sometimes resemble the slender, brilliant tails of a comet; sometimes, like exploding fireworks, they burst forth into countless tiny specks of color. The camera follows steadily and persistently, capturing every moment of color splashing. The entire picture is full of uncontrolled dynamic beauty and pure creative energy, like a silent but intense color symphony.", + "seed": 31506, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_5/index_1152_seed_31506_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11593894474208355, + "vlm_logprob": -0.11593894474208355, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.11593894474208355, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22957774996757507 + ], + "opposite_logprob_tokens": [ + -1.6045777797698975 + ], + "logprob_reward": -0.22957774996757507, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22957774996757507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5287230610847473, + 0.03104267828166485, + 0.06564462929964066, + 0.028916554525494576, + 0.046251747757196426, + 0.12288980931043625, + 0.03382068872451782, + 0.024752328172326088, + 0.09692106395959854, + 0.021037450060248375 + ], + "credit_map": [ + 0.06755343824625015, + 0.04093599319458008, + 0.03303774818778038, + 0.18106867372989655, + 0.026578836143016815, + 0.04361468181014061, + 0.05045846104621887, + 0.08547525852918625, + 0.0036629235837608576, + 0.0049622454680502415, + 0.0024138689041137695, + 0.0037787752225995064, + 0.003779760329052806, + 0.00433052983134985, + 0.0029090577736496925, + 0.005205516237765551, + 0.004247893113642931, + 0.0037963283248245716, + 0.006129342131316662, + 0.004167912062257528, + 0.005307427141815424, + 0.01377771608531475, + 0.0085938461124897, + 0.01962416060268879, + 0.0024985461495816708, + 0.0032869139686226845, + 0.004549430683255196, + 0.0025787390768527985, + 0.0028031414840370417, + 0.00466596195474267, + 0.003251704154536128, + 0.005282118450850248, + 0.004211890511214733, + 0.004951550625264645, + 0.003607621183618903, + 0.005613549146801233, + 0.0033859023824334145, + 0.008828441612422466, + 0.004080368205904961, + 0.011572420597076416, + 0.006561288610100746, + 0.012463191524147987, + 0.007875971496105194, + 0.02454245276749134, + 0.008162256330251694, + 0.02010183595120907, + 0.013980642892420292, + 0.029202155768871307, + 0.0046106926165521145, + 0.0038658822886645794, + 0.00573283527046442, + 0.005215405486524105, + 0.0032406526152044535, + 0.0033837498631328344, + 0.0031079943291842937, + 0.004663475789129734, + 0.0034155757166445255, + 0.002593275625258684, + 0.0036578739527612925, + 0.002425386104732752, + 0.00272352434694767, + 0.004678283352404833, + 0.0025143076200038195, + 0.0027441028505563736, + 0.010366976261138916, + 0.010866740718483925, + 0.012063642963767052, + 0.01663237251341343, + 0.009242170490324497, + 0.015161716379225254, + 0.011289882473647594, + 0.011297557502985, + 0.0017768448451533914, + 0.002288712887093425, + 0.0048125507310032845, + 0.001297958311624825, + 0.0014759397599846125, + 0.0039517926052212715, + 0.00173173937946558, + 0.0037019101437181234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.22957774996757507 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604701578617096 + ], + "opposite_logprob_tokens": [ + -1.9104701280593872 + ], + "logprob_reward": -0.1604701578617096, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604701578617096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18282778561115265, + 0.06868834048509598, + 0.1347903460264206, + 0.08998186141252518, + 0.07052287459373474, + 0.16814561188220978, + 0.055050574243068695, + 0.05219872668385506, + 0.07382170855998993, + 0.1039721742272377 + ], + "credit_map": [ + 0.023633018136024475, + 0.015805482864379883, + 0.012668682262301445, + 0.05215020850300789, + 0.009472724981606007, + 0.019602879881858826, + 0.015589478425681591, + 0.03390532732009888, + 0.0059236520901322365, + 0.008729994297027588, + 0.006484510842710733, + 0.009198336862027645, + 0.007590030319988728, + 0.012769521214067936, + 0.006940614432096481, + 0.011051684617996216, + 0.007098858710378408, + 0.0057765464298427105, + 0.012868267484009266, + 0.008199756033718586, + 0.010119747370481491, + 0.02773413620889187, + 0.016864342615008354, + 0.04612868279218674, + 0.0074574886821210384, + 0.008421956561505795, + 0.013420721516013145, + 0.01275317370891571, + 0.00857526995241642, + 0.013729321770370007, + 0.010283559560775757, + 0.015340368263423443, + 0.005953111685812473, + 0.007206663955003023, + 0.007063011173158884, + 0.00932654645293951, + 0.0080967266112566, + 0.012283443473279476, + 0.0052826679311692715, + 0.015310709364712238, + 0.00946066528558731, + 0.018421480432152748, + 0.013038506731390953, + 0.031255513429641724, + 0.013131356798112392, + 0.02431677281856537, + 0.026222536340355873, + 0.03229876980185509, + 0.007384964264929295, + 0.006299471482634544, + 0.009336942806839943, + 0.007877996191382408, + 0.004580468405038118, + 0.006573656108230352, + 0.00521220825612545, + 0.007784866727888584, + 0.005880068987607956, + 0.0050915866158902645, + 0.007539501879364252, + 0.004426011350005865, + 0.009236001409590244, + 0.009233449585735798, + 0.005559495184570551, + 0.005232612602412701, + 0.007800903171300888, + 0.008333391509950161, + 0.00880045723170042, + 0.008787722326815128, + 0.01054992713034153, + 0.010593464598059654, + 0.007737338542938232, + 0.011218506842851639, + 0.006963783875107765, + 0.006526942364871502, + 0.030194148421287537, + 0.004338046070188284, + 0.005075009539723396, + 0.024005847051739693, + 0.0038865034002810717, + 0.022981887683272362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.1604701578617096 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008654113858938217 + ], + "opposite_logprob_tokens": [ + -4.7586541175842285 + ], + "logprob_reward": -0.008654113858938217, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008654113858938217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09706812351942062, + 0.07972341030836105, + 0.07628113031387329, + 0.12183458358049393, + 0.1390741467475891, + 0.13913187384605408, + 0.07154049724340439, + 0.08451399207115173, + 0.09451192617416382, + 0.09632028639316559 + ], + "credit_map": [ + 0.012434489093720913, + 0.01271770242601633, + 0.007656881120055914, + 0.019280336797237396, + 0.006917744874954224, + 0.011685337871313095, + 0.01026149746030569, + 0.016114136204123497, + 0.009408882819116116, + 0.01242553349584341, + 0.006682348903268576, + 0.006295212544500828, + 0.007918021641671658, + 0.013366264291107655, + 0.008294114843010902, + 0.015333037823438644, + 0.006843132898211479, + 0.006957357283681631, + 0.012109706178307533, + 0.007913870736956596, + 0.008361761458218098, + 0.01169538777321577, + 0.009585937485098839, + 0.01281398069113493, + 0.009529202245175838, + 0.012525724247097969, + 0.01035004761070013, + 0.012965883128345013, + 0.01644887588918209, + 0.021501045674085617, + 0.017133386805653572, + 0.021380439400672913, + 0.013488168828189373, + 0.013974493369460106, + 0.014081182889640331, + 0.02401825040578842, + 0.011201555840671062, + 0.023684414103627205, + 0.010104403831064701, + 0.028521686792373657, + 0.011787203140556812, + 0.014469354413449764, + 0.016482079401612282, + 0.019463106989860535, + 0.017782116308808327, + 0.013229679316282272, + 0.024983689188957214, + 0.020934654399752617, + 0.008288090117275715, + 0.005442311521619558, + 0.014331704936921597, + 0.010552520863711834, + 0.007021858356893063, + 0.006525055505335331, + 0.009179587475955486, + 0.010199382901191711, + 0.008657574653625488, + 0.009003262035548687, + 0.013037562370300293, + 0.005895696580410004, + 0.009007264859974384, + 0.018827134743332863, + 0.009016490541398525, + 0.011069020256400108, + 0.00962943397462368, + 0.011693591251969337, + 0.011183517053723335, + 0.01623186282813549, + 0.010168753564357758, + 0.013753199018537998, + 0.010764491744339466, + 0.01108709629625082, + 0.010117856785655022, + 0.006871480029076338, + 0.027723725885152817, + 0.00659154262393713, + 0.007611560169607401, + 0.01755240187048912, + 0.006445628125220537, + 0.013406096957623959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.008654113858938217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038984060287476 + ], + "opposite_logprob_tokens": [ + -1.9103899002075195 + ], + "logprob_reward": -0.16038984060287476, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038984060287476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3781835436820984, + 0.0467604361474514, + 0.1094663143157959, + 0.07343699783086777, + 0.17894022166728973, + 0.05592837184667587, + 0.018508143723011017, + 0.04162359982728958, + 0.043072741478681564, + 0.05407963693141937 + ], + "credit_map": [ + 0.04515564441680908, + 0.026224525645375252, + 0.02450404316186905, + 0.12884563207626343, + 0.019004151225090027, + 0.033582478761672974, + 0.03592729568481445, + 0.06493981927633286, + 0.0045858267694711685, + 0.0077054440043866634, + 0.0032061634119600058, + 0.0053278617560863495, + 0.007032929453998804, + 0.0061168973334133625, + 0.005159798078238964, + 0.007625514175742865, + 0.00813209917396307, + 0.006427709478884935, + 0.0084943026304245, + 0.008319512009620667, + 0.01146000437438488, + 0.01942078024148941, + 0.014813858084380627, + 0.0323980487883091, + 0.005072299391031265, + 0.00835984293371439, + 0.008129751309752464, + 0.010118624195456505, + 0.008258593268692493, + 0.011064620688557625, + 0.00813447218388319, + 0.014298791065812111, + 0.01564117707312107, + 0.014346407726407051, + 0.01595841720700264, + 0.024363625794649124, + 0.02049715258181095, + 0.03446032851934433, + 0.01154286228120327, + 0.042130254209041595, + 0.003267906606197357, + 0.006117294076830149, + 0.003905809950083494, + 0.011656207963824272, + 0.003443864407017827, + 0.008545397780835629, + 0.007586401421576738, + 0.011405483819544315, + 0.002646118402481079, + 0.0016486543463543057, + 0.004204879514873028, + 0.0015819495311006904, + 0.0019697684329003096, + 0.001992003759369254, + 0.0013374180998653173, + 0.003127351403236389, + 0.005324652884155512, + 0.0036100184079259634, + 0.006210219115018845, + 0.002359446370974183, + 0.0060349926352500916, + 0.010142596438527107, + 0.0033869664184749126, + 0.004554702900350094, + 0.005257940851151943, + 0.003255419433116913, + 0.00681029399856925, + 0.0022956696338951588, + 0.003374930238351226, + 0.010982027277350426, + 0.0030840651597827673, + 0.008012397214770317, + 0.0043550762347877026, + 0.002921435981988907, + 0.01894206553697586, + 0.003316950285807252, + 0.00230622966773808, + 0.011930504813790321, + 0.003236617660149932, + 0.00707075884565711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.16038984060287476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602861419320107 + ], + "opposite_logprob_tokens": [ + -3.8956029415130615 + ], + "logprob_reward": -0.020602861419320107, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602861419320107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18676064908504486, + 0.0422939769923687, + 0.07455559819936752, + 0.13716740906238556, + 0.12645606696605682, + 0.0764281153678894, + 0.044138792902231216, + 0.11427277326583862, + 0.08454208076000214, + 0.11338446289300919 + ], + "credit_map": [ + 0.01982405036687851, + 0.015563996508717537, + 0.014049527235329151, + 0.058736659586429596, + 0.012637478299438953, + 0.016707776114344597, + 0.016864769160747528, + 0.03237641975283623, + 0.005215456709265709, + 0.005461821798235178, + 0.004012511111795902, + 0.005471935961395502, + 0.005527960602194071, + 0.007104006130248308, + 0.0033023108262568712, + 0.0061979773454368114, + 0.007690018508583307, + 0.004976848140358925, + 0.011339981108903885, + 0.0065590678714215755, + 0.010007739067077637, + 0.01143084280192852, + 0.009478539228439331, + 0.01307256706058979, + 0.01144888810813427, + 0.011273952201008797, + 0.025365281850099564, + 0.016229964792728424, + 0.011022770777344704, + 0.019147366285324097, + 0.018971599638462067, + 0.023707587271928787, + 0.013154597021639347, + 0.011038162745535374, + 0.01296915952116251, + 0.012484055943787098, + 0.020185599103569984, + 0.023702723905444145, + 0.010011044330894947, + 0.022910727187991142, + 0.0052259028889238834, + 0.00839706975966692, + 0.00988712813705206, + 0.011340660974383354, + 0.007547757588326931, + 0.00800982117652893, + 0.017186889424920082, + 0.008832893334329128, + 0.004931480623781681, + 0.004934796132147312, + 0.006886619608849287, + 0.004987969528883696, + 0.004911760333925486, + 0.004824856761842966, + 0.0058997212909162045, + 0.0067615932784974575, + 0.012720034457743168, + 0.014692578464746475, + 0.018905693665146828, + 0.010327167809009552, + 0.016868755221366882, + 0.0198980662971735, + 0.010165243409574032, + 0.010695230215787888, + 0.009197385050356388, + 0.006968614179641008, + 0.012543411925435066, + 0.01113772951066494, + 0.008249180391430855, + 0.01474430039525032, + 0.011248644441366196, + 0.010452816262841225, + 0.01200848538428545, + 0.00875782035291195, + 0.02868369035422802, + 0.008462532423436642, + 0.013005645014345646, + 0.02067902311682701, + 0.009035731665790081, + 0.012751549482345581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.020602861419320107 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, uninterrupted, dynamic tracking shot locks onto and circles a weathered wooden easel spinning rapidly in mid-air. The old easel's surface is covered in scratches, dents, and the fine cracks of age. Layers of thick, dried-up paint, from deep Prussian blue and fiery cadmium red to brilliant lemon yellow and pure titanium white, some of which have cracked, creating a mottled texture. Amidst a pure black, boundless void, the easel undergoes a wild, erratic rotation about its central axis. With each violent tumble and twist, the intense centrifugal force hurls the still-dry, viscous paint, creating dazzling streaks of color and splashes of liquid against the inky black background. These color trails sometimes resemble the slender, brilliant tails of a comet; sometimes, like exploding fireworks, they burst forth into countless tiny specks of color. The camera follows steadily and persistently, capturing every moment of color splashing. The entire picture is full of uncontrolled dynamic beauty and pure creative energy, like a silent but intense color symphony.", + "seed": 31507, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_5/index_1152_seed_31507_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2167865749448538, + "vlm_logprob": -0.2167865749448538, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.2167865749448538, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28697723150253296 + ], + "opposite_logprob_tokens": [ + -1.4119772911071777 + ], + "logprob_reward": -0.28697723150253296, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28697723150253296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16010715067386627, + 0.17572662234306335, + 0.08336588740348816, + 0.028516989201307297, + 0.05269191786646843, + 0.06750203669071198, + 0.09621021151542664, + 0.23350146412849426, + 0.04960091412067413, + 0.05277681723237038 + ], + "credit_map": [ + 0.027655819430947304, + 0.01527363434433937, + 0.025486478582024574, + 0.022859539836645126, + 0.01498075108975172, + 0.018619468435645103, + 0.019484179094433784, + 0.015747280791401863, + 0.019955040886998177, + 0.018670158460736275, + 0.021495170891284943, + 0.019849209114909172, + 0.017087625339627266, + 0.025188853964209557, + 0.03291963413357735, + 0.02056092768907547, + 0.007110669743269682, + 0.0067512765526771545, + 0.007551726885139942, + 0.006598359905183315, + 0.006521913222968578, + 0.021632157266139984, + 0.01785900443792343, + 0.009340780787169933, + 0.00286472518928349, + 0.002822753507643938, + 0.004066006280481815, + 0.0027591988909989595, + 0.0037663148250430822, + 0.003850275883451104, + 0.0030654622241854668, + 0.005322253797203302, + 0.008770283311605453, + 0.005021118093281984, + 0.005681959912180901, + 0.004718891344964504, + 0.006660391576588154, + 0.009979015216231346, + 0.006546689197421074, + 0.005313564091920853, + 0.011590275913476944, + 0.0061105042695999146, + 0.006919639650732279, + 0.0064835213124752045, + 0.006394743453711271, + 0.01570441946387291, + 0.00634963670745492, + 0.007949301972985268, + 0.019885027781128883, + 0.017687110230326653, + 0.009673193097114563, + 0.01172694656997919, + 0.009673882275819778, + 0.008148339577019215, + 0.007732430472970009, + 0.011683276854455471, + 0.04669847711920738, + 0.013211768120527267, + 0.013456946238875389, + 0.015893209725618362, + 0.04591953009366989, + 0.016642453148961067, + 0.021795913577079773, + 0.05988314747810364, + 0.0056153833866119385, + 0.006968795787543058, + 0.006054609548300505, + 0.005662539042532444, + 0.005497622769325972, + 0.007374970242381096, + 0.006028900388628244, + 0.006398097146302462, + 0.005649520084261894, + 0.00742819020524621, + 0.007310142740607262, + 0.006358691956847906, + 0.007546587381511927, + 0.0067809647880494595, + 0.007085259538143873, + 0.004617460072040558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.28697723150253296 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10044995695352554 + ], + "opposite_logprob_tokens": [ + -2.350450038909912 + ], + "logprob_reward": -0.10044995695352554, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10044995695352554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2679094970226288, + 0.2228812873363495, + 0.05604427680373192, + 0.051243819296360016, + 0.044749241322278976, + 0.07810313999652863, + 0.1092727854847908, + 0.03985334932804108, + 0.06535714119672775, + 0.06458541750907898 + ], + "credit_map": [ + 0.045134853571653366, + 0.029124237596988678, + 0.0446954146027565, + 0.03870994597673416, + 0.029415123164653778, + 0.025902938097715378, + 0.03297518938779831, + 0.02195183001458645, + 0.02308782748878002, + 0.020925801247358322, + 0.024338139221072197, + 0.020312119275331497, + 0.022016353905200958, + 0.030771829187870026, + 0.032927826046943665, + 0.048501379787921906, + 0.0046234834007918835, + 0.007497173734009266, + 0.006842286791652441, + 0.007138521410524845, + 0.005249998066574335, + 0.008519638329744339, + 0.008188205771148205, + 0.007984976284205914, + 0.00428070267662406, + 0.003990781959146261, + 0.006845137104392052, + 0.005306858103722334, + 0.004518055822700262, + 0.005776797421276569, + 0.0055254921317100525, + 0.014999997802078724, + 0.011127425357699394, + 0.005046486388891935, + 0.0044011808931827545, + 0.002790895290672779, + 0.0038667120970785618, + 0.008365320041775703, + 0.0035770046524703503, + 0.005574219394475222, + 0.013121433556079865, + 0.00827030185610056, + 0.009581286460161209, + 0.008100578561425209, + 0.008022165857255459, + 0.011913809925317764, + 0.009287495166063309, + 0.009806076996028423, + 0.01979646272957325, + 0.016755178570747375, + 0.012041803449392319, + 0.01422853022813797, + 0.013450365513563156, + 0.009174958802759647, + 0.010717376135289669, + 0.013108114711940289, + 0.0058777593076229095, + 0.004291865974664688, + 0.003314334200695157, + 0.005865823477506638, + 0.005368738900870085, + 0.0039054695516824722, + 0.0044702449813485146, + 0.006759115494787693, + 0.007880579680204391, + 0.009456970728933811, + 0.008327596820890903, + 0.00649141613394022, + 0.005472698714584112, + 0.012366972863674164, + 0.005348621401935816, + 0.01001229602843523, + 0.008835503831505775, + 0.009254506789147854, + 0.007998418062925339, + 0.006905296817421913, + 0.009248966351151466, + 0.008863587863743305, + 0.007812950760126114, + 0.00566619448363781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.10044995695352554 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008658958598971367 + ], + "opposite_logprob_tokens": [ + -4.7586588859558105 + ], + "logprob_reward": -0.008658958598971367, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008658958598971367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19372378289699554, + 0.1947939693927765, + 0.05846326798200607, + 0.03232187777757645, + 0.04311911016702652, + 0.09465894848108292, + 0.048437051475048065, + 0.2213832288980484, + 0.043636590242385864, + 0.06946220248937607 + ], + "credit_map": [ + 0.03571363911032677, + 0.024649852886795998, + 0.03488357737660408, + 0.027292102575302124, + 0.019430184736847878, + 0.013594402931630611, + 0.02185189351439476, + 0.016308117657899857, + 0.023526478558778763, + 0.01970633864402771, + 0.022596588358283043, + 0.020538531243801117, + 0.02248194068670273, + 0.028397999703884125, + 0.029262570664286613, + 0.028283515945076942, + 0.004618421196937561, + 0.008083566091954708, + 0.006360108032822609, + 0.007097621448338032, + 0.008028926327824593, + 0.0086513701826334, + 0.008932237513363361, + 0.00669101532548666, + 0.004300527740269899, + 0.004133368842303753, + 0.0038956725038588047, + 0.0037192644085735083, + 0.004481657408177853, + 0.0034083847422152758, + 0.004959185142070055, + 0.0034238165244460106, + 0.005965733900666237, + 0.004404434934258461, + 0.004521379247307777, + 0.00472658546641469, + 0.003618786344304681, + 0.009950829669833183, + 0.005480935797095299, + 0.004450424574315548, + 0.014851221814751625, + 0.0091682905331254, + 0.01111859641969204, + 0.010133977979421616, + 0.010081418789923191, + 0.016818182542920113, + 0.009348085150122643, + 0.013139167800545692, + 0.008266974240541458, + 0.007897371426224709, + 0.005197054706513882, + 0.006366550922393799, + 0.005110705737024546, + 0.004679353907704353, + 0.005124224815517664, + 0.005794817116111517, + 0.04205682873725891, + 0.013214143924415112, + 0.014012462459504604, + 0.01832379400730133, + 0.04119241237640381, + 0.017980070784687996, + 0.020754899829626083, + 0.05384860560297966, + 0.004589694086462259, + 0.0033839689567685127, + 0.004345076624304056, + 0.005292525049299002, + 0.00408801157027483, + 0.009921673685312271, + 0.004613560624420643, + 0.0074020749889314175, + 0.00892366748303175, + 0.010375159792602062, + 0.007509299088269472, + 0.010120321065187454, + 0.010212150402367115, + 0.009193588979542255, + 0.007633317727595568, + 0.0054946960881352425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.008658958598971367 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328523755073547 + ], + "opposite_logprob_tokens": [ + -0.7578523755073547 + ], + "logprob_reward": -0.6328523755073547, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328523755073547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2857522666454315, + 0.16862043738365173, + 0.06502073258161545, + 0.07669885456562042, + 0.028770118951797485, + 0.16375602781772614, + 0.0319625623524189, + 0.08157528936862946, + 0.07789726555347443, + 0.019946422427892685 + ], + "credit_map": [ + 0.05158582329750061, + 0.03169047832489014, + 0.04985351115465164, + 0.03301975876092911, + 0.02990007773041725, + 0.03552043437957764, + 0.02916146069765091, + 0.025020744651556015, + 0.022579139098525047, + 0.018751811236143112, + 0.019726749509572983, + 0.01683814264833927, + 0.021749980747699738, + 0.0252027940005064, + 0.022452939301729202, + 0.02131887711584568, + 0.007042421028017998, + 0.0057826717384159565, + 0.009894230403006077, + 0.0061829909682273865, + 0.006893644575029612, + 0.012515338137745857, + 0.00685584032908082, + 0.009853597730398178, + 0.007132741156965494, + 0.004496114328503609, + 0.02125387080013752, + 0.007865918800234795, + 0.006179118528962135, + 0.010510515421628952, + 0.008250986225903034, + 0.011009599082171917, + 0.006794670596718788, + 0.0038407042156904936, + 0.003156619379296899, + 0.0026731325779110193, + 0.0026182944420725107, + 0.0031200747471302748, + 0.0026298193261027336, + 0.0039368001744151115, + 0.025957638397812843, + 0.017252115532755852, + 0.015526801347732544, + 0.014871631748974323, + 0.01646689511835575, + 0.02994304709136486, + 0.0214786846190691, + 0.02225920930504799, + 0.005932857282459736, + 0.004478496499359608, + 0.004953100346028805, + 0.003764946712180972, + 0.0030245266389101744, + 0.002782190451398492, + 0.0027751349844038486, + 0.004251307342201471, + 0.016038551926612854, + 0.00466026458889246, + 0.005031351465731859, + 0.005898202769458294, + 0.015349755994975567, + 0.006373515352606773, + 0.007649058010429144, + 0.02057458646595478, + 0.008107886649668217, + 0.007856138981878757, + 0.00898858904838562, + 0.007299571763724089, + 0.006162538658827543, + 0.019664134830236435, + 0.006377682089805603, + 0.013440719805657864, + 0.0031853774562478065, + 0.0019737069960683584, + 0.002389369299635291, + 0.0025618025101721287, + 0.0021067808847874403, + 0.002897674450650811, + 0.0025280858390033245, + 0.0023036261554807425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.6328523755073547 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499435216188431 + ], + "opposite_logprob_tokens": [ + -2.9299943447113037 + ], + "logprob_reward": -0.05499435216188431, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499435216188431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13718870282173157, + 0.08884671330451965, + 0.11798474937677383, + 0.08917362987995148, + 0.06802305579185486, + 0.14668427407741547, + 0.10171139240264893, + 0.06614558398723602, + 0.06459604203701019, + 0.11964590102434158 + ], + "credit_map": [ + 0.026958482339978218, + 0.014057554304599762, + 0.02499230206012726, + 0.01597624644637108, + 0.01079501397907734, + 0.017334241420030594, + 0.013644363731145859, + 0.013430498540401459, + 0.009541686624288559, + 0.00781808141618967, + 0.010781320743262768, + 0.008109469898045063, + 0.008437449112534523, + 0.013823042623698711, + 0.01564442366361618, + 0.014691240154206753, + 0.008712400682270527, + 0.007709984201937914, + 0.008387400768697262, + 0.009912163019180298, + 0.011042162775993347, + 0.029023826122283936, + 0.02449674904346466, + 0.018700072541832924, + 0.007730885874480009, + 0.005621367134153843, + 0.016390599310398102, + 0.009698386304080486, + 0.008352965116500854, + 0.01249298918992281, + 0.011178577318787575, + 0.017707861959934235, + 0.012386472895741463, + 0.00598550122231245, + 0.007563586812466383, + 0.004839944653213024, + 0.00760364904999733, + 0.011765480041503906, + 0.00805850699543953, + 0.009819909930229187, + 0.022208770737051964, + 0.01646536961197853, + 0.018738238140940666, + 0.013979126699268818, + 0.01437325682491064, + 0.023691410198807716, + 0.019224639981985092, + 0.01800346188247204, + 0.019039232283830643, + 0.016971709206700325, + 0.012370545417070389, + 0.013377521187067032, + 0.008828740566968918, + 0.009062426164746284, + 0.008847747929394245, + 0.013213468715548515, + 0.013157355599105358, + 0.004718896932899952, + 0.003949692007154226, + 0.0046370308846235275, + 0.013827088288962841, + 0.004120649304240942, + 0.006769346073269844, + 0.014965523965656757, + 0.005543553736060858, + 0.007622966542840004, + 0.005239457357674837, + 0.006845999043434858, + 0.005532999988645315, + 0.01626673899590969, + 0.005865059792995453, + 0.011679265648126602, + 0.01629430241882801, + 0.020226778462529182, + 0.011756154708564281, + 0.01724029891192913, + 0.018594102934002876, + 0.016112102195620537, + 0.010405056178569794, + 0.009017093107104301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.05499435216188431 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, uninterrupted, dynamic tracking shot locks onto and circles a weathered wooden easel spinning rapidly in mid-air. The old easel's surface is covered in scratches, dents, and the fine cracks of age. Layers of thick, dried-up paint, from deep Prussian blue and fiery cadmium red to brilliant lemon yellow and pure titanium white, some of which have cracked, creating a mottled texture. Amidst a pure black, boundless void, the easel undergoes a wild, erratic rotation about its central axis. With each violent tumble and twist, the intense centrifugal force hurls the still-dry, viscous paint, creating dazzling streaks of color and splashes of liquid against the inky black background. These color trails sometimes resemble the slender, brilliant tails of a comet; sometimes, like exploding fireworks, they burst forth into countless tiny specks of color. The camera follows steadily and persistently, capturing every moment of color splashing. The entire picture is full of uncontrolled dynamic beauty and pure creative energy, like a silent but intense color symphony.", + "seed": 31508, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_5/index_1152_seed_31508_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1310387911275029, + "vlm_logprob": -0.1310387911275029, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.1310387911275029, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3540556728839874 + ], + "opposite_logprob_tokens": [ + -1.229055643081665 + ], + "logprob_reward": -0.3540556728839874, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3540556728839874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19468842446804047, + 0.06519850343465805, + 0.05110475793480873, + 0.08811173588037491, + 0.15826299786567688, + 0.14777831733226776, + 0.05486350506544113, + 0.1081600934267044, + 0.0736396312713623, + 0.05819204822182655 + ], + "credit_map": [ + 0.020535020157694817, + 0.015710171312093735, + 0.01676298677921295, + 0.025263570249080658, + 0.018705138936638832, + 0.03536335006356239, + 0.04232772812247276, + 0.020020434632897377, + 0.009792509488761425, + 0.005476417485624552, + 0.008533882908523083, + 0.006395959760993719, + 0.007276126649230719, + 0.009238095954060555, + 0.010468774475157261, + 0.008016734384000301, + 0.007024820894002914, + 0.006888887844979763, + 0.0103793665766716, + 0.004535662941634655, + 0.00485398480668664, + 0.007340529467910528, + 0.00428575836122036, + 0.005795745179057121, + 0.017373522743582726, + 0.008162644691765308, + 0.012711375951766968, + 0.0074457949958741665, + 0.008279263973236084, + 0.010042012669146061, + 0.011647139675915241, + 0.012449977919459343, + 0.02179173193871975, + 0.0064487275667488575, + 0.03695611283183098, + 0.010401417501270771, + 0.009995659813284874, + 0.029134664684534073, + 0.017385147511959076, + 0.026149537414312363, + 0.012480338104069233, + 0.008722921833395958, + 0.017715156078338623, + 0.013348369859158993, + 0.009635970927774906, + 0.04575202241539955, + 0.009085627272725105, + 0.031037913635373116, + 0.004750469699501991, + 0.002782627008855343, + 0.00393156660720706, + 0.004370672162622213, + 0.0031405275221914053, + 0.014280402101576328, + 0.006942691281437874, + 0.0146645437926054, + 0.01382534485310316, + 0.009315380826592445, + 0.014025400392711163, + 0.00929241068661213, + 0.01569187454879284, + 0.01903388276696205, + 0.008940870873630047, + 0.018034925684332848, + 0.010524127632379532, + 0.007743343245238066, + 0.006670566275715828, + 0.007001935504376888, + 0.007597226183861494, + 0.01159853395074606, + 0.006110108457505703, + 0.016393790021538734, + 0.00808167178183794, + 0.00506390817463398, + 0.006734535098075867, + 0.006555402651429176, + 0.005906037986278534, + 0.00814015045762062, + 0.008526689372956753, + 0.00918364617973566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.3540556728839874 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043125398457050323 + ], + "opposite_logprob_tokens": [ + -3.1681253910064697 + ], + "logprob_reward": -0.043125398457050323, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043125398457050323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23447030782699585, + 0.07972277700901031, + 0.03638739138841629, + 0.07228781282901764, + 0.17809249460697174, + 0.1552353948354721, + 0.0842452198266983, + 0.07212803512811661, + 0.043720345944166183, + 0.043710269033908844 + ], + "credit_map": [ + 0.029832931235432625, + 0.027464609593153, + 0.025506427511572838, + 0.03574618697166443, + 0.026434216648340225, + 0.03544774278998375, + 0.02817126363515854, + 0.025866925716400146, + 0.01065726950764656, + 0.007808252703398466, + 0.013564650900661945, + 0.008534286171197891, + 0.011227752082049847, + 0.011153566651046276, + 0.008758584968745708, + 0.00801840890198946, + 0.004430421628057957, + 0.005002992227673531, + 0.006180572789162397, + 0.0031582543160766363, + 0.004775327630341053, + 0.0052685001865029335, + 0.0029285333584994078, + 0.004642791580408812, + 0.01510623563081026, + 0.005453568883240223, + 0.00878754910081625, + 0.0069725471548736095, + 0.006695258431136608, + 0.006650902796536684, + 0.011427163146436214, + 0.01119458768516779, + 0.030301762744784355, + 0.007428308017551899, + 0.04060215875506401, + 0.012750116176903248, + 0.009993024170398712, + 0.03086734749376774, + 0.01654246635735035, + 0.02960730716586113, + 0.007513857912272215, + 0.013017905876040459, + 0.014970981515944004, + 0.008343221619725227, + 0.007204272318631411, + 0.05304013192653656, + 0.009485087357461452, + 0.04165994003415108, + 0.008194530382752419, + 0.008004316128790379, + 0.0070197321474552155, + 0.011279106140136719, + 0.005994751583784819, + 0.014917111955583096, + 0.01402873732149601, + 0.014806926250457764, + 0.00888754054903984, + 0.007343118544667959, + 0.010598503984510899, + 0.005942944902926683, + 0.011559770442545414, + 0.010847878642380238, + 0.006916627753525972, + 0.010031646117568016, + 0.006130591034889221, + 0.004822984803467989, + 0.0043479083105921745, + 0.0040166075341403484, + 0.0043928371742367744, + 0.006831957492977381, + 0.004467776045203209, + 0.0087096793577075, + 0.004274547565728426, + 0.006277912762016058, + 0.004209021106362343, + 0.004096042364835739, + 0.0061837537214159966, + 0.004866833332926035, + 0.008177090436220169, + 0.005625065881758928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.043125398457050323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209638349711895 + ], + "opposite_logprob_tokens": [ + -4.264209747314453 + ], + "logprob_reward": -0.014209638349711895, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209638349711895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17636200785636902, + 0.05527639016509056, + 0.05063088983297348, + 0.05824627727270126, + 0.10097673535346985, + 0.3225286304950714, + 0.07551302760839462, + 0.05471319332718849, + 0.04291287437081337, + 0.06283999234437943 + ], + "credit_map": [ + 0.02302646078169346, + 0.018296457827091217, + 0.019033614546060562, + 0.025341186672449112, + 0.018238592892885208, + 0.028332138434052467, + 0.02402552217245102, + 0.02006802149116993, + 0.00693496922031045, + 0.0036935789976269007, + 0.011017736047506332, + 0.005969902500510216, + 0.005859619006514549, + 0.009676762856543064, + 0.006176341325044632, + 0.005947478581219912, + 0.0059294914826750755, + 0.005449189804494381, + 0.010762382298707962, + 0.004093752708286047, + 0.005843351129442453, + 0.0070671020075678825, + 0.005381559021770954, + 0.006104061380028725, + 0.011697648093104362, + 0.0050696125254035, + 0.008945085108280182, + 0.005550820380449295, + 0.006520811468362808, + 0.007114000152796507, + 0.006715336348861456, + 0.0066329617984592915, + 0.0132463239133358, + 0.006262687034904957, + 0.022925367578864098, + 0.008076881989836693, + 0.007156269159168005, + 0.017933063209056854, + 0.010358928702771664, + 0.015017220750451088, + 0.014178971759974957, + 0.027363799512386322, + 0.03346572071313858, + 0.01694835163652897, + 0.017238067463040352, + 0.1134318858385086, + 0.019521063193678856, + 0.08038076758384705, + 0.01106911338865757, + 0.005335339345037937, + 0.00765988091006875, + 0.009149142540991306, + 0.005643811076879501, + 0.011394620873034, + 0.012240373529493809, + 0.013020744547247887, + 0.010210704989731312, + 0.005488986149430275, + 0.009946719743311405, + 0.004120287485420704, + 0.007973438128829002, + 0.00632869079709053, + 0.004179783631116152, + 0.006464580073952675, + 0.0062922462821006775, + 0.005817747209221125, + 0.0037188830319792032, + 0.004762530792504549, + 0.003993890713900328, + 0.0061772894114255905, + 0.0038095726631581783, + 0.008340715430676937, + 0.0069422186352312565, + 0.00724327377974987, + 0.006081725005060434, + 0.006550915073603392, + 0.005253008101135492, + 0.009674218483269215, + 0.006915344391018152, + 0.01417929120361805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.014209638349711895 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560177743434906 + ], + "opposite_logprob_tokens": [ + -1.6006017923355103 + ], + "logprob_reward": -0.22560177743434906, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560177743434906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14525683224201202, + 0.11762459576129913, + 0.1687321960926056, + 0.12251021713018417, + 0.09497977793216705, + 0.05509362742304802, + 0.06829604506492615, + 0.07372358441352844, + 0.1077839806675911, + 0.04599916562438011 + ], + "credit_map": [ + 0.02003565989434719, + 0.018658334389328957, + 0.01564536802470684, + 0.024349287152290344, + 0.013419496826827526, + 0.025485849007964134, + 0.012538313865661621, + 0.015124522149562836, + 0.01737242192029953, + 0.01455436460673809, + 0.019226467236876488, + 0.010443737730383873, + 0.010623651556670666, + 0.016576282680034637, + 0.015202326700091362, + 0.013625340536236763, + 0.032641176134347916, + 0.016592519357800484, + 0.03935275971889496, + 0.012525741942226887, + 0.016023172065615654, + 0.02447337843477726, + 0.011150272563099861, + 0.015973180532455444, + 0.026245126500725746, + 0.009906831197440624, + 0.02160950005054474, + 0.009572482667863369, + 0.01025079470127821, + 0.012009522877633572, + 0.017463786527514458, + 0.01545217540115118, + 0.02459547109901905, + 0.003545232117176056, + 0.022158406674861908, + 0.004034548997879028, + 0.0029352065175771713, + 0.017289962619543076, + 0.004936265293508768, + 0.015484682284295559, + 0.006302940659224987, + 0.004457495640963316, + 0.007177804131060839, + 0.004023998975753784, + 0.00485867029055953, + 0.013774838298559189, + 0.004102255683392286, + 0.010395626537501812, + 0.003366067074239254, + 0.004600179847329855, + 0.004222340416163206, + 0.006268330384045839, + 0.0037003387697041035, + 0.018428517505526543, + 0.009252475574612617, + 0.018457796424627304, + 0.010426283814013004, + 0.007558350916951895, + 0.012026418931782246, + 0.0076555716805160046, + 0.009013906121253967, + 0.010186360217630863, + 0.007258592173457146, + 0.00959810521453619, + 0.011808840557932854, + 0.013184388168156147, + 0.010562182404100895, + 0.014546944759786129, + 0.011746715754270554, + 0.01646270975470543, + 0.010930567979812622, + 0.01854163408279419, + 0.004997981246560812, + 0.004909760318696499, + 0.0032851616851985455, + 0.0036415625363588333, + 0.005799759179353714, + 0.006503525655716658, + 0.007909970358014107, + 0.008951445110142231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.22560177743434906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201468512415886 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201468512415886, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201468512415886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29641619324684143, + 0.03963151201605797, + 0.09629598259925842, + 0.06059705838561058, + 0.04473155364394188, + 0.06693560630083084, + 0.03531622141599655, + 0.20455317199230194, + 0.08719062060117722, + 0.06833209097385406 + ], + "credit_map": [ + 0.03587552160024643, + 0.04083722457289696, + 0.03376753628253937, + 0.04700379818677902, + 0.03050108440220356, + 0.04385199397802353, + 0.032403744757175446, + 0.03217530995607376, + 0.00591680034995079, + 0.005022523459047079, + 0.005114986561238766, + 0.004253581631928682, + 0.004030456766486168, + 0.0055174631997942924, + 0.005433994811028242, + 0.004341708961874247, + 0.013429129496216774, + 0.010516314767301083, + 0.02162470482289791, + 0.00991768203675747, + 0.009843026287853718, + 0.015228426083922386, + 0.00612286664545536, + 0.009613830596208572, + 0.010118488222360611, + 0.005213931668549776, + 0.008859380148351192, + 0.006815421860665083, + 0.0068658702075481415, + 0.005890404339879751, + 0.009423287585377693, + 0.007410275284200907, + 0.006859119050204754, + 0.0024734409525990486, + 0.010189452208578587, + 0.0028006930369883776, + 0.003925341181457043, + 0.00801515206694603, + 0.004248303361237049, + 0.0062200515531003475, + 0.00986552331596613, + 0.006175765302032232, + 0.011833920143544674, + 0.007457081228494644, + 0.006595544051378965, + 0.012728986330330372, + 0.005797469988465309, + 0.0064813196659088135, + 0.0044383397325873375, + 0.0035730458330363035, + 0.0044032069854438305, + 0.0038820586632937193, + 0.003981325309723616, + 0.0050484552048146725, + 0.005386145785450935, + 0.004603642504662275, + 0.0270095057785511, + 0.020826872438192368, + 0.027050171047449112, + 0.019314533099532127, + 0.03201669454574585, + 0.02978997863829136, + 0.018096482381224632, + 0.030448924750089645, + 0.011590201407670975, + 0.00861219223588705, + 0.00911659561097622, + 0.007866853848099709, + 0.012546509504318237, + 0.013123970478773117, + 0.007929727435112, + 0.016404571011662483, + 0.008305136114358902, + 0.007406373508274555, + 0.006502658128738403, + 0.006292520556598902, + 0.009199597872793674, + 0.008841372095048428, + 0.010380162857472897, + 0.011404273100197315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.018201468512415886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, uninterrupted, dynamic tracking shot locks onto and circles a weathered wooden easel spinning rapidly in mid-air. The old easel's surface is covered in scratches, dents, and the fine cracks of age. Layers of thick, dried-up paint, from deep Prussian blue and fiery cadmium red to brilliant lemon yellow and pure titanium white, some of which have cracked, creating a mottled texture. Amidst a pure black, boundless void, the easel undergoes a wild, erratic rotation about its central axis. With each violent tumble and twist, the intense centrifugal force hurls the still-dry, viscous paint, creating dazzling streaks of color and splashes of liquid against the inky black background. These color trails sometimes resemble the slender, brilliant tails of a comet; sometimes, like exploding fireworks, they burst forth into countless tiny specks of color. The camera follows steadily and persistently, capturing every moment of color splashing. The entire picture is full of uncontrolled dynamic beauty and pure creative energy, like a silent but intense color symphony.", + "seed": 31509, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_5/index_1152_seed_31509_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2452089235186577, + "vlm_logprob": -0.2452089235186577, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.2452089235186577, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7008511424064636 + ], + "opposite_logprob_tokens": [ + -0.7008511424064636 + ], + "logprob_reward": -0.7008511424064636, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7008511424064636, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1404174566268921, + 0.037322476506233215, + 0.05225589498877525, + 0.07687417417764664, + 0.10026947408914566, + 0.1486232727766037, + 0.22859863936901093, + 0.11237500607967377, + 0.06065306067466736, + 0.04261058568954468 + ], + "credit_map": [ + 0.02289848029613495, + 0.009374205954372883, + 0.02859118953347206, + 0.011647071689367294, + 0.014950512908399105, + 0.01747414469718933, + 0.018482500687241554, + 0.016999343410134315, + 0.004807986784726381, + 0.0026925639249384403, + 0.006490130443125963, + 0.0034278908278793097, + 0.003151959041133523, + 0.005808664485812187, + 0.004958012141287327, + 0.005985266529023647, + 0.0078044175170362, + 0.004433090798556805, + 0.010422002524137497, + 0.0038413868751376867, + 0.0055830213241279125, + 0.007245476357638836, + 0.006450648419559002, + 0.006475852336734533, + 0.008464447222650051, + 0.0056502074003219604, + 0.012664415873587132, + 0.005223273299634457, + 0.010474221780896187, + 0.012368295341730118, + 0.012913791462779045, + 0.009115510620176792, + 0.00982758030295372, + 0.007307449355721474, + 0.01856006681919098, + 0.007534720003604889, + 0.00648388359695673, + 0.02265150286257267, + 0.006835190113633871, + 0.021069074049592018, + 0.016786638647317886, + 0.012789115309715271, + 0.0456923171877861, + 0.009590432047843933, + 0.010087384842336178, + 0.02737732231616974, + 0.009054834954440594, + 0.017245210707187653, + 0.025512589141726494, + 0.024724261835217476, + 0.03170764446258545, + 0.026494204998016357, + 0.019504999741911888, + 0.0499659888446331, + 0.02175898477435112, + 0.028929954394698143, + 0.015620999969542027, + 0.011608959175646305, + 0.013924377970397472, + 0.014228117652237415, + 0.008866066113114357, + 0.01685246080160141, + 0.00872885249555111, + 0.022545168176293373, + 0.008793973363935947, + 0.007177100516855717, + 0.00928642600774765, + 0.005915082525461912, + 0.006614220328629017, + 0.008892162702977657, + 0.00483910646289587, + 0.009134980849921703, + 0.0066300141625106335, + 0.00247614155523479, + 0.010435295291244984, + 0.0036074775271117687, + 0.002517196349799633, + 0.006397834047675133, + 0.0037827661726623774, + 0.006763855926692486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.7008511424064636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04312722757458687 + ], + "opposite_logprob_tokens": [ + -3.1681272983551025 + ], + "logprob_reward": -0.04312722757458687, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04312722757458687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11695865541696548, + 0.060255665332078934, + 0.046307530254125595, + 0.05470817908644676, + 0.14294373989105225, + 0.10373520851135254, + 0.1568574607372284, + 0.17640233039855957, + 0.04034052789211273, + 0.10149073600769043 + ], + "credit_map": [ + 0.014625923708081245, + 0.0076406244188547134, + 0.020773958414793015, + 0.007647397927939892, + 0.010462075471878052, + 0.022505613043904305, + 0.014204699546098709, + 0.019098367542028427, + 0.009198338724672794, + 0.004573107231408358, + 0.011797142215073109, + 0.005224297288805246, + 0.005123127717524767, + 0.006861128378659487, + 0.008699668571352959, + 0.008778853341937065, + 0.0061363657005131245, + 0.0026132308412343264, + 0.008225113153457642, + 0.002863375935703516, + 0.0030679851770401, + 0.009915962815284729, + 0.004396893084049225, + 0.009088604710996151, + 0.006728641223162413, + 0.004427785519510508, + 0.01104630809277296, + 0.003758018370717764, + 0.007084144279360771, + 0.008953351527452469, + 0.004359004087746143, + 0.008350921794772148, + 0.006995169911533594, + 0.007930099032819271, + 0.007261385675519705, + 0.008185744285583496, + 0.008506079204380512, + 0.04987062141299248, + 0.007939382456243038, + 0.046255264431238174, + 0.011143566109240055, + 0.006777665577828884, + 0.026339223608374596, + 0.011140314862132072, + 0.007846779190003872, + 0.02080090530216694, + 0.007346214726567268, + 0.012340545654296875, + 0.02759452350437641, + 0.015745827928185463, + 0.020764008164405823, + 0.015905508771538734, + 0.017943410202860832, + 0.01926358975470066, + 0.021467413753271103, + 0.01817317306995392, + 0.03167286142706871, + 0.012553421780467033, + 0.03586706891655922, + 0.013201921246945858, + 0.012819508090615273, + 0.029567869380116463, + 0.010378326289355755, + 0.030341356992721558, + 0.006047538015991449, + 0.004853263031691313, + 0.0063592069782316685, + 0.00403601722791791, + 0.0034002861939370632, + 0.00625738175585866, + 0.002867786679416895, + 0.006519048009067774, + 0.014317813329398632, + 0.004357683472335339, + 0.024343984201550484, + 0.00906852725893259, + 0.005576835013926029, + 0.015023550018668175, + 0.014109720475971699, + 0.01469262782484293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.04312722757458687 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023311948403716087 + ], + "opposite_logprob_tokens": [ + -3.7733118534088135 + ], + "logprob_reward": -0.023311948403716087, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023311948403716087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12488299608230591, + 0.07806333899497986, + 0.05595797300338745, + 0.04482153058052063, + 0.0644017681479454, + 0.09634548425674438, + 0.2735872268676758, + 0.18538092076778412, + 0.030968626961112022, + 0.04559013620018959 + ], + "credit_map": [ + 0.017595861107110977, + 0.006423198152333498, + 0.027880748733878136, + 0.015608031302690506, + 0.009984693489968777, + 0.016485707834362984, + 0.016256239265203476, + 0.014648513868451118, + 0.014123226515948772, + 0.005709618330001831, + 0.015153604559600353, + 0.008145310916006565, + 0.005486206151545048, + 0.011520659551024437, + 0.009040242061018944, + 0.008884469978511333, + 0.00435892166569829, + 0.002625122433528304, + 0.0044401888735592365, + 0.002727617509663105, + 0.003399056615307927, + 0.018129801377654076, + 0.0032610278576612473, + 0.01701623573899269, + 0.004970187786966562, + 0.003621326293796301, + 0.00815318152308464, + 0.003834253177046776, + 0.005451798439025879, + 0.007050555665045977, + 0.005909857805818319, + 0.005830369424074888, + 0.008674226701259613, + 0.004608425311744213, + 0.010371909476816654, + 0.006506277713924646, + 0.008252465166151524, + 0.009044370613992214, + 0.008039254695177078, + 0.008904841728508472, + 0.010798148810863495, + 0.00541742192581296, + 0.022042125463485718, + 0.008121966384351254, + 0.005840221885591745, + 0.014866520650684834, + 0.012011738494038582, + 0.017247341573238373, + 0.03306981548666954, + 0.030583227053284645, + 0.033045005053281784, + 0.035503920167684555, + 0.02624370902776718, + 0.04987146705389023, + 0.029776055365800858, + 0.03549404814839363, + 0.021938001736998558, + 0.016690723598003387, + 0.03398594260215759, + 0.023872213438153267, + 0.018871458247303963, + 0.030131785199046135, + 0.014338579960167408, + 0.025552228093147278, + 0.004456410650163889, + 0.0022831466048955917, + 0.005473082885146141, + 0.0038686103653162718, + 0.003443371970206499, + 0.004236652981489897, + 0.0033585920464247465, + 0.0038487580604851246, + 0.006701902486383915, + 0.0019490118138492107, + 0.009162644855678082, + 0.004232983570545912, + 0.002656339667737484, + 0.006765337660908699, + 0.006274002604186535, + 0.007847914472222328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.023311948403716087 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42892658710479736 + ], + "opposite_logprob_tokens": [ + -1.0539265871047974 + ], + "logprob_reward": -0.42892658710479736, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42892658710479736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21698062121868134, + 0.07888833433389664, + 0.09100288152694702, + 0.034341584891080856, + 0.039511483162641525, + 0.06164630129933357, + 0.09944520890712738, + 0.12298926711082458, + 0.05885700136423111, + 0.19633738696575165 + ], + "credit_map": [ + 0.03991006314754486, + 0.012302243150770664, + 0.03767581284046173, + 0.025298412889242172, + 0.015691561624407768, + 0.03196721896529198, + 0.02501143515110016, + 0.02912387251853943, + 0.014783349819481373, + 0.00472595589235425, + 0.018047265708446503, + 0.005914592649787664, + 0.003961999900639057, + 0.012223852798342705, + 0.00821604672819376, + 0.01101526990532875, + 0.01865786500275135, + 0.004304396454244852, + 0.028304878622293472, + 0.0037876951973885298, + 0.004339201375842094, + 0.014944728463888168, + 0.004418103955686092, + 0.01224600337445736, + 0.005680440459400415, + 0.0032088819425553083, + 0.006640298292040825, + 0.0028643286786973476, + 0.004048733040690422, + 0.004415810573846102, + 0.003370462916791439, + 0.004112632013857365, + 0.0036465530283749104, + 0.0027907169423997402, + 0.0035828419495373964, + 0.0038553865160793066, + 0.004597691353410482, + 0.008747363463044167, + 0.004346449393779039, + 0.007944480516016483, + 0.009272160939872265, + 0.004527499433606863, + 0.014428559690713882, + 0.0056038955226540565, + 0.005092439241707325, + 0.011169306933879852, + 0.004894288722425699, + 0.006658143829554319, + 0.02080799825489521, + 0.009929210878908634, + 0.01711519993841648, + 0.008323773741722107, + 0.007073205430060625, + 0.017068007960915565, + 0.0079736253246665, + 0.01115418877452612, + 0.023056181147694588, + 0.004892092663794756, + 0.027903994545340538, + 0.0063042896799743176, + 0.006087302230298519, + 0.021382175385951996, + 0.006492826621979475, + 0.026870407164096832, + 0.013195177540183067, + 0.0028893263079226017, + 0.015004729852080345, + 0.005802018567919731, + 0.002947817323729396, + 0.009341336786746979, + 0.004154487047344446, + 0.005522110965102911, + 0.03051995486021042, + 0.006513625383377075, + 0.056751150637865067, + 0.018083356320858, + 0.008070685900747776, + 0.0332927480340004, + 0.014481117017567158, + 0.028624732047319412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.42892658710479736 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a weathered wooden easel visible spinning rapidly in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the easel surface show scratches, dents, and fine cracks of age?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are layers of thick, dried-up paint in Prussian blue, cadmium red, lemon yellow, and titanium white visible on the easel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a pure black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow steadily and persistently while the easel rotates and paint splashes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982771210372448 + ], + "opposite_logprob_tokens": [ + -3.52982759475708 + ], + "logprob_reward": -0.02982771210372448, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982771210372448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18166005611419678, + 0.05807989090681076, + 0.12306097149848938, + 0.08959256112575531, + 0.11279995739459991, + 0.03912537172436714, + 0.1870965212583542, + 0.11996553838253021, + 0.027495311573147774, + 0.0611238107085228 + ], + "credit_map": [ + 0.03363057225942612, + 0.01217438280582428, + 0.043452825397253036, + 0.014064011164009571, + 0.01385561004281044, + 0.02513529546558857, + 0.01726125366985798, + 0.02208610437810421, + 0.011728933081030846, + 0.003142519621178508, + 0.011374779976904392, + 0.0032865910325199366, + 0.0031867469660937786, + 0.010644492693245411, + 0.0050219581462442875, + 0.009693869389593601, + 0.018264099955558777, + 0.005974146071821451, + 0.027836646884679794, + 0.005926771089434624, + 0.006349779665470123, + 0.026916513219475746, + 0.006969281937927008, + 0.02482372522354126, + 0.012663775123655796, + 0.006392267532646656, + 0.01727108284831047, + 0.00733800046145916, + 0.01068048644810915, + 0.012594957835972309, + 0.013566432520747185, + 0.009085557423532009, + 0.011080549098551273, + 0.008557616733014584, + 0.017191795632243156, + 0.010770649649202824, + 0.008138415403664112, + 0.024773722514510155, + 0.009800201281905174, + 0.022487010806798935, + 0.007571298163384199, + 0.002917750971391797, + 0.005227471236139536, + 0.002583540743216872, + 0.0033014933578670025, + 0.006445369683206081, + 0.0031681207474321127, + 0.007910327985882759, + 0.03870382905006409, + 0.01847309060394764, + 0.03193170577287674, + 0.012249861843883991, + 0.013911250978708267, + 0.03286850452423096, + 0.01553051546216011, + 0.023427754640579224, + 0.012467648833990097, + 0.011914587579667568, + 0.020953424274921417, + 0.01614438183605671, + 0.011004890315234661, + 0.021877137944102287, + 0.008243493735790253, + 0.017359977588057518, + 0.004237906541675329, + 0.0021966039203107357, + 0.005596125964075327, + 0.003986953757703304, + 0.0021281898953020573, + 0.003841187572106719, + 0.0027228884864598513, + 0.0027854545041918755, + 0.00833031628280878, + 0.0035568217281252146, + 0.013585604727268219, + 0.006188937928527594, + 0.003639309201389551, + 0.009700573049485683, + 0.005704832263290882, + 0.01041741669178009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.02982771210372448 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_6/index_tensor([6731, 6731])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_6/index_tensor([6731, 6731])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a6de82e42e91b1de9f37dcaa5915b3c675b60947 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_6/index_tensor([6731, 6731])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium close-up shot, taken from a level angle, captures the silhouette of a young soldier, his head wearily resting against the cold window of an old train. He wears a well-worn, dark green woolen uniform with a stiff collar. His short, black hair, slightly damp with sweat, clung to his forehead. Crystal-clear tears silently fall from his bloodshot eyes, streaking wet trails along his dusty, young cheeks. His jaw is set, as if he's straining to suppress an impending collapse. The carriage sways gently and rhythmically with the rhythm of the rails, the dark wooden window frames and interior gleaming with a worn patina in the dim light. Outside, the distant mountains and interlaced fields are stretched and blurred by the rapid movement, completely losing their concrete form and vibrant colors, transforming into a dynamic, backward-flowing scroll of varying shades of gray, dark green, and ochre. The soft, cloudy diffuse light pouring in from the window sets a tranquil and sad cold tone for the entire scene. The picture is full of silent tension, which contrasts sharply with the soldier's heavy sorrow and the poetic and illusory scenery outside the window.", + "seed": 10061, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_6/index_6731_seed_10061_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.033731281757354736, + "vlm_logprob": -0.033731281757354736, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.033731281757354736, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05164388567209244 + ], + "opposite_logprob_tokens": [ + -4.8016438484191895 + ], + "logprob_reward": -0.05164388567209244, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.05164388567209244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13671012222766876, + 0.18933863937854767, + 0.09229352325201035, + 0.04153585061430931, + 0.06927059590816498, + 0.07336067408323288, + 0.18697497248649597, + 0.07770086824893951, + 0.05603645369410515, + 0.07677827775478363 + ], + "credit_map": [ + 0.020901065319776535, + 0.015944860875606537, + 0.024722397327423096, + 0.01718531921505928, + 0.012175173498690128, + 0.015610238537192345, + 0.009223508648574352, + 0.020947566255927086, + 0.021019235253334045, + 0.019982455298304558, + 0.05069037526845932, + 0.029971275478601456, + 0.008050117641687393, + 0.036272939294576645, + 0.009134790860116482, + 0.014217459596693516, + 0.012242668308317661, + 0.008571537211537361, + 0.023117607459425926, + 0.012877080589532852, + 0.006224594544619322, + 0.015851009637117386, + 0.004534903913736343, + 0.008874121122062206, + 0.004817324224859476, + 0.007302459795027971, + 0.007051211781799793, + 0.006064970511943102, + 0.003365421202033758, + 0.003966725897043943, + 0.004974129609763622, + 0.00399360666051507, + 0.007612698711454868, + 0.008841759525239468, + 0.01628970354795456, + 0.013521494343876839, + 0.005010977853089571, + 0.007428452372550964, + 0.0037161996588110924, + 0.006849312223494053, + 0.006955633405596018, + 0.007627136539667845, + 0.020613839849829674, + 0.012249047867953777, + 0.003745228750631213, + 0.013945982791483402, + 0.002855468075722456, + 0.005368338432163, + 0.019300321117043495, + 0.019428500905632973, + 0.05271800234913826, + 0.02884078584611416, + 0.009366055019199848, + 0.029134275391697884, + 0.01196111086755991, + 0.01622593402862549, + 0.010602974332869053, + 0.009986876510083675, + 0.013760626316070557, + 0.010858145542442799, + 0.005992384627461433, + 0.008385414257645607, + 0.006748899817466736, + 0.01136554591357708, + 0.007349098101258278, + 0.0071288407780230045, + 0.011254015378654003, + 0.009174005128443241, + 0.0059418994933366776, + 0.006161279045045376, + 0.0041732327081263065, + 0.004854083526879549, + 0.014032610692083836, + 0.01065007969737053, + 0.01105915941298008, + 0.011918800882995129, + 0.007986093871295452, + 0.005322783254086971, + 0.007176029961556196, + 0.008632717654109001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.05164388567209244 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043083496391773224 + ], + "opposite_logprob_tokens": [ + -3.168083429336548 + ], + "logprob_reward": -0.043083496391773224, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043083496391773224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1025579422712326, + 0.06376729160547256, + 0.11268189549446106, + 0.07210482656955719, + 0.12453407794237137, + 0.03363265097141266, + 0.08648301661014557, + 0.2362159788608551, + 0.0803840234875679, + 0.08763832598924637 + ], + "credit_map": [ + 0.014379367232322693, + 0.012176535092294216, + 0.01810722053050995, + 0.01484366599470377, + 0.009341156110167503, + 0.009394227527081966, + 0.013225024566054344, + 0.011090738698840141, + 0.005833066999912262, + 0.010000544600188732, + 0.011866094544529915, + 0.012162850238382816, + 0.004200918599963188, + 0.0071070496924221516, + 0.0061415936797857285, + 0.006455168593674898, + 0.010544500313699245, + 0.011142324656248093, + 0.02626948617398739, + 0.015233629383146763, + 0.009061302989721298, + 0.016468079760670662, + 0.015496787615120411, + 0.008465779945254326, + 0.007375374436378479, + 0.0076357475481927395, + 0.01859036274254322, + 0.01016538217663765, + 0.003675284795463085, + 0.014134875498712063, + 0.004711271729320288, + 0.005816522985696793, + 0.015609844587743282, + 0.015331660397350788, + 0.026313122361898422, + 0.023460909724235535, + 0.010891408659517765, + 0.010848825797438622, + 0.0078088692389428616, + 0.01426943950355053, + 0.0031081310007721186, + 0.0031593842431902885, + 0.00853276252746582, + 0.005606681574136019, + 0.0021482266020029783, + 0.0057122232392430305, + 0.0017607848858460784, + 0.003604454919695854, + 0.009322868660092354, + 0.007568535394966602, + 0.021218648180365562, + 0.012496920302510262, + 0.006022660061717033, + 0.012366598471999168, + 0.007288675289601088, + 0.01019811350852251, + 0.02825908735394478, + 0.02104976214468479, + 0.05811113491654396, + 0.038433756679296494, + 0.021068111062049866, + 0.030529119074344635, + 0.018674615770578384, + 0.020090389996767044, + 0.008790909312665462, + 0.008132213726639748, + 0.020440077409148216, + 0.012216801755130291, + 0.004699672572314739, + 0.01230469811707735, + 0.003917543683201075, + 0.00988211389631033, + 0.01368015818297863, + 0.01317654363811016, + 0.018015969544649124, + 0.012840026058256626, + 0.007196268532425165, + 0.008567620068788528, + 0.005515726748853922, + 0.008646012283861637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.043083496391773224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290403187274933 + ], + "opposite_logprob_tokens": [ + -3.7732903957366943 + ], + "logprob_reward": -0.023290403187274933, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290403187274933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10016579926013947, + 0.11506128311157227, + 0.08033743500709534, + 0.059043049812316895, + 0.04745391756296158, + 0.08741895854473114, + 0.1605713814496994, + 0.22321800887584686, + 0.039024922996759415, + 0.08770529180765152 + ], + "credit_map": [ + 0.009251248091459274, + 0.01103564165532589, + 0.01586981676518917, + 0.013355087488889694, + 0.007605991326272488, + 0.012533200904726982, + 0.011125066317617893, + 0.019389737397432327, + 0.013974362052977085, + 0.01844996213912964, + 0.020194992423057556, + 0.017340825870633125, + 0.008958946913480759, + 0.01111934706568718, + 0.014030447229743004, + 0.010992401279509068, + 0.007344541139900684, + 0.010463347658514977, + 0.011559304781258106, + 0.010005095042288303, + 0.008856629952788353, + 0.01017884723842144, + 0.011438965797424316, + 0.010490703396499157, + 0.005410952027887106, + 0.005367210600525141, + 0.01665513589978218, + 0.00864782091230154, + 0.0028256827499717474, + 0.011387471109628677, + 0.004300421569496393, + 0.004448353312909603, + 0.00585585180670023, + 0.004944707732647657, + 0.012857720255851746, + 0.007424847688525915, + 0.003153787925839424, + 0.00680272001773119, + 0.0025304381269961596, + 0.003883844241499901, + 0.007165601011365652, + 0.008684772998094559, + 0.025247590616345406, + 0.015210880897939205, + 0.005534177180379629, + 0.017445987090468407, + 0.0035343372728675604, + 0.004595605190843344, + 0.01676572673022747, + 0.015455149114131927, + 0.045210517942905426, + 0.025223087519407272, + 0.008430561982095242, + 0.024807440117001534, + 0.011362539604306221, + 0.013316365890204906, + 0.028157100081443787, + 0.024488789960741997, + 0.04473094269633293, + 0.03805818408727646, + 0.024545693770051003, + 0.025905171409249306, + 0.01724996417760849, + 0.020082170143723488, + 0.004123293794691563, + 0.003274944145232439, + 0.010406817309558392, + 0.006350694224238396, + 0.002677682787179947, + 0.0059495833702385426, + 0.0022170040756464005, + 0.004024902358651161, + 0.009251520968973637, + 0.009881794452667236, + 0.02399919368326664, + 0.013372009620070457, + 0.004057527054101229, + 0.013729560188949108, + 0.004813756328076124, + 0.008599929511547089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.023290403187274933 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810315206646919 + ], + "opposite_logprob_tokens": [ + -3.2881031036376953 + ], + "logprob_reward": -0.03810315206646919, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810315206646919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08070360124111176, + 0.21120890974998474, + 0.08760417252779007, + 0.03779805451631546, + 0.021193692460656166, + 0.04826699197292328, + 0.1370166391134262, + 0.19864855706691742, + 0.08823716640472412, + 0.08932221680879593 + ], + "credit_map": [ + 0.011750356294214725, + 0.010530101135373116, + 0.014435213059186935, + 0.01107028964906931, + 0.005675171036273241, + 0.012982260435819626, + 0.004591207019984722, + 0.009669001214206219, + 0.019932854920625687, + 0.019976690411567688, + 0.057537991553545, + 0.03450751304626465, + 0.00909876823425293, + 0.045139458030462265, + 0.008310954086482525, + 0.01670469157397747, + 0.009625437669456005, + 0.008692985400557518, + 0.020898325368762016, + 0.013101885095238686, + 0.005213331896811724, + 0.018817072734236717, + 0.0034689106978476048, + 0.007786223199218512, + 0.007511675823479891, + 0.004943043924868107, + 0.005713988095521927, + 0.0051230029202997684, + 0.003506016917526722, + 0.003262369194999337, + 0.0034533729776740074, + 0.004284584894776344, + 0.00255944742821157, + 0.0027939616702497005, + 0.0042886738665401936, + 0.004272877238690853, + 0.001987505005672574, + 0.0016040557529777288, + 0.0015971304383128881, + 0.0020900394301861525, + 0.004221671726554632, + 0.006839638110250235, + 0.012505781836807728, + 0.008583023212850094, + 0.0031977633479982615, + 0.008297359570860863, + 0.0019195377826690674, + 0.0027022152207791805, + 0.015039023011922836, + 0.011555559933185577, + 0.03590363636612892, + 0.020421411842107773, + 0.009239544160664082, + 0.02004106529057026, + 0.011626272462308407, + 0.01319013349711895, + 0.027701379731297493, + 0.02124759741127491, + 0.03054414875805378, + 0.031872499734163284, + 0.02548396587371826, + 0.01852887123823166, + 0.018656659871339798, + 0.024613432586193085, + 0.010304360650479794, + 0.006498638540506363, + 0.025027846917510033, + 0.014283325523138046, + 0.004103014711290598, + 0.013341826386749744, + 0.004569736309349537, + 0.01010841690003872, + 0.009347296319901943, + 0.0075262621976435184, + 0.027613388374447823, + 0.013885858468711376, + 0.0036856813821941614, + 0.015765493735671043, + 0.003923093434423208, + 0.007575138937681913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.03810315206646919 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535471469163895 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535471469163895, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535471469163895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07825279980897903, + 0.04207523539662361, + 0.05017776042222977, + 0.0974193811416626, + 0.07900388538837433, + 0.10097161680459976, + 0.12667687237262726, + 0.2894955277442932, + 0.03317909315228462, + 0.10274779796600342 + ], + "credit_map": [ + 0.011739798821508884, + 0.01132080890238285, + 0.011394697241485119, + 0.010176110081374645, + 0.005792038515210152, + 0.00532744824886322, + 0.011252199299633503, + 0.011249702423810959, + 0.005488497205078602, + 0.006307639647275209, + 0.005947164725512266, + 0.006193037144839764, + 0.0037760422565042973, + 0.004448677878826857, + 0.0034182490780949593, + 0.006495926063507795, + 0.004819544963538647, + 0.0069152177311480045, + 0.008366521447896957, + 0.01103502232581377, + 0.0033501514699310064, + 0.00400787265971303, + 0.005505381152033806, + 0.006178052164614201, + 0.00813616905361414, + 0.008443808183073997, + 0.02325703576207161, + 0.01348945964127779, + 0.0065910411067306995, + 0.01966884359717369, + 0.007606879808008671, + 0.01022613886743784, + 0.011079439893364906, + 0.008892476558685303, + 0.018190238624811172, + 0.012531238608062267, + 0.005612938664853573, + 0.008663701824843884, + 0.005347531288862228, + 0.008686322718858719, + 0.00875130109488964, + 0.008929152972996235, + 0.029768189415335655, + 0.016665438190102577, + 0.005476230755448341, + 0.020635971799492836, + 0.00427768612280488, + 0.006467644590884447, + 0.01373188104480505, + 0.013977820053696632, + 0.03326538950204849, + 0.019442126154899597, + 0.006700725294649601, + 0.0181894488632679, + 0.009784797206521034, + 0.01158467773348093, + 0.030887549743056297, + 0.030917098745703697, + 0.07915420085191727, + 0.04866960272192955, + 0.02193087339401245, + 0.0420629158616066, + 0.017786145210266113, + 0.018087133765220642, + 0.00400942750275135, + 0.003642899915575981, + 0.007272468414157629, + 0.005370686296373606, + 0.002594543620944023, + 0.004390493500977755, + 0.002534417435526848, + 0.003364157397300005, + 0.0119323143735528, + 0.013231756165623665, + 0.01964676007628441, + 0.015776604413986206, + 0.007140733301639557, + 0.012679518200457096, + 0.007440493907779455, + 0.014899618923664093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.012535471469163895 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium close-up shot, taken from a level angle, captures the silhouette of a young soldier, his head wearily resting against the cold window of an old train. He wears a well-worn, dark green woolen uniform with a stiff collar. His short, black hair, slightly damp with sweat, clung to his forehead. Crystal-clear tears silently fall from his bloodshot eyes, streaking wet trails along his dusty, young cheeks. His jaw is set, as if he's straining to suppress an impending collapse. The carriage sways gently and rhythmically with the rhythm of the rails, the dark wooden window frames and interior gleaming with a worn patina in the dim light. Outside, the distant mountains and interlaced fields are stretched and blurred by the rapid movement, completely losing their concrete form and vibrant colors, transforming into a dynamic, backward-flowing scroll of varying shades of gray, dark green, and ochre. The soft, cloudy diffuse light pouring in from the window sets a tranquil and sad cold tone for the entire scene. The picture is full of silent tension, which contrasts sharply with the soldier's heavy sorrow and the poetic and illusory scenery outside the window.", + "seed": 10062, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_6/index_6731_seed_10062_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026080177817493676, + "vlm_logprob": -0.026080177817493676, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.026080177817493676, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034355323761701584 + ], + "opposite_logprob_tokens": [ + -4.534355163574219 + ], + "logprob_reward": -0.034355323761701584, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.034355323761701584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4847995340824127, + 0.05934934690594673, + 0.0524357408285141, + 0.06423739343881607, + 0.022085128352046013, + 0.032902609556913376, + 0.06775260716676712, + 0.04169212281703949, + 0.11585718393325806, + 0.058888304978609085 + ], + "credit_map": [ + 0.1080184206366539, + 0.03291056677699089, + 0.02540801838040352, + 0.01723143830895424, + 0.022260798141360283, + 0.0498337559401989, + 0.0619896724820137, + 0.16714686155319214, + 0.004697045776993036, + 0.005539163947105408, + 0.007729308679699898, + 0.005315246991813183, + 0.004202879499644041, + 0.006853028666228056, + 0.014631657861173153, + 0.01038101315498352, + 0.006680097430944443, + 0.005410079378634691, + 0.0064888992346823215, + 0.0067030987702310085, + 0.005347573664039373, + 0.00661463662981987, + 0.0075447396375238895, + 0.007646619342267513, + 0.009096897207200527, + 0.005768369417637587, + 0.006676694843918085, + 0.0039776768535375595, + 0.004923698492348194, + 0.0071275257505476475, + 0.012970561161637306, + 0.013695969246327877, + 0.0029666442424058914, + 0.0034753410145640373, + 0.0024875234812498093, + 0.0019409238593652844, + 0.0017555536469444633, + 0.0029140065889805555, + 0.0035418453626334667, + 0.003003290155902505, + 0.0032953708432614803, + 0.002647717483341694, + 0.0069236913695931435, + 0.003298634896054864, + 0.0015443590236827731, + 0.0050349426455795765, + 0.005397621542215347, + 0.004760270472615957, + 0.011232327669858932, + 0.00724799046292901, + 0.007336200214922428, + 0.0032817022874951363, + 0.004605486523360014, + 0.006702855695039034, + 0.012113950215280056, + 0.015232093632221222, + 0.00435927277430892, + 0.005588159896433353, + 0.003763751359656453, + 0.004301072563976049, + 0.003681507194414735, + 0.0034916505683213472, + 0.010953949764370918, + 0.005552759859710932, + 0.02027053013443947, + 0.0066065602004528046, + 0.00803440809249878, + 0.003797688288614154, + 0.003699738532304764, + 0.007159100379794836, + 0.022794773802161217, + 0.04349438473582268, + 0.0062194811180233955, + 0.008101324550807476, + 0.007074183318763971, + 0.00494246743619442, + 0.004834681749343872, + 0.0033453921787440777, + 0.013444809243083, + 0.010925961658358574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.034355323761701584 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307812824845314 + ], + "opposite_logprob_tokens": [ + -3.1680781841278076 + ], + "logprob_reward": -0.04307812824845314, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307812824845314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6062723398208618, + 0.06878650933504105, + 0.12621751427650452, + 0.01348193734884262, + 0.029549935832619667, + 0.016048086807131767, + 0.03253607079386711, + 0.027337584644556046, + 0.04795446991920471, + 0.03181561455130577 + ], + "credit_map": [ + 0.14057183265686035, + 0.03507880121469498, + 0.0336238332092762, + 0.017330970615148544, + 0.027544202283024788, + 0.06412090361118317, + 0.07507098466157913, + 0.21293078362941742, + 0.005279106087982655, + 0.008606621995568275, + 0.00479013891890645, + 0.0057185315527021885, + 0.0065229954198002815, + 0.005550977308303118, + 0.01991121657192707, + 0.012406924739480019, + 0.026693478226661682, + 0.00973968580365181, + 0.006175864487886429, + 0.0076804193668067455, + 0.0095473388209939, + 0.01296668779104948, + 0.01520520355552435, + 0.03820882365107536, + 0.0009342513512820005, + 0.0019054978620260954, + 0.0013655929360538721, + 0.0014587767655029893, + 0.0013700793497264385, + 0.0017395378090441227, + 0.0027107889764010906, + 0.001997413346543908, + 0.0031327656470239162, + 0.004319348838180304, + 0.003108167089521885, + 0.0024645368102937937, + 0.0026762301567941904, + 0.0027186311781406403, + 0.006765116471797228, + 0.004365141969174147, + 0.002208166755735874, + 0.001489815884269774, + 0.003145136870443821, + 0.0012697227066382766, + 0.0007552884053438902, + 0.002068550558760762, + 0.00257270410656929, + 0.0025387019850313663, + 0.005600448232144117, + 0.002875274047255516, + 0.004065896850079298, + 0.001687997835688293, + 0.001710066688247025, + 0.0026389211416244507, + 0.005338144022971392, + 0.008619320578873158, + 0.002824279712513089, + 0.0027326587587594986, + 0.0032317631412297487, + 0.003232161980122328, + 0.0023315188009291887, + 0.0025832431856542826, + 0.006689026951789856, + 0.0037129304837435484, + 0.006373237352818251, + 0.0035458793863654137, + 0.0025600637309253216, + 0.0027867048047482967, + 0.003064498770982027, + 0.003935304936021566, + 0.011958932504057884, + 0.013729850761592388, + 0.00517992815002799, + 0.0035829744301736355, + 0.004232976585626602, + 0.00234004738740623, + 0.0022381243761628866, + 0.004112145397812128, + 0.0058240387588739395, + 0.004305377136915922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04307812824845314 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370155766606331 + ], + "opposite_logprob_tokens": [ + -3.4087016582489014 + ], + "logprob_reward": -0.03370155766606331, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370155766606331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5829477906227112, + 0.05374941974878311, + 0.10301560908555984, + 0.030438384041190147, + 0.05911407619714737, + 0.02841244265437126, + 0.0433431938290596, + 0.019161362200975418, + 0.06217403709888458, + 0.017643820494413376 + ], + "credit_map": [ + 0.1340903490781784, + 0.03347674012184143, + 0.030579308047890663, + 0.016721688210964203, + 0.02810230851173401, + 0.06228889897465706, + 0.07354726642370224, + 0.20414112508296967, + 0.00755707873031497, + 0.006192173808813095, + 0.008020262233912945, + 0.003980366047471762, + 0.003959564492106438, + 0.003989367745816708, + 0.011769343167543411, + 0.008281256072223186, + 0.02109471708536148, + 0.009834437631070614, + 0.00494561530649662, + 0.006408536806702614, + 0.007502412423491478, + 0.01217695977538824, + 0.013445541262626648, + 0.027607379481196404, + 0.0020390681456774473, + 0.005562483798712492, + 0.0026812274008989334, + 0.0034764029551297426, + 0.003413294441998005, + 0.0024738716892898083, + 0.006636668927967548, + 0.004155363421887159, + 0.006220556795597076, + 0.008120724931359291, + 0.007931175641715527, + 0.005879588890820742, + 0.0053011043928563595, + 0.005523314233869314, + 0.01182936318218708, + 0.00830823928117752, + 0.0057647149078547955, + 0.002292722463607788, + 0.0038714653346687555, + 0.0018486485350877047, + 0.001268278923816979, + 0.002728514140471816, + 0.004006857052445412, + 0.006631238851696253, + 0.007342716678977013, + 0.0036457465030252934, + 0.00515561830252409, + 0.0026751381810754538, + 0.0022098864428699017, + 0.004166441969573498, + 0.006498939357697964, + 0.011648700572550297, + 0.0022963269148021936, + 0.0023091931361705065, + 0.002258258406072855, + 0.0020257937721908092, + 0.0018968521617352962, + 0.0016087896656244993, + 0.0035686898045241833, + 0.0031974553130567074, + 0.007346510887145996, + 0.004768662620335817, + 0.0037616633344441652, + 0.003873909590765834, + 0.005008852109313011, + 0.004633227828890085, + 0.017895963042974472, + 0.014885243028402328, + 0.002254262799397111, + 0.0021026604808866978, + 0.002175472443923354, + 0.001421330263838172, + 0.000998810981400311, + 0.002119668060913682, + 0.004118842538446188, + 0.0024527704808861017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.03370155766606331 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734057795256376 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006734057795256376, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734057795256376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48415306210517883, + 0.03900875896215439, + 0.044284798204898834, + 0.0795668289065361, + 0.11725945770740509, + 0.12219548970460892, + 0.0287734754383564, + 0.03293728455901146, + 0.03983874246478081, + 0.011982088908553123 + ], + "credit_map": [ + 0.10962202399969101, + 0.028069790452718735, + 0.028689628466963768, + 0.013495809398591518, + 0.021249236539006233, + 0.05117306113243103, + 0.06213754788041115, + 0.16971595585346222, + 0.006844489369541407, + 0.004585671238601208, + 0.006467124912887812, + 0.0030115616973489523, + 0.0023256989661604166, + 0.0031743249855935574, + 0.0065637924708426, + 0.006036100443452597, + 0.005378880072385073, + 0.0049447896890342236, + 0.003992362879216671, + 0.004534806124866009, + 0.0033282027579844, + 0.007490289863198996, + 0.006780131254345179, + 0.007835340686142445, + 0.010514055378735065, + 0.007320345379412174, + 0.008226236328482628, + 0.004680059850215912, + 0.006333448458462954, + 0.012052696198225021, + 0.013165152631700039, + 0.017274845391511917, + 0.012574520893394947, + 0.014049410820007324, + 0.012145302258431911, + 0.006722551770508289, + 0.01115143671631813, + 0.016127172857522964, + 0.026216596364974976, + 0.01827247627079487, + 0.023609576746821404, + 0.008594178594648838, + 0.018581386655569077, + 0.007573644630610943, + 0.004893919453024864, + 0.014975872822105885, + 0.017703276127576828, + 0.0262636486440897, + 0.0048627182841300964, + 0.0024379438254982233, + 0.004906637594103813, + 0.0018217546166852117, + 0.0017737805610522628, + 0.00250416761264205, + 0.004474941175431013, + 0.005991536192595959, + 0.0037195526529103518, + 0.003554190741851926, + 0.0031193657778203487, + 0.004305264446884394, + 0.0026286239735782146, + 0.0030650319531559944, + 0.008310060016810894, + 0.0042352015152573586, + 0.007024187594652176, + 0.0021991173271089792, + 0.0031590647995471954, + 0.0014203079044818878, + 0.0012838812544941902, + 0.0027196102309972048, + 0.007393889129161835, + 0.014638686552643776, + 0.0016639924142509699, + 0.0012688200222328305, + 0.0012081388849765062, + 0.0010416285367682576, + 0.0006917159771546721, + 0.0010549277067184448, + 0.0030701865907758474, + 0.001982679357752204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.006734057795256376 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531821615993977 + ], + "opposite_logprob_tokens": [ + -4.387531757354736 + ], + "logprob_reward": -0.012531821615993977, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531821615993977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3981575667858124, + 0.049595706164836884, + 0.07094816863536835, + 0.03285503014922142, + 0.05675381422042847, + 0.03488598018884659, + 0.20673657953739166, + 0.032081544399261475, + 0.05353742837905884, + 0.06444820016622543 + ], + "credit_map": [ + 0.08025190979242325, + 0.026385586708784103, + 0.03078615292906761, + 0.015655744820833206, + 0.019134143367409706, + 0.041939571499824524, + 0.05523879453539848, + 0.12876568734645844, + 0.005804179701954126, + 0.005894975736737251, + 0.010871369391679764, + 0.004512956831604242, + 0.003359735244885087, + 0.004418065771460533, + 0.007596615236252546, + 0.007137808948755264, + 0.016585303470492363, + 0.005819595418870449, + 0.00990634597837925, + 0.0036826988216489553, + 0.0032116558868438005, + 0.00763998506590724, + 0.007839287631213665, + 0.016263296827673912, + 0.0036411911714822054, + 0.0038761578034609556, + 0.0037783472798764706, + 0.002664671977981925, + 0.0021057743579149246, + 0.004575089085847139, + 0.006319565698504448, + 0.005894235335290432, + 0.006021438632160425, + 0.007552223280072212, + 0.004643850959837437, + 0.004527462180703878, + 0.0063418177887797356, + 0.007846617139875889, + 0.012028702534735203, + 0.007791697047650814, + 0.003250627778470516, + 0.004026493523269892, + 0.004771807696670294, + 0.0035987375304102898, + 0.0019173193722963333, + 0.0048845927231013775, + 0.006794807035475969, + 0.005641593132168055, + 0.02974572218954563, + 0.016544384881854057, + 0.03374756872653961, + 0.01474010944366455, + 0.015765590593218803, + 0.025629311800003052, + 0.04269219934940338, + 0.027871698141098022, + 0.005041587166488171, + 0.0029751844704151154, + 0.005475006997585297, + 0.00321229943074286, + 0.0021591740660369396, + 0.0034064108040183783, + 0.005237807985395193, + 0.00457407534122467, + 0.00770171731710434, + 0.0042473915964365005, + 0.003939894028007984, + 0.0028993331361562014, + 0.003413330065086484, + 0.003875388065353036, + 0.01108681783080101, + 0.01637355238199234, + 0.011557496152818203, + 0.007349270861595869, + 0.007752757985144854, + 0.004198420327156782, + 0.004292136523872614, + 0.007334258873015642, + 0.012322334572672844, + 0.009641523472964764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.012531821615993977 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium close-up shot, taken from a level angle, captures the silhouette of a young soldier, his head wearily resting against the cold window of an old train. He wears a well-worn, dark green woolen uniform with a stiff collar. His short, black hair, slightly damp with sweat, clung to his forehead. Crystal-clear tears silently fall from his bloodshot eyes, streaking wet trails along his dusty, young cheeks. His jaw is set, as if he's straining to suppress an impending collapse. The carriage sways gently and rhythmically with the rhythm of the rails, the dark wooden window frames and interior gleaming with a worn patina in the dim light. Outside, the distant mountains and interlaced fields are stretched and blurred by the rapid movement, completely losing their concrete form and vibrant colors, transforming into a dynamic, backward-flowing scroll of varying shades of gray, dark green, and ochre. The soft, cloudy diffuse light pouring in from the window sets a tranquil and sad cold tone for the entire scene. The picture is full of silent tension, which contrasts sharply with the soldier's heavy sorrow and the poetic and illusory scenery outside the window.", + "seed": 10063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_6/index_6731_seed_10063_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03974105911329388, + "vlm_logprob": -0.03974105911329388, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.03974105911329388, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030930591747164726 + ], + "opposite_logprob_tokens": [ + -4.905930519104004 + ], + "logprob_reward": -0.030930591747164726, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.030930591747164726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10060550272464752, + 0.10237406939268112, + 0.02606899105012417, + 0.02505054883658886, + 0.03907875716686249, + 0.4135550856590271, + 0.07116162031888962, + 0.07201482355594635, + 0.12531360983848572, + 0.024776993319392204 + ], + "credit_map": [ + 0.015727344900369644, + 0.00817773025482893, + 0.01492021232843399, + 0.010308532044291496, + 0.01051272451877594, + 0.01486020814627409, + 0.01523903850466013, + 0.010859709233045578, + 0.006568087264895439, + 0.007474290672689676, + 0.013103715144097805, + 0.010058939456939697, + 0.01736346445977688, + 0.010838772170245647, + 0.024723712354898453, + 0.012243086472153664, + 0.002440772717818618, + 0.0023657402489334345, + 0.004227059427648783, + 0.002967499429360032, + 0.003341983538120985, + 0.003694816492497921, + 0.0045500705018639565, + 0.002481050556525588, + 0.002265943679958582, + 0.0024634075816720724, + 0.0046969265677034855, + 0.003054338740184903, + 0.004550546407699585, + 0.002395675517618656, + 0.00318197556771338, + 0.0024417368695139885, + 0.003077623900026083, + 0.003715134458616376, + 0.005535946227610111, + 0.004311356693506241, + 0.005187590606510639, + 0.004369743634015322, + 0.007979639805853367, + 0.004901721142232418, + 0.05772097408771515, + 0.014600874856114388, + 0.0391724668443203, + 0.01894805207848549, + 0.02802351489663124, + 0.047409508377313614, + 0.04064342752099037, + 0.167036235332489, + 0.004921363666653633, + 0.004944658372551203, + 0.00545069295912981, + 0.008697355166077614, + 0.010792401619255543, + 0.0061452509835362434, + 0.02150096371769905, + 0.008708936162292957, + 0.008027281612157822, + 0.00313648977316916, + 0.008582651615142822, + 0.004457230679690838, + 0.0056950803846120834, + 0.007858985103666782, + 0.008503641933202744, + 0.025753460824489594, + 0.011744699440896511, + 0.0056122527457773685, + 0.04622676223516464, + 0.016118478029966354, + 0.009763759560883045, + 0.007508137263357639, + 0.015510144643485546, + 0.012829373590648174, + 0.0025851281825453043, + 0.002446116181090474, + 0.003978559281677008, + 0.0024955521803349257, + 0.00315348943695426, + 0.0032628022599965334, + 0.003142458852380514, + 0.0037128885742276907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.030930591747164726 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309365898370743 + ], + "opposite_logprob_tokens": [ + -3.168093681335449 + ], + "logprob_reward": -0.04309365898370743, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309365898370743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1447172462940216, + 0.1483321189880371, + 0.049540355801582336, + 0.06511383503675461, + 0.05985599756240845, + 0.2817150354385376, + 0.06108459457755089, + 0.03256610780954361, + 0.11660517752170563, + 0.04046950489282608 + ], + "credit_map": [ + 0.01436250563710928, + 0.013579330407083035, + 0.01829962432384491, + 0.024992642924189568, + 0.013978867791593075, + 0.01839873008430004, + 0.02471752278506756, + 0.01638803817331791, + 0.013056316412985325, + 0.010722055099904537, + 0.029182009398937225, + 0.015063776634633541, + 0.0193464532494545, + 0.011166512034833431, + 0.030701622366905212, + 0.019093362614512444, + 0.003959511872380972, + 0.0031805229373276234, + 0.006306495051831007, + 0.003989935852587223, + 0.00839014258235693, + 0.005982245318591595, + 0.011877027340233326, + 0.005854470189660788, + 0.00517336605116725, + 0.004549622535705566, + 0.006447922438383102, + 0.005781332962214947, + 0.012149968184530735, + 0.004345605615526438, + 0.010799520649015903, + 0.015866493806242943, + 0.0051496210508048534, + 0.004120341036468744, + 0.010215257294476032, + 0.0063271839171648026, + 0.008614422753453255, + 0.005026976577937603, + 0.010631323792040348, + 0.009770873002707958, + 0.03885352984070778, + 0.011583348736166954, + 0.026829920709133148, + 0.013102056458592415, + 0.019862087443470955, + 0.031927041709423065, + 0.027006514370441437, + 0.1125505343079567, + 0.004280842375010252, + 0.008186555467545986, + 0.006743751000612974, + 0.00623351288959384, + 0.010225916281342506, + 0.005826046224683523, + 0.011847253888845444, + 0.0077407145872712135, + 0.0024717131163924932, + 0.002362312050536275, + 0.004400044213980436, + 0.0029334211722016335, + 0.004725078120827675, + 0.002673621289432049, + 0.008107520639896393, + 0.004892394877970219, + 0.011383725330233574, + 0.005571132525801659, + 0.04618579521775246, + 0.015948375687003136, + 0.008241607807576656, + 0.006382117513567209, + 0.011250743642449379, + 0.011641673743724823, + 0.002976583782583475, + 0.004208140075206757, + 0.005772564094513655, + 0.0057977549731731415, + 0.0043529001995921135, + 0.006932662799954414, + 0.00568179739639163, + 0.004747098311781883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.04309365898370743 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620587058365345 + ], + "opposite_logprob_tokens": [ + -2.812058687210083 + ], + "logprob_reward": -0.0620587058365345, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620587058365345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19082337617874146, + 0.1521700769662857, + 0.08316312730312347, + 0.06508889049291611, + 0.08082346618175507, + 0.13080020248889923, + 0.028645148500800133, + 0.061915140599012375, + 0.14910241961479187, + 0.057468123733997345 + ], + "credit_map": [ + 0.015676790848374367, + 0.01745356060564518, + 0.024513235315680504, + 0.035569604486227036, + 0.02419847808778286, + 0.019564634189009666, + 0.034837573766708374, + 0.019009487703442574, + 0.011649695225059986, + 0.013600719161331654, + 0.024196982383728027, + 0.013753569684922695, + 0.02649296633899212, + 0.015737565234303474, + 0.02912585623562336, + 0.017612723633646965, + 0.007208207622170448, + 0.008569438010454178, + 0.012234656140208244, + 0.010085836984217167, + 0.012915429659187794, + 0.012408236972987652, + 0.011403961107134819, + 0.008337362669408321, + 0.006265994161367416, + 0.005265952553600073, + 0.01152077317237854, + 0.006911616772413254, + 0.01019088551402092, + 0.004097986500710249, + 0.008649850264191628, + 0.012185826897621155, + 0.0063245864585042, + 0.00930657610297203, + 0.008733222261071205, + 0.01108612958341837, + 0.013317035511136055, + 0.006437179632484913, + 0.016419723629951477, + 0.009199012070894241, + 0.017750052735209465, + 0.004774951841682196, + 0.013005882501602173, + 0.00583137571811676, + 0.009516647085547447, + 0.015561357140541077, + 0.013971690088510513, + 0.05038825422525406, + 0.0033976223785430193, + 0.003255495335906744, + 0.004023352637887001, + 0.003322670701891184, + 0.003780352883040905, + 0.0035790198016911745, + 0.003848267486318946, + 0.0034383663441985846, + 0.0058387634344398975, + 0.003306424245238304, + 0.008122129365801811, + 0.005136431194841862, + 0.00779414176940918, + 0.0052251084707677364, + 0.010593309067189693, + 0.01589883118867874, + 0.01501261442899704, + 0.006929428782314062, + 0.056303881108760834, + 0.020213158801198006, + 0.01225932314991951, + 0.008370595052838326, + 0.014459224417805672, + 0.015554196201264858, + 0.004288297146558762, + 0.004493630025535822, + 0.010463154874742031, + 0.00843635480850935, + 0.00625056866556406, + 0.006398359779268503, + 0.010559037327766418, + 0.006578718777745962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0620587058365345 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499469116330147 + ], + "opposite_logprob_tokens": [ + -2.929994583129883 + ], + "logprob_reward": -0.05499469116330147, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499469116330147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1153220683336258, + 0.20781369507312775, + 0.04164009913802147, + 0.0678010955452919, + 0.13457433879375458, + 0.14921332895755768, + 0.07984999567270279, + 0.05930423364043236, + 0.03804147616028786, + 0.1064397543668747 + ], + "credit_map": [ + 0.014170406386256218, + 0.008701935410499573, + 0.014660466462373734, + 0.012648526579141617, + 0.010970440693199635, + 0.021472088992595673, + 0.014523847959935665, + 0.01817433536052704, + 0.01251497957855463, + 0.018445100635290146, + 0.026643997058272362, + 0.019234662875533104, + 0.028986159712076187, + 0.026883915066719055, + 0.04564706236124039, + 0.02945779450237751, + 0.003858741605654359, + 0.004106719512492418, + 0.005291616078466177, + 0.004662584979087114, + 0.007222774438560009, + 0.003901067888364196, + 0.005347984377294779, + 0.0072486079297959805, + 0.004391086287796497, + 0.003908379934728146, + 0.009123041294515133, + 0.006202883552759886, + 0.011186039075255394, + 0.0044624460861086845, + 0.018173789605498314, + 0.010353412479162216, + 0.013019039295613766, + 0.00967730488628149, + 0.031221363693475723, + 0.014627711847424507, + 0.011658714152872562, + 0.021037805825471878, + 0.019274480640888214, + 0.014057891443371773, + 0.018926767632365227, + 0.007991358637809753, + 0.01683141104876995, + 0.010044294409453869, + 0.011406747624278069, + 0.015225040726363659, + 0.014374196529388428, + 0.054413482546806335, + 0.0056608268059790134, + 0.007569064851850271, + 0.008815762586891651, + 0.00924625527113676, + 0.01363326795399189, + 0.00633035646751523, + 0.020306993275880814, + 0.008287456817924976, + 0.003798855235800147, + 0.0036083550658077, + 0.008924621157348156, + 0.0073878527618944645, + 0.007473353296518326, + 0.0064728339202702045, + 0.014179534278810024, + 0.007458818610757589, + 0.0034934221766889095, + 0.002512115053832531, + 0.013588600791990757, + 0.005392523016780615, + 0.0027240475174039602, + 0.0021938912104815245, + 0.004627636168152094, + 0.0035092344041913748, + 0.007600344251841307, + 0.010727833025157452, + 0.01262847799807787, + 0.012544911354780197, + 0.012852443382143974, + 0.016245650127530098, + 0.01816393807530403, + 0.015676140785217285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.05499469116330147 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627647835761309 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.007627647835761309, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627647835761309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20719045400619507, + 0.18895897269248962, + 0.04596978798508644, + 0.06059940159320831, + 0.09070616215467453, + 0.15571445226669312, + 0.048239290714263916, + 0.06800819933414459, + 0.07797087728977203, + 0.05664244294166565 + ], + "credit_map": [ + 0.04628993570804596, + 0.021316900849342346, + 0.02725084498524666, + 0.02014988474547863, + 0.01561498362571001, + 0.022191638126969337, + 0.03220776841044426, + 0.022168494760990143, + 0.018384095281362534, + 0.014294156804680824, + 0.026988696306943893, + 0.013216076418757439, + 0.027173813432455063, + 0.02137587033212185, + 0.030450336635112762, + 0.03707593306899071, + 0.008642465807497501, + 0.003772688563913107, + 0.00899623241275549, + 0.003938715904951096, + 0.004596153739839792, + 0.004458085633814335, + 0.005220803432166576, + 0.006344640627503395, + 0.005613714922219515, + 0.00397861935198307, + 0.01140256505459547, + 0.00694439560174942, + 0.011303937062621117, + 0.0035321121104061604, + 0.01030256412923336, + 0.007521489635109901, + 0.005626064259558916, + 0.006472872104495764, + 0.014325185678899288, + 0.01143505610525608, + 0.010710477828979492, + 0.008921687491238117, + 0.02043190598487854, + 0.012782905250787735, + 0.021495396271348, + 0.006016585044562817, + 0.01585007645189762, + 0.00866153184324503, + 0.011596866883337498, + 0.016057953238487244, + 0.016362736001610756, + 0.059673309326171875, + 0.008488332852721214, + 0.004571882076561451, + 0.00815516710281372, + 0.004698838107287884, + 0.006424389313906431, + 0.002633269177749753, + 0.008052160032093525, + 0.005215250886976719, + 0.006044864654541016, + 0.004018037114292383, + 0.010741373524069786, + 0.005259895231574774, + 0.00761454738676548, + 0.005344795528799295, + 0.013243039138615131, + 0.01574164442718029, + 0.005194566678255796, + 0.004865963011980057, + 0.023146146908402443, + 0.009305339306592941, + 0.007903520949184895, + 0.004799445625394583, + 0.014790070243179798, + 0.00796582456678152, + 0.006066245026886463, + 0.005035471171140671, + 0.008894766680896282, + 0.005553053691983223, + 0.006382201798260212, + 0.009147979319095612, + 0.0071189189329743385, + 0.008443806320428848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.007627647835761309 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium close-up shot, taken from a level angle, captures the silhouette of a young soldier, his head wearily resting against the cold window of an old train. He wears a well-worn, dark green woolen uniform with a stiff collar. His short, black hair, slightly damp with sweat, clung to his forehead. Crystal-clear tears silently fall from his bloodshot eyes, streaking wet trails along his dusty, young cheeks. His jaw is set, as if he's straining to suppress an impending collapse. The carriage sways gently and rhythmically with the rhythm of the rails, the dark wooden window frames and interior gleaming with a worn patina in the dim light. Outside, the distant mountains and interlaced fields are stretched and blurred by the rapid movement, completely losing their concrete form and vibrant colors, transforming into a dynamic, backward-flowing scroll of varying shades of gray, dark green, and ochre. The soft, cloudy diffuse light pouring in from the window sets a tranquil and sad cold tone for the entire scene. The picture is full of silent tension, which contrasts sharply with the soldier's heavy sorrow and the poetic and illusory scenery outside the window.", + "seed": 10064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_6/index_6731_seed_10064_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.021873726136982442, + "vlm_logprob": -0.021873726136982442, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.021873726136982442, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030042091384530067 + ], + "opposite_logprob_tokens": [ + -5.0300421714782715 + ], + "logprob_reward": -0.030042091384530067, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.030042091384530067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3892308473587036, + 0.10697518289089203, + 0.2179064303636551, + 0.07020040601491928, + 0.028887338936328888, + 0.03557052090764046, + 0.0393645204603672, + 0.04457312077283859, + 0.04413628950715065, + 0.023155370727181435 + ], + "credit_map": [ + 0.011545445770025253, + 0.005360208917409182, + 0.01914520375430584, + 0.029495781287550926, + 0.019048472866415977, + 0.08777287602424622, + 0.059056270867586136, + 0.1578066051006317, + 0.0036149946972727776, + 0.002462074626237154, + 0.006530323531478643, + 0.010723150335252285, + 0.007693049963563681, + 0.025486351922154427, + 0.006429573521018028, + 0.044035669416189194, + 0.009030764922499657, + 0.0071130190044641495, + 0.015816180035471916, + 0.022033393383026123, + 0.012641169130802155, + 0.04649316892027855, + 0.017631791532039642, + 0.08714693039655685, + 0.002919733989983797, + 0.004276283551007509, + 0.005689490120857954, + 0.008385797031223774, + 0.006624405272305012, + 0.012664234265685081, + 0.007512463256716728, + 0.022127997130155563, + 0.002773887710645795, + 0.0022775232791900635, + 0.0041301860474050045, + 0.0031329540070146322, + 0.0024650206323713064, + 0.004804592579603195, + 0.0024087291676551104, + 0.006894446909427643, + 0.0015597189776599407, + 0.0014720394974574447, + 0.0031404325272887945, + 0.00414649723097682, + 0.0020005269907414913, + 0.008760049007833004, + 0.002423131139948964, + 0.012068123556673527, + 0.002475835382938385, + 0.0031595381442457438, + 0.0042862012051045895, + 0.004649182315915823, + 0.003216851968318224, + 0.008892042562365532, + 0.0030570211820304394, + 0.00962784793227911, + 0.0033493484370410442, + 0.002508865436539054, + 0.00478965463116765, + 0.005739274434745312, + 0.0028973964508622885, + 0.01015855185687542, + 0.0028419781010597944, + 0.012288050726056099, + 0.0038099722005426884, + 0.0024419536348432302, + 0.006181066390126944, + 0.0041013737209141254, + 0.002879032166674733, + 0.008336505852639675, + 0.004006959963589907, + 0.012379425577819347, + 0.001936950022354722, + 0.0025973881129175425, + 0.0032354947179555893, + 0.0027060876600444317, + 0.002386271022260189, + 0.0022176429629325867, + 0.003841811092570424, + 0.00423372583463788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.030042091384530067 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033722419291734695 + ], + "opposite_logprob_tokens": [ + -3.408722400665283 + ], + "logprob_reward": -0.033722419291734695, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033722419291734695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1431054174900055, + 0.06595703959465027, + 0.5689256191253662, + 0.032282810658216476, + 0.024523066356778145, + 0.014649062417447567, + 0.09346409142017365, + 0.02039915882050991, + 0.0281931571662426, + 0.00850063469260931 + ], + "credit_map": [ + 0.005165111273527145, + 0.004452192224562168, + 0.009966546669602394, + 0.016910599544644356, + 0.008087518624961376, + 0.027548018842935562, + 0.021728988736867905, + 0.049246445298194885, + 0.0024087997153401375, + 0.0023287932854145765, + 0.004172069486230612, + 0.0070231398567557335, + 0.004497052635997534, + 0.015363738872110844, + 0.004312216304242611, + 0.02585122361779213, + 0.02169659361243248, + 0.009337225928902626, + 0.03822418674826622, + 0.05544847995042801, + 0.032558441162109375, + 0.1331583708524704, + 0.04504662752151489, + 0.23345567286014557, + 0.0014758408069610596, + 0.002947762608528137, + 0.003103366820141673, + 0.005592236295342445, + 0.003104313975200057, + 0.004699193872511387, + 0.0029659464489668608, + 0.008394152857363224, + 0.0011154935928061604, + 0.0013447708915919065, + 0.0030021436978131533, + 0.003192790085449815, + 0.0019020001636818051, + 0.003990491386502981, + 0.0037136184982955456, + 0.006261756177991629, + 0.0011048014275729656, + 0.0014863046817481518, + 0.002262379275634885, + 0.0018050272483378649, + 0.001339436275884509, + 0.0028285968583077192, + 0.0012687575072050095, + 0.0025537603069096804, + 0.005792974028736353, + 0.008862986229360104, + 0.01311549823731184, + 0.012415741570293903, + 0.009797144681215286, + 0.016421454027295113, + 0.009054196998476982, + 0.018004100769758224, + 0.0010152168106287718, + 0.0011486129369586706, + 0.0015984087949618697, + 0.002823285758495331, + 0.0017807574477046728, + 0.004382324405014515, + 0.0030459288973361254, + 0.004604623187333345, + 0.002926941029727459, + 0.0013940103817731142, + 0.00396713288500905, + 0.0024908166378736496, + 0.0014928465243428946, + 0.005491998977959156, + 0.0020490798633545637, + 0.008380330167710781, + 0.0007985398988239467, + 0.0008347849943675101, + 0.0014126294991001487, + 0.0011451945174485445, + 0.0007329325890168548, + 0.0011288764653727412, + 0.0007551652379333973, + 0.001692510792054236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.033722419291734695 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328900620341301 + ], + "opposite_logprob_tokens": [ + -3.7732889652252197 + ], + "logprob_reward": -0.02328900620341301, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328900620341301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4606928527355194, + 0.215948686003685, + 0.09878598898649216, + 0.04886874556541443, + 0.04863082244992256, + 0.04614126309752464, + 0.027498465031385422, + 0.029199352487921715, + 0.01538771390914917, + 0.008846115320920944 + ], + "credit_map": [ + 0.013657251372933388, + 0.008674977347254753, + 0.023966390639543533, + 0.036735594272613525, + 0.02298586443066597, + 0.11516986042261124, + 0.050293564796447754, + 0.18920932710170746, + 0.006474784109741449, + 0.004082974512130022, + 0.01163443736732006, + 0.021961310878396034, + 0.014968063682317734, + 0.05279896408319473, + 0.01261028740555048, + 0.09141788631677628, + 0.004988931119441986, + 0.005587918218225241, + 0.0073142824694514275, + 0.01222181599587202, + 0.008861933834850788, + 0.014105288311839104, + 0.015994613990187645, + 0.029711203649640083, + 0.001684323768131435, + 0.0017699952004477382, + 0.004593587946146727, + 0.00504120159894228, + 0.0021855586674064398, + 0.009488014504313469, + 0.0037502243649214506, + 0.020355841144919395, + 0.0019412162946537137, + 0.002051714574918151, + 0.005553157534450293, + 0.00557839497923851, + 0.003012370318174362, + 0.008952360600233078, + 0.00710321543738246, + 0.014438392594456673, + 0.0019097714684903622, + 0.004179567098617554, + 0.002915505087003112, + 0.006084240507334471, + 0.003681303234770894, + 0.009144226089119911, + 0.0033383022528141737, + 0.014888346195220947, + 0.0027194952126592398, + 0.0032478796783834696, + 0.004643664229661226, + 0.004487816244363785, + 0.002568192780017853, + 0.00418385025113821, + 0.002557512139901519, + 0.0030900556594133377, + 0.002038263948634267, + 0.0013146005803719163, + 0.002931000431999564, + 0.002768178703263402, + 0.0012404386652633548, + 0.007247659843415022, + 0.001958161359652877, + 0.009701049886643887, + 0.001329441205598414, + 0.0012350494507700205, + 0.002465883968397975, + 0.002177351852878928, + 0.001145046902820468, + 0.0025147933047264814, + 0.001513592666015029, + 0.003006554674357176, + 0.0006267029093578458, + 0.0009485699119977653, + 0.0009473053505644202, + 0.001294769230298698, + 0.0008411344024352729, + 0.000934025039896369, + 0.0012807531747967005, + 0.0019728546030819416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02328900620341301 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978199765086174 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.00978199765086174, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978199765086174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2635011672973633, + 0.16899307072162628, + 0.3328569531440735, + 0.10561804473400116, + 0.03389503434300423, + 0.010135599412024021, + 0.022151727229356766, + 0.017104782164096832, + 0.021241264417767525, + 0.024502338841557503 + ], + "credit_map": [ + 0.007861297577619553, + 0.004355572629719973, + 0.012180102989077568, + 0.020292483270168304, + 0.013002204708755016, + 0.06291011720895767, + 0.03443223983049393, + 0.10846713185310364, + 0.005821707658469677, + 0.003422458190470934, + 0.009251638315618038, + 0.016949165612459183, + 0.012112008407711983, + 0.04114539176225662, + 0.00951844360679388, + 0.07077226042747498, + 0.01347170490771532, + 0.00747879222035408, + 0.02452848106622696, + 0.032570790499448776, + 0.021371573209762573, + 0.07442007213830948, + 0.029326606541872025, + 0.12968891859054565, + 0.0038639053236693144, + 0.007351729553192854, + 0.008400008082389832, + 0.013082902878522873, + 0.01230754517018795, + 0.015423432923853397, + 0.013992111198604107, + 0.031196406111121178, + 0.0016234859358519316, + 0.0029585114680230618, + 0.0037826818879693747, + 0.0051543861627578735, + 0.002808609278872609, + 0.004568788688629866, + 0.006964509375393391, + 0.006034061312675476, + 0.0006651721778325737, + 0.0010393706616014242, + 0.0016913353465497494, + 0.0011701133335009217, + 0.000904753222130239, + 0.0019562477245926857, + 0.0008360891370102763, + 0.0018725180998444557, + 0.0019460226176306605, + 0.0018322075484320521, + 0.0036146966740489006, + 0.0030302435625344515, + 0.001871159765869379, + 0.003947349265217781, + 0.0023230360820889473, + 0.0035870117135345936, + 0.0012606888776645064, + 0.0008418901124969125, + 0.0013830643147230148, + 0.0017645284533500671, + 0.0009790754411369562, + 0.0036965063773095608, + 0.0014340681955218315, + 0.005744959693402052, + 0.0011767640244215727, + 0.0012380179250612855, + 0.001817414304241538, + 0.002191659528762102, + 0.0016486727399751544, + 0.004808375611901283, + 0.002182205207645893, + 0.006178154610097408, + 0.0033659813925623894, + 0.0026707351207733154, + 0.0031874526757746935, + 0.0034594859462231398, + 0.0012993619311600924, + 0.0030313394963741302, + 0.0020526100415736437, + 0.005435370374470949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.00978199765086174 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young soldier wearing a dark green woolen uniform with a stiff collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the soldier's short black hair, bloodshot eyes, and tears visible on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting the interior of an old train carriage with dark wooden window frames and a worn patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the view outside the window show blurred mountains and fields in shades of gray, dark green, and ochre moving backward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533116154372692 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.012533116154372692, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533116154372692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5732050538063049, + 0.06584320217370987, + 0.10789436101913452, + 0.029116233810782433, + 0.08331191539764404, + 0.04633299261331558, + 0.02371211349964142, + 0.03905200958251953, + 0.021432245150208473, + 0.010099914856255054 + ], + "credit_map": [ + 0.017508363351225853, + 0.008554142899811268, + 0.028191158547997475, + 0.04410478472709656, + 0.028505409136414528, + 0.1407388597726822, + 0.06990781426429749, + 0.23569445312023163, + 0.002033220836892724, + 0.0017097106901928782, + 0.004049869254231453, + 0.007088137790560722, + 0.0044342875480651855, + 0.015195967629551888, + 0.003955122083425522, + 0.02737688645720482, + 0.004380180966109037, + 0.0022464352659881115, + 0.007550576236099005, + 0.010452879592776299, + 0.006444038823246956, + 0.02479044906795025, + 0.009054755792021751, + 0.0429750420153141, + 0.0011186632327735424, + 0.001102990354411304, + 0.002565152943134308, + 0.003073771484196186, + 0.001420066924765706, + 0.005599600728601217, + 0.0024840282276272774, + 0.011751958169043064, + 0.0032882201485335827, + 0.0025857447180896997, + 0.008582809008657932, + 0.008385246619582176, + 0.0035646327305585146, + 0.020373720675706863, + 0.00515160383656621, + 0.03137993440032005, + 0.0033923808950930834, + 0.001523568294942379, + 0.0037342356517910957, + 0.005231658928096294, + 0.001927424455061555, + 0.011449014768004417, + 0.0024955158587545156, + 0.016579192131757736, + 0.0036409711465239525, + 0.001957498723641038, + 0.003288058564066887, + 0.002583634341135621, + 0.0018720881780609488, + 0.002897164784371853, + 0.0024314403999596834, + 0.005041258409619331, + 0.003372404957190156, + 0.0016304964665323496, + 0.0034704559948295355, + 0.0037269401364028454, + 0.001462304382584989, + 0.009386254474520683, + 0.002260761335492134, + 0.013742388226091862, + 0.0020188482012599707, + 0.0012414039811119437, + 0.003529628738760948, + 0.0022057811729609966, + 0.0014561965363100171, + 0.0036218599416315556, + 0.002231003949418664, + 0.005127521697431803, + 0.0010245929006487131, + 0.0005920036346651614, + 0.0010493187000975013, + 0.001066800905391574, + 0.000664629798848182, + 0.001606567995622754, + 0.001033987500704825, + 0.0030620121397078037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.012533116154372692 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_7/index_tensor([3062, 3062])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_7/index_tensor([3062, 3062])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..59fe000815f37491d1effdd51655c46074fe4942 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_7/index_tensor([3062, 3062])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, captured from a slightly elevated angle, slowly pans to the right, following a simple wooden boat as it slowly plows across the broad, calm river. A bright full moon hangs high in the inky night sky, its cool glow the sole source of light, casting a shimmering silver mirror on the water. The distant mountains appear as mere rolling black silhouettes. Atop the boat's bow, a poet stands steady, clad in a flowing moon-white silk robe. His long hair is held back by a simple headband, a few strands brushing against his thin face in the cool night breeze. With one hand clasped behind his back, he gazes at the moon's reflection on the distant water, his brow etched with deep thought and a subtle hint of loneliness. The boat glides silently, the waves plowed by the bow shimmering silver in the moonlight, rippling outwards. His wide robe sleeves flutter gently in the breeze, as if to blend into the boundless night. The entire picture is dominated by deep indigo and silver-white colors, creating an atmosphere of extreme tranquility, loneliness and classical poetry.", + "seed": 19644, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_7/index_3062_seed_19644_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12072438327595592, + "vlm_logprob": -0.12072438327595592, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.12072438327595592, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35685086250305176 + ], + "opposite_logprob_tokens": [ + -1.2318508625030518 + ], + "logprob_reward": -0.35685086250305176, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35685086250305176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41263124346733093, + 0.13833217322826385, + 0.037834133952856064, + 0.03289405629038811, + 0.057341162115335464, + 0.04621012136340141, + 0.09150417894124985, + 0.038136791437864304, + 0.04283352568745613, + 0.10228265821933746 + ], + "credit_map": [ + 0.05495879799127579, + 0.03347339853644371, + 0.05497642979025841, + 0.03506150469183922, + 0.03333822637796402, + 0.07682590931653976, + 0.04076007008552551, + 0.08323688805103302, + 0.014262381009757519, + 0.015665777027606964, + 0.017620457336306572, + 0.013842886313796043, + 0.012103447690606117, + 0.022594396024942398, + 0.021460365504026413, + 0.020782470703125, + 0.004104976076632738, + 0.004841173067688942, + 0.004558075685054064, + 0.003401388181373477, + 0.004281447734683752, + 0.005285655613988638, + 0.005664808675646782, + 0.0056966086849570274, + 0.0032891975715756416, + 0.003092175116762519, + 0.006100783124566078, + 0.0031050352845340967, + 0.002808926161378622, + 0.006055969744920731, + 0.003659873967990279, + 0.004782095551490784, + 0.007934235036373138, + 0.0046196142211556435, + 0.0076663983054459095, + 0.005811349488794804, + 0.005507928319275379, + 0.00975300557911396, + 0.007585612591356039, + 0.008463019505143166, + 0.007725245784968138, + 0.005218944977968931, + 0.00624118372797966, + 0.004850884899497032, + 0.003891110187396407, + 0.007029819767922163, + 0.005191158968955278, + 0.0060617742128670216, + 0.014601833187043667, + 0.0077513353899121284, + 0.011926611885428429, + 0.009471921250224113, + 0.007530537433922291, + 0.015843374654650688, + 0.01035491842776537, + 0.014023649506270885, + 0.0034312144853174686, + 0.005442099180072546, + 0.0045185391791164875, + 0.005828834604471922, + 0.0035190177150070667, + 0.0050521777011454105, + 0.0045084464363753796, + 0.005836465395987034, + 0.00561754684895277, + 0.0037402906455099583, + 0.004996097180992365, + 0.004266060888767242, + 0.0052106305956840515, + 0.008733348920941353, + 0.004563221707940102, + 0.005706325173377991, + 0.010140582919120789, + 0.007658571004867554, + 0.014044941402971745, + 0.009077907539904118, + 0.009426972828805447, + 0.021544739603996277, + 0.013162168674170971, + 0.017226768657565117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.35685086250305176 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00596374599263072 + ], + "opposite_logprob_tokens": [ + -5.1309638023376465 + ], + "logprob_reward": -0.00596374599263072, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00596374599263072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3758220672607422, + 0.14861689507961273, + 0.06252101063728333, + 0.04679892584681511, + 0.07782900333404541, + 0.0530744269490242, + 0.04585396870970726, + 0.058585695922374725, + 0.056401073932647705, + 0.07449691742658615 + ], + "credit_map": [ + 0.05856306850910187, + 0.03323680907487869, + 0.03526927903294563, + 0.04154738411307335, + 0.0306700449436903, + 0.0676908791065216, + 0.04951973259449005, + 0.05932486429810524, + 0.01991511695086956, + 0.015838773921132088, + 0.01705017127096653, + 0.016045596450567245, + 0.011620881035923958, + 0.02561742067337036, + 0.018420204520225525, + 0.024108733981847763, + 0.008065734058618546, + 0.007121462374925613, + 0.00516154570505023, + 0.008114523254334927, + 0.005072299391031265, + 0.00881197489798069, + 0.011300666257739067, + 0.008872808888554573, + 0.004809307400137186, + 0.00476017827168107, + 0.005541256628930569, + 0.004481475334614515, + 0.004184782039374113, + 0.009966752491891384, + 0.004907272756099701, + 0.008147901855409145, + 0.006902790162712336, + 0.008508170954883099, + 0.007780193351209164, + 0.00929506216198206, + 0.00736653245985508, + 0.015933996066451073, + 0.008594285696744919, + 0.013447978533804417, + 0.005772798787802458, + 0.00570922065526247, + 0.0051317838951945305, + 0.006389066111296415, + 0.005405794829130173, + 0.009787002578377724, + 0.005928180646151304, + 0.008950578980147839, + 0.00687484722584486, + 0.0041334661655128, + 0.00607322808355093, + 0.004598888568580151, + 0.0032576341181993484, + 0.007808085065335035, + 0.005570536945015192, + 0.007537282537668943, + 0.005997700151056051, + 0.008228754624724388, + 0.008209454827010632, + 0.008077599108219147, + 0.00541488965973258, + 0.0074780043214559555, + 0.006736191920936108, + 0.008443104103207588, + 0.005840957630425692, + 0.005087932571768761, + 0.009347695857286453, + 0.006136967334896326, + 0.006170930340886116, + 0.008996703661978245, + 0.006597826257348061, + 0.008222066797316074, + 0.006735974457114935, + 0.009326746687293053, + 0.009716439992189407, + 0.008148791268467903, + 0.006549406796693802, + 0.013580767437815666, + 0.008994570933282375, + 0.01144422683864832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.00596374599263072 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632971741259098 + ], + "opposite_logprob_tokens": [ + -4.882633209228516 + ], + "logprob_reward": -0.007632971741259098, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632971741259098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3830350935459137, + 0.12842710316181183, + 0.08516369760036469, + 0.0794786885380745, + 0.06085830181837082, + 0.03527098149061203, + 0.05514650046825409, + 0.03926253318786621, + 0.05829421058297157, + 0.0750628411769867 + ], + "credit_map": [ + 0.05659256502985954, + 0.031471315771341324, + 0.049323178827762604, + 0.03734669089317322, + 0.03472891077399254, + 0.06424292922019958, + 0.04011420160531998, + 0.0692153126001358, + 0.018074600026011467, + 0.012679395265877247, + 0.0165698379278183, + 0.013665411621332169, + 0.010175750590860844, + 0.018371395766735077, + 0.018630212172865868, + 0.020260518416762352, + 0.009372195228934288, + 0.008184470236301422, + 0.009931758977472782, + 0.009529227390885353, + 0.006967086810618639, + 0.016758881509304047, + 0.01176429819315672, + 0.012655781581997871, + 0.00949144922196865, + 0.008126656524837017, + 0.010501448065042496, + 0.007639626506716013, + 0.006882519461214542, + 0.012484588660299778, + 0.009249706752598286, + 0.015102699398994446, + 0.006886136252433062, + 0.008450064808130264, + 0.007585043087601662, + 0.005793888121843338, + 0.0053626312874257565, + 0.009473919868469238, + 0.008517589420080185, + 0.008789035491645336, + 0.0038436728063970804, + 0.004689383320510387, + 0.003796328091993928, + 0.003300946671515703, + 0.0030269725248217583, + 0.006544794887304306, + 0.004776250571012497, + 0.005292637273669243, + 0.008975249715149403, + 0.004203466698527336, + 0.008742498233914375, + 0.0056074014864861965, + 0.004011903423815966, + 0.008940878324210644, + 0.0057433051988482475, + 0.008921797387301922, + 0.004790321923792362, + 0.006529118400067091, + 0.0045762695372104645, + 0.004364232067018747, + 0.0034592244774103165, + 0.005806598346680403, + 0.0042188470251858234, + 0.0055179232731461525, + 0.006464394740760326, + 0.006933278404176235, + 0.0063789174892008305, + 0.008069939911365509, + 0.006999765522778034, + 0.009570428170263767, + 0.006761331111192703, + 0.007116158492863178, + 0.007622052449733019, + 0.006503901444375515, + 0.010739949531853199, + 0.006089778151363134, + 0.008521326817572117, + 0.015759844332933426, + 0.007579735480248928, + 0.012246260419487953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.007632971741259098 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255505919456482 + ], + "opposite_logprob_tokens": [ + -1.600550651550293 + ], + "logprob_reward": -0.2255505919456482, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255505919456482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.139640212059021, + 0.07326657325029373, + 0.11305925250053406, + 0.10863931477069855, + 0.10681424289941788, + 0.05887794867157936, + 0.07262104004621506, + 0.22226814925670624, + 0.046360015869140625, + 0.05845329165458679 + ], + "credit_map": [ + 0.012802310287952423, + 0.020095085725188255, + 0.019901596009731293, + 0.01870102435350418, + 0.015811091288924217, + 0.017521174624562263, + 0.017066091299057007, + 0.017741838470101357, + 0.006991290487349033, + 0.009598017670214176, + 0.007072737440466881, + 0.01046944223344326, + 0.0076745301485061646, + 0.01137651689350605, + 0.009840749204158783, + 0.010243291035294533, + 0.00946175679564476, + 0.012115705758333206, + 0.009798301383852959, + 0.02070964127779007, + 0.013169013895094395, + 0.014116816222667694, + 0.01278874184936285, + 0.020899271592497826, + 0.011509900912642479, + 0.012345097959041595, + 0.011968057602643967, + 0.015930965542793274, + 0.011596170254051685, + 0.01849840022623539, + 0.008899809792637825, + 0.01789090782403946, + 0.013390188105404377, + 0.016043243929743767, + 0.013154872693121433, + 0.0119943143799901, + 0.008073925040662289, + 0.016396373510360718, + 0.01403387077152729, + 0.013727459125220776, + 0.006594523787498474, + 0.008128590881824493, + 0.005221969448029995, + 0.006896506063640118, + 0.006118312012404203, + 0.009488922543823719, + 0.007927937433123589, + 0.008501186035573483, + 0.008449546992778778, + 0.00829688087105751, + 0.007553088944405317, + 0.007610435597598553, + 0.008063838817179203, + 0.012682884931564331, + 0.008219491690397263, + 0.01174487266689539, + 0.026859814301133156, + 0.020532768219709396, + 0.027060002088546753, + 0.02610979601740837, + 0.026579512283205986, + 0.033405203372240067, + 0.02571089193224907, + 0.03601016104221344, + 0.006874010432511568, + 0.0068779573775827885, + 0.004566975869238377, + 0.0066712419502437115, + 0.004020326305180788, + 0.006269522476941347, + 0.005352557171136141, + 0.005727423820644617, + 0.0058061243034899235, + 0.007087992038577795, + 0.006858541164547205, + 0.006735390983521938, + 0.005258405115455389, + 0.0097355293110013, + 0.006673119962215424, + 0.010298185050487518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.2255505919456482 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623744197189808 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.007623744197189808, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623744197189808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39577868580818176, + 0.16595898568630219, + 0.07477953284978867, + 0.041800737380981445, + 0.09076040983200073, + 0.06323694437742233, + 0.022405557334423065, + 0.04320420324802399, + 0.039438605308532715, + 0.06263631582260132 + ], + "credit_map": [ + 0.04988100752234459, + 0.028858866542577744, + 0.06549497693777084, + 0.035404179245233536, + 0.03177966549992561, + 0.06172466650605202, + 0.03929492086172104, + 0.08334039896726608, + 0.02234957180917263, + 0.016892878338694572, + 0.02468942664563656, + 0.017265377566218376, + 0.013811451382935047, + 0.024646146222949028, + 0.01975887268781662, + 0.026545273140072823, + 0.008996625430881977, + 0.006869187578558922, + 0.010202068835496902, + 0.005318565294146538, + 0.0063577559776604176, + 0.017377430573105812, + 0.00794028677046299, + 0.011717617511749268, + 0.00482817180454731, + 0.00333874998614192, + 0.005086452234536409, + 0.0034339202102273703, + 0.0027732730377465487, + 0.009349651634693146, + 0.005805816035717726, + 0.007184699643403292, + 0.017450451850891113, + 0.006646749097853899, + 0.012752741575241089, + 0.007905757986009121, + 0.007251090835779905, + 0.016564980149269104, + 0.008427322842180729, + 0.013761317357420921, + 0.011368238367140293, + 0.004150531254708767, + 0.007357921451330185, + 0.004980682395398617, + 0.0054021053947508335, + 0.012666309252381325, + 0.006739065051078796, + 0.010572095401585102, + 0.0029117222875356674, + 0.0027048150077462196, + 0.002802417613565922, + 0.002207939513027668, + 0.0019815564155578613, + 0.0040782783180475235, + 0.0025765919126570225, + 0.003142236266285181, + 0.006089248228818178, + 0.003881481708958745, + 0.005974071566015482, + 0.004832664038985968, + 0.004232693929225206, + 0.007121324073523283, + 0.004070458933711052, + 0.007002264726907015, + 0.0048294030129909515, + 0.004407282453030348, + 0.0049286652356386185, + 0.00470754224807024, + 0.0035152549389749765, + 0.0074310582131147385, + 0.003731334349140525, + 0.005888063460588455, + 0.00706818513572216, + 0.0056308419443666935, + 0.00868386309593916, + 0.005028665065765381, + 0.007209249306470156, + 0.013464663177728653, + 0.005568925756961107, + 0.009981927461922169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.007623744197189808 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, captured from a slightly elevated angle, slowly pans to the right, following a simple wooden boat as it slowly plows across the broad, calm river. A bright full moon hangs high in the inky night sky, its cool glow the sole source of light, casting a shimmering silver mirror on the water. The distant mountains appear as mere rolling black silhouettes. Atop the boat's bow, a poet stands steady, clad in a flowing moon-white silk robe. His long hair is held back by a simple headband, a few strands brushing against his thin face in the cool night breeze. With one hand clasped behind his back, he gazes at the moon's reflection on the distant water, his brow etched with deep thought and a subtle hint of loneliness. The boat glides silently, the waves plowed by the bow shimmering silver in the moonlight, rippling outwards. His wide robe sleeves flutter gently in the breeze, as if to blend into the boundless night. The entire picture is dominated by deep indigo and silver-white colors, creating an atmosphere of extreme tranquility, loneliness and classical poetry.", + "seed": 19645, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_7/index_3062_seed_19645_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02656565448269248, + "vlm_logprob": -0.02656565448269248, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.02656565448269248, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023973237723112106 + ], + "opposite_logprob_tokens": [ + -4.27397346496582 + ], + "logprob_reward": -0.023973237723112106, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.023973237723112106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09678665548563004, + 0.060997456312179565, + 0.03806426003575325, + 0.13784676790237427, + 0.21732495725154877, + 0.05235438793897629, + 0.049660466611385345, + 0.06552986800670624, + 0.1964227259159088, + 0.0850125178694725 + ], + "credit_map": [ + 0.006854664534330368, + 0.011721438728272915, + 0.016419988125562668, + 0.009497785940766335, + 0.012595267966389656, + 0.01195463165640831, + 0.010873764753341675, + 0.01686910167336464, + 0.0054166861809790134, + 0.005084885284304619, + 0.011286205612123013, + 0.003897336544468999, + 0.00897408090531826, + 0.007502072025090456, + 0.009334709495306015, + 0.00950147956609726, + 0.0029734736308455467, + 0.0036745071411132812, + 0.0051673646084964275, + 0.0068205539137125015, + 0.0057160030119121075, + 0.004617504309862852, + 0.0038540931418538094, + 0.0052407593466341496, + 0.00687051098793745, + 0.009024416096508503, + 0.04184388369321823, + 0.005977585446089506, + 0.006108712870627642, + 0.012910373508930206, + 0.005561317782849073, + 0.049549952149391174, + 0.011846933513879776, + 0.006790074985474348, + 0.08817531168460846, + 0.010185793042182922, + 0.00506236357614398, + 0.014627690427005291, + 0.006057462189346552, + 0.07457929849624634, + 0.004169262479990721, + 0.003684108844026923, + 0.014259136281907558, + 0.005218567792326212, + 0.00510355643928051, + 0.006393210496753454, + 0.0039506945759058, + 0.009575842879712582, + 0.0032444153912365437, + 0.0038342003244906664, + 0.010219073854386806, + 0.006747706793248653, + 0.00429768580943346, + 0.004427710548043251, + 0.0038662515580654144, + 0.013023416511714458, + 0.004063508938997984, + 0.007119765970855951, + 0.00770154083147645, + 0.006764475256204605, + 0.0141707519069314, + 0.006322586443275213, + 0.010768569074571133, + 0.008618668653070927, + 0.012555151246488094, + 0.017363188788294792, + 0.023959394544363022, + 0.019574183970689774, + 0.024105427786707878, + 0.04347950592637062, + 0.013584067113697529, + 0.041801780462265015, + 0.005819479003548622, + 0.007334048394113779, + 0.014940636232495308, + 0.010835127905011177, + 0.007568829692900181, + 0.012455384247004986, + 0.01121667679399252, + 0.014842330478131771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.023973237723112106 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01109242532402277 + ], + "opposite_logprob_tokens": [ + -4.511092662811279 + ], + "logprob_reward": -0.01109242532402277, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01109242532402277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15186253190040588, + 0.03799152001738548, + 0.029795542359352112, + 0.07501162588596344, + 0.10614507645368576, + 0.0369204618036747, + 0.020286299288272858, + 0.12393954396247864, + 0.26763176918029785, + 0.15041567385196686 + ], + "credit_map": [ + 0.02007896825671196, + 0.016602706164121628, + 0.03208981826901436, + 0.01613633707165718, + 0.014263675548136234, + 0.014373164623975754, + 0.01125563308596611, + 0.02706221304833889, + 0.0030864791478961706, + 0.003094266401603818, + 0.008455307222902775, + 0.0033226008526980877, + 0.005933328066021204, + 0.0037520478945225477, + 0.0046804421581327915, + 0.005667044781148434, + 0.0019511532736942172, + 0.00265330309048295, + 0.00415195245295763, + 0.0024284389801323414, + 0.0043417480774223804, + 0.00551869161427021, + 0.003153139492496848, + 0.005597114097326994, + 0.0034284021239727736, + 0.0032274737022817135, + 0.02569281868636608, + 0.0033867554739117622, + 0.005011261906474829, + 0.004767707549035549, + 0.0035391724668443203, + 0.025958025828003883, + 0.0058885421603918076, + 0.0067517682909965515, + 0.03504735231399536, + 0.009502884931862354, + 0.005606925114989281, + 0.011540938168764114, + 0.006387329660356045, + 0.02541932463645935, + 0.0023921970278024673, + 0.0032060672529041767, + 0.009330998174846172, + 0.00323838135227561, + 0.005314255133271217, + 0.004187170881778002, + 0.0033603478223085403, + 0.005891043692827225, + 0.0018921063747256994, + 0.0020239197183400393, + 0.003726300550624728, + 0.0024228054098784924, + 0.001973099773749709, + 0.0031225376296788454, + 0.0019360420992597938, + 0.0031894869171082973, + 0.0072566126473248005, + 0.007877789437770844, + 0.025293899700045586, + 0.012094532139599323, + 0.01807476207613945, + 0.008422533050179482, + 0.01770850643515587, + 0.0272109005600214, + 0.017652736976742744, + 0.02381679229438305, + 0.030558794736862183, + 0.047371890395879745, + 0.039530109614133835, + 0.04306910187005997, + 0.026333820074796677, + 0.03929850459098816, + 0.009019880555570126, + 0.015141257084906101, + 0.016285410150885582, + 0.025375157594680786, + 0.019081633538007736, + 0.029974214732646942, + 0.014367884024977684, + 0.021170223131775856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.01109242532402277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046426597982645035 + ], + "opposite_logprob_tokens": [ + -5.379642486572266 + ], + "logprob_reward": -0.0046426597982645035, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046426597982645035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2506994307041168, + 0.09086739271879196, + 0.07170481234788895, + 0.04700569808483124, + 0.17166772484779358, + 0.02110288478434086, + 0.042423177510499954, + 0.09905799478292465, + 0.11616452038288116, + 0.08930638432502747 + ], + "credit_map": [ + 0.04478771612048149, + 0.030979285016655922, + 0.04495372995734215, + 0.02329404093325138, + 0.022410916164517403, + 0.030544491484761238, + 0.017270946875214577, + 0.036458324640989304, + 0.0068866838701069355, + 0.011791097931563854, + 0.013196965679526329, + 0.010256505571305752, + 0.013275889679789543, + 0.009069392457604408, + 0.012609899044036865, + 0.013780957087874413, + 0.006065084133297205, + 0.0071938601322472095, + 0.011228764429688454, + 0.0057010105811059475, + 0.006595478858798742, + 0.014253993518650532, + 0.0067606219090521336, + 0.013905995525419712, + 0.002504791133105755, + 0.007390153128653765, + 0.005645036231726408, + 0.004320133943110704, + 0.005772064905613661, + 0.006488699000328779, + 0.0054262843914330006, + 0.00945853628218174, + 0.008996912278234959, + 0.007239396218210459, + 0.06250189244747162, + 0.014652885496616364, + 0.0072505902498960495, + 0.01356465183198452, + 0.00851445458829403, + 0.04894694313406944, + 0.001480336650274694, + 0.0020216465927660465, + 0.003177199512720108, + 0.0021759779192507267, + 0.004344538319855928, + 0.0022331937216222286, + 0.002465392928570509, + 0.003204596694558859, + 0.0028011263348162174, + 0.00454728165641427, + 0.010010093450546265, + 0.0038229902274906635, + 0.0028893123380839825, + 0.004214600659906864, + 0.0027401086408644915, + 0.011397664435207844, + 0.005105762742459774, + 0.006201060023158789, + 0.03055688366293907, + 0.0061904024332761765, + 0.008054232224822044, + 0.008602268062531948, + 0.008848294615745544, + 0.025499092414975166, + 0.009855536743998528, + 0.012625913135707378, + 0.018802093341946602, + 0.021102147176861763, + 0.01710023544728756, + 0.012312737293541431, + 0.011635356582701206, + 0.01273050345480442, + 0.006163456477224827, + 0.008571779355406761, + 0.00787390023469925, + 0.01106941420584917, + 0.009815562516450882, + 0.024310775101184845, + 0.006129784509539604, + 0.015371711924672127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.0046426597982645035 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902670443058014 + ], + "opposite_logprob_tokens": [ + -2.464026689529419 + ], + "logprob_reward": -0.08902670443058014, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902670443058014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05996837094426155, + 0.029974201694130898, + 0.029564769938588142, + 0.1899169534444809, + 0.0891055092215538, + 0.025125646963715553, + 0.01935972273349762, + 0.2874620854854584, + 0.17413544654846191, + 0.09538727253675461 + ], + "credit_map": [ + 0.005978433880954981, + 0.006585937459021807, + 0.00991230458021164, + 0.008047820068895817, + 0.0095950523391366, + 0.008248528465628624, + 0.0035077189095318317, + 0.008092578500509262, + 0.0024126966018229723, + 0.0029568548779934645, + 0.005247047636657953, + 0.002683625789359212, + 0.0037062617484480143, + 0.0054705459624528885, + 0.0029691432137042284, + 0.004528027027845383, + 0.002176959766075015, + 0.002450154162943363, + 0.006664232816547155, + 0.00269987597130239, + 0.00447533605620265, + 0.004935404751449823, + 0.0027613176498562098, + 0.00340148713439703, + 0.009843595325946808, + 0.004536968190222979, + 0.07554154098033905, + 0.0068311532959342, + 0.0038011439610272646, + 0.010102396830916405, + 0.005667971912771463, + 0.07359219342470169, + 0.004850259516388178, + 0.0033980081789195538, + 0.03093201294541359, + 0.005722265690565109, + 0.0038516030181199312, + 0.006442195735871792, + 0.004461916629225016, + 0.029447242617607117, + 0.0018430092604830861, + 0.0021082106977701187, + 0.005099254194647074, + 0.00208885851316154, + 0.0022303916048258543, + 0.0027233234141021967, + 0.002239532768726349, + 0.006793066393584013, + 0.0019831142853945494, + 0.0017107349121943116, + 0.003196816425770521, + 0.0027104695327579975, + 0.002877134131267667, + 0.002328589092940092, + 0.0018871532520279288, + 0.002665710635483265, + 0.01445069070905447, + 0.014878520742058754, + 0.09320519119501114, + 0.01617482863366604, + 0.02137652412056923, + 0.0192111749202013, + 0.023899560794234276, + 0.08426559716463089, + 0.010251226834952831, + 0.01700565405189991, + 0.018457839265465736, + 0.016533778980374336, + 0.021036608144640923, + 0.03482033684849739, + 0.013386721722781658, + 0.042643286287784576, + 0.006213134154677391, + 0.00861929077655077, + 0.011663227342069149, + 0.009139890782535076, + 0.010515651665627956, + 0.01424077246338129, + 0.011899839155375957, + 0.02309546247124672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.08902670443058014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040932451374828815 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.0040932451374828815, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040932451374828815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0771234780550003, + 0.06520380079746246, + 0.047937147319316864, + 0.042978908866643906, + 0.12100246548652649, + 0.03935442864894867, + 0.03940253704786301, + 0.2154572308063507, + 0.21391615271568298, + 0.1376238316297531 + ], + "credit_map": [ + 0.013959022238850594, + 0.00819198414683342, + 0.01015909481793642, + 0.0070396000519394875, + 0.011271320283412933, + 0.010743469931185246, + 0.006115201860666275, + 0.009643787518143654, + 0.004454502370208502, + 0.007109396625310183, + 0.011446677148342133, + 0.006698467768728733, + 0.010790728963911533, + 0.007212381809949875, + 0.007973391562700272, + 0.009518258273601532, + 0.004630774259567261, + 0.0047662402503192425, + 0.00794174149632454, + 0.004370959009975195, + 0.006478558760136366, + 0.009218720719218254, + 0.004410092253237963, + 0.0061200568452477455, + 0.001865679514594376, + 0.0008975230157375336, + 0.016190102323889732, + 0.0020899251103401184, + 0.0014771124115213752, + 0.002194669097661972, + 0.0019426001235842705, + 0.016321299597620964, + 0.006799240130931139, + 0.0031433303374797106, + 0.04806280881166458, + 0.006698304321616888, + 0.004094790201634169, + 0.007686669938266277, + 0.0034935344010591507, + 0.04102377966046333, + 0.0023618589621037245, + 0.003608893370255828, + 0.005597077775746584, + 0.0069409990683197975, + 0.006009558215737343, + 0.004072057083249092, + 0.004552313592284918, + 0.0062116715125739574, + 0.002771205035969615, + 0.00457047950476408, + 0.008532931096851826, + 0.004262489732354879, + 0.005012155044823885, + 0.003004153026267886, + 0.004061122424900532, + 0.0071880025789141655, + 0.013163664378225803, + 0.018214888870716095, + 0.060635797679424286, + 0.013208362273871899, + 0.023142216727137566, + 0.02761349454522133, + 0.015584331937134266, + 0.04389448091387749, + 0.01850726641714573, + 0.0262446328997612, + 0.03261778876185417, + 0.037848617881536484, + 0.01876136288046837, + 0.03505035862326622, + 0.021103763952851295, + 0.02378237247467041, + 0.006979288533329964, + 0.01230007503181696, + 0.013400995172560215, + 0.02439909428358078, + 0.010117151774466038, + 0.030653292313218117, + 0.017393790185451508, + 0.022380152717232704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.0040932451374828815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, captured from a slightly elevated angle, slowly pans to the right, following a simple wooden boat as it slowly plows across the broad, calm river. A bright full moon hangs high in the inky night sky, its cool glow the sole source of light, casting a shimmering silver mirror on the water. The distant mountains appear as mere rolling black silhouettes. Atop the boat's bow, a poet stands steady, clad in a flowing moon-white silk robe. His long hair is held back by a simple headband, a few strands brushing against his thin face in the cool night breeze. With one hand clasped behind his back, he gazes at the moon's reflection on the distant water, his brow etched with deep thought and a subtle hint of loneliness. The boat glides silently, the waves plowed by the bow shimmering silver in the moonlight, rippling outwards. His wide robe sleeves flutter gently in the breeze, as if to blend into the boundless night. The entire picture is dominated by deep indigo and silver-white colors, creating an atmosphere of extreme tranquility, loneliness and classical poetry.", + "seed": 19646, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_7/index_3062_seed_19646_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.033828160137636586, + "vlm_logprob": -0.033828160137636586, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.033828160137636586, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01653139665722847 + ], + "opposite_logprob_tokens": [ + -4.641531467437744 + ], + "logprob_reward": -0.01653139665722847, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01653139665722847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6952637434005737, + 0.04124909266829491, + 0.03978023678064346, + 0.012983890250325203, + 0.061707574874162674, + 0.016516152769327164, + 0.0358181893825531, + 0.02583092451095581, + 0.05337827280163765, + 0.017471905797719955 + ], + "credit_map": [ + 0.05853067338466644, + 0.01849057339131832, + 0.19247190654277802, + 0.030420908704400063, + 0.04453154653310776, + 0.257550448179245, + 0.03987982124090195, + 0.05338786169886589, + 0.005352301057428122, + 0.004626941867172718, + 0.0032389811240136623, + 0.001982372719794512, + 0.008481666445732117, + 0.006023953203111887, + 0.005537431687116623, + 0.0060054403729736805, + 0.007629908621311188, + 0.003195509547367692, + 0.0029353396967053413, + 0.0029662924353033304, + 0.008677050471305847, + 0.005857957061380148, + 0.005180323496460915, + 0.0033378531225025654, + 0.002143747406080365, + 0.0012843410950154066, + 0.001325722667388618, + 0.0008348078699782491, + 0.001775842858478427, + 0.0018943384056910872, + 0.0025537919718772173, + 0.0011712987907230854, + 0.008464953862130642, + 0.004421061370521784, + 0.003462107153609395, + 0.0035117738880217075, + 0.013637282885611057, + 0.01019287109375, + 0.01189461350440979, + 0.006122910417616367, + 0.0025086274836212397, + 0.0011708131060004234, + 0.001318872207775712, + 0.0012181231286376715, + 0.0034971844870597124, + 0.0027797159273177385, + 0.002328399568796158, + 0.0016944167437031865, + 0.004245216492563486, + 0.00479961046949029, + 0.0038843434303998947, + 0.0030737686902284622, + 0.005512000992894173, + 0.004976148717105389, + 0.00551558006554842, + 0.0038115205243229866, + 0.003235412761569023, + 0.002618058817461133, + 0.003827302949503064, + 0.0022319802083075047, + 0.003350625978782773, + 0.0035053568426519632, + 0.004277521278709173, + 0.00278466590680182, + 0.00968346931040287, + 0.012196293100714684, + 0.003634125692769885, + 0.0021725192200392485, + 0.004787340760231018, + 0.008401602506637573, + 0.006934265606105328, + 0.005568658001720905, + 0.0014062185073271394, + 0.0024468612391501665, + 0.0012606446398422122, + 0.0015276952181011438, + 0.0030132269021123648, + 0.0028359247371554375, + 0.003111781319603324, + 0.0018695540493354201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.01653139665722847 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089242994785309 + ], + "opposite_logprob_tokens": [ + -4.511089324951172 + ], + "logprob_reward": -0.011089242994785309, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089242994785309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18301160633563995, + 0.09927879273891449, + 0.08975068479776382, + 0.07240738719701767, + 0.030741818249225616, + 0.19227640330791473, + 0.09230151772499084, + 0.10823822766542435, + 0.07226172834634781, + 0.05973184108734131 + ], + "credit_map": [ + 0.016151508316397667, + 0.010601322166621685, + 0.03834613040089607, + 0.009808075614273548, + 0.015426845289766788, + 0.04540412500500679, + 0.026490425691008568, + 0.02078317105770111, + 0.011344702914357185, + 0.010362244211137295, + 0.008122230879962444, + 0.005345564801245928, + 0.01876283809542656, + 0.014137404039502144, + 0.015407937578856945, + 0.015795867890119553, + 0.014579476788640022, + 0.009392812848091125, + 0.007369395345449448, + 0.006544093601405621, + 0.017268287017941475, + 0.012923063710331917, + 0.013075234368443489, + 0.008598322980105877, + 0.009337803348898888, + 0.008257288485765457, + 0.0057448651641607285, + 0.005008575972169638, + 0.011707590892910957, + 0.009450250305235386, + 0.015822287648916245, + 0.007078729569911957, + 0.003345303703099489, + 0.002743353135883808, + 0.0022096368484199047, + 0.0026829380076378584, + 0.004412887152284384, + 0.005150030367076397, + 0.007119530811905861, + 0.003078136593103409, + 0.02224096842110157, + 0.013295922428369522, + 0.013136726804077625, + 0.01650742068886757, + 0.026855837553739548, + 0.03473043814301491, + 0.04355023801326752, + 0.0219588465988636, + 0.01017574593424797, + 0.007156559266149998, + 0.010064443573355675, + 0.008061868138611317, + 0.015808744356036186, + 0.012775062583386898, + 0.020003998652100563, + 0.008255092427134514, + 0.011793078854680061, + 0.011042634025216103, + 0.007728857453912497, + 0.006039705593138933, + 0.02200806885957718, + 0.015645423904061317, + 0.023463834077119827, + 0.010516626760363579, + 0.007544441148638725, + 0.008951249532401562, + 0.005129052326083183, + 0.004123373422771692, + 0.012693231925368309, + 0.009905461221933365, + 0.01621483638882637, + 0.007700085639953613, + 0.004523813258856535, + 0.008677689358592033, + 0.004846458788961172, + 0.005346112884581089, + 0.010907337069511414, + 0.009424427524209023, + 0.00988028571009636, + 0.006125718355178833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.011089242994785309 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00030560590676032007 + ], + "opposite_logprob_tokens": [ + -8.12530517578125 + ], + "logprob_reward": -0.00030560590676032007, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.00030560590676032007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4517924189567566, + 0.15257729589939117, + 0.10033242404460907, + 0.08438657224178314, + 0.02086981199681759, + 0.039282627403736115, + 0.03599213436245918, + 0.03588808700442314, + 0.051446061581373215, + 0.02743254229426384 + ], + "credit_map": [ + 0.03919123485684395, + 0.02579553611576557, + 0.11701640486717224, + 0.023361636325716972, + 0.02273261919617653, + 0.1499990075826645, + 0.03074881248176098, + 0.042947206646203995, + 0.017353154718875885, + 0.018522093072533607, + 0.0125114805996418, + 0.007685086689889431, + 0.027156619355082512, + 0.0234188511967659, + 0.022617921233177185, + 0.023312093690037727, + 0.016332855448126793, + 0.0080408388748765, + 0.007915194146335125, + 0.006124362349510193, + 0.020158790051937103, + 0.014228404499590397, + 0.01732063852250576, + 0.010211337357759476, + 0.011712799780070782, + 0.009222760796546936, + 0.007687703240662813, + 0.005863236263394356, + 0.012947211042046547, + 0.014068706892430782, + 0.015338188037276268, + 0.007545970845967531, + 0.0029958472587168217, + 0.0019058316247537732, + 0.001959847053512931, + 0.0015182190109044313, + 0.0035279719159007072, + 0.0031952187418937683, + 0.00349625782109797, + 0.002270619384944439, + 0.004956560209393501, + 0.00362670561298728, + 0.003633959451690316, + 0.002717404393479228, + 0.004686966072767973, + 0.006545557640492916, + 0.00775722274556756, + 0.005358252674341202, + 0.004514636471867561, + 0.003050172235816717, + 0.003712818492203951, + 0.0028666884172707796, + 0.00715188542380929, + 0.005010984372347593, + 0.006484382785856724, + 0.0032005682587623596, + 0.004108387976884842, + 0.0035612545907497406, + 0.003926293458789587, + 0.002297737868502736, + 0.006282132118940353, + 0.005037146154791117, + 0.006650119554251432, + 0.004025017376989126, + 0.007500074803829193, + 0.007748672738671303, + 0.005240330938249826, + 0.002745475387200713, + 0.006035088561475277, + 0.007790742442011833, + 0.009200348518788815, + 0.005185327026993036, + 0.0033185346983373165, + 0.004982622340321541, + 0.0021803947165608406, + 0.0025388230569660664, + 0.002848832169547677, + 0.0045304917730391026, + 0.0038447417318820953, + 0.003188101574778557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00030560590676032007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270274817943573 + ], + "opposite_logprob_tokens": [ + -2.1270275115966797 + ], + "logprob_reward": -0.1270274817943573, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270274817943573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2507532835006714, + 0.15281829237937927, + 0.1778547167778015, + 0.08656049519777298, + 0.019449539482593536, + 0.018877791240811348, + 0.03654325753450394, + 0.055787552148103714, + 0.150959774851799, + 0.05039525404572487 + ], + "credit_map": [ + 0.02278043143451214, + 0.016426142305135727, + 0.04912504553794861, + 0.011837566271424294, + 0.02059563808143139, + 0.06776662915945053, + 0.03335335850715637, + 0.028868472203612328, + 0.017212867736816406, + 0.01579129695892334, + 0.011925438418984413, + 0.008237545378506184, + 0.02931862883269787, + 0.02278842404484749, + 0.02365218475461006, + 0.023891927674412727, + 0.02945355884730816, + 0.017302025109529495, + 0.011604638770222664, + 0.008600822649896145, + 0.03930611535906792, + 0.02645680122077465, + 0.02769351191818714, + 0.017437268048524857, + 0.011505714617669582, + 0.010762088000774384, + 0.006303735543042421, + 0.005668931640684605, + 0.015267851762473583, + 0.013005813583731651, + 0.016239779070019722, + 0.007806587498635054, + 0.0024096898268908262, + 0.0015133084962144494, + 0.002309875562787056, + 0.002022070810198784, + 0.0027982378378510475, + 0.0028199818916618824, + 0.0027870466001331806, + 0.0027893304359167814, + 0.002361979568377137, + 0.0016448647947981954, + 0.00158908951561898, + 0.0019076374592259526, + 0.0030102520249783993, + 0.0028373929671943188, + 0.0032057473435997963, + 0.0023208283819258213, + 0.004127883818000555, + 0.004222049377858639, + 0.004259348381310701, + 0.003053823485970497, + 0.005144485272467136, + 0.0054898676462471485, + 0.0051670619286596775, + 0.005078740417957306, + 0.007554886396974325, + 0.006913099437952042, + 0.0034268368035554886, + 0.0027272431179881096, + 0.01024872250854969, + 0.008692745119333267, + 0.011022751219570637, + 0.005201265681535006, + 0.019072918221354485, + 0.016193214803934097, + 0.008878232911229134, + 0.009480739012360573, + 0.025967737659811974, + 0.01995897851884365, + 0.03578859940171242, + 0.015619379468262196, + 0.005127718672156334, + 0.007889856584370136, + 0.004117812030017376, + 0.00466008298099041, + 0.0051001813262701035, + 0.009445969946682453, + 0.007921207696199417, + 0.00613242806866765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.1270274817943573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014187073335051537 + ], + "opposite_logprob_tokens": [ + -4.264186859130859 + ], + "logprob_reward": -0.014187073335051537, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014187073335051537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5455774068832397, + 0.055963821709156036, + 0.16980651021003723, + 0.04851595312356949, + 0.030855994671583176, + 0.027143407613039017, + 0.020982546731829643, + 0.019411463290452957, + 0.04955236241221428, + 0.032190535217523575 + ], + "credit_map": [ + 0.04597426950931549, + 0.027696700766682625, + 0.12833625078201294, + 0.023396918550133705, + 0.048537664115428925, + 0.17138254642486572, + 0.05203000828623772, + 0.04822302237153053, + 0.007597573101520538, + 0.007980103604495525, + 0.005487366113811731, + 0.0039027645252645016, + 0.01152469590306282, + 0.007305022794753313, + 0.006047678645700216, + 0.006118618417531252, + 0.026021933183073997, + 0.01859305426478386, + 0.02108372375369072, + 0.014344453811645508, + 0.024649728089571, + 0.02609550580382347, + 0.023698976263403893, + 0.015319114550948143, + 0.007761276327073574, + 0.0042505403980612755, + 0.003540600650012493, + 0.003260477678850293, + 0.00772785022854805, + 0.007580773439258337, + 0.008389285765588284, + 0.006005144212394953, + 0.004851414356380701, + 0.002869211370125413, + 0.002250025048851967, + 0.002147180959582329, + 0.005826673470437527, + 0.004812872502952814, + 0.0047876182943582535, + 0.0033109942451119423, + 0.0035994078498333693, + 0.0025909654796123505, + 0.0024352455511689186, + 0.0021270248107612133, + 0.0032348460517823696, + 0.004475730471313, + 0.004371272400021553, + 0.004308911971747875, + 0.002479789312928915, + 0.0031064897775650024, + 0.002331201685592532, + 0.0018809691537171602, + 0.002449317369610071, + 0.0024485941976308823, + 0.0037160557694733143, + 0.0025701289996504784, + 0.003584174672141671, + 0.004350742790848017, + 0.0015814995858818293, + 0.0009825542801991105, + 0.002110362285748124, + 0.002970372326672077, + 0.0020331796258687973, + 0.0017985746962949634, + 0.008478241972625256, + 0.010612201876938343, + 0.00475707370787859, + 0.0028921389020979404, + 0.003921904601156712, + 0.007895992137491703, + 0.006211915519088507, + 0.004782889038324356, + 0.004006864968687296, + 0.004940310027450323, + 0.002394583076238632, + 0.0026499887462705374, + 0.004226610530167818, + 0.005756103899329901, + 0.00471848901361227, + 0.003497579600661993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014187073335051537 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, captured from a slightly elevated angle, slowly pans to the right, following a simple wooden boat as it slowly plows across the broad, calm river. A bright full moon hangs high in the inky night sky, its cool glow the sole source of light, casting a shimmering silver mirror on the water. The distant mountains appear as mere rolling black silhouettes. Atop the boat's bow, a poet stands steady, clad in a flowing moon-white silk robe. His long hair is held back by a simple headband, a few strands brushing against his thin face in the cool night breeze. With one hand clasped behind his back, he gazes at the moon's reflection on the distant water, his brow etched with deep thought and a subtle hint of loneliness. The boat glides silently, the waves plowed by the bow shimmering silver in the moonlight, rippling outwards. His wide robe sleeves flutter gently in the breeze, as if to blend into the boundless night. The entire picture is dominated by deep indigo and silver-white colors, creating an atmosphere of extreme tranquility, loneliness and classical poetry.", + "seed": 19647, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_7/index_3062_seed_19647_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.033746808650903404, + "vlm_logprob": -0.033746808650903404, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.033746808650903404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030978208407759666 + ], + "opposite_logprob_tokens": [ + -3.655978202819824 + ], + "logprob_reward": -0.030978208407759666, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.030978208407759666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14024169743061066, + 0.09528832137584686, + 0.08265897631645203, + 0.04826263710856438, + 0.1243460550904274, + 0.04031849279999733, + 0.06042173132300377, + 0.0951831191778183, + 0.24189849197864532, + 0.07138039171695709 + ], + "credit_map": [ + 0.011830619536340237, + 0.01595124788582325, + 0.019214287400245667, + 0.014040755107998848, + 0.017506467178463936, + 0.018828010186553, + 0.021542776376008987, + 0.021327540278434753, + 0.007186752278357744, + 0.017558857798576355, + 0.012689827010035515, + 0.011042050085961819, + 0.010514678433537483, + 0.010148993693292141, + 0.010259732604026794, + 0.015887437388300896, + 0.0076831006444990635, + 0.007882666774094105, + 0.015821615234017372, + 0.006893889978528023, + 0.007879339158535004, + 0.013576139695942402, + 0.009878185577690601, + 0.013044050894677639, + 0.005564834922552109, + 0.010767261497676373, + 0.00852254033088684, + 0.0031063437927514315, + 0.003930166829377413, + 0.00840509682893753, + 0.0030236809980124235, + 0.00494271470233798, + 0.012297027744352818, + 0.015528109855949879, + 0.0148289380595088, + 0.012094269506633282, + 0.014694718644022942, + 0.023648101836442947, + 0.01264556311070919, + 0.018609337508678436, + 0.004260696936398745, + 0.004321561194956303, + 0.005562009289860725, + 0.003608505707234144, + 0.005891907028853893, + 0.00792795978486538, + 0.0034343055449426174, + 0.005311551503837109, + 0.00557732442393899, + 0.00793476589024067, + 0.012929007411003113, + 0.00590168870985508, + 0.005334272049367428, + 0.01042307261377573, + 0.005305163562297821, + 0.007016442250460386, + 0.010728291235864162, + 0.007390297017991543, + 0.02374907396733761, + 0.007437311112880707, + 0.008355962112545967, + 0.017350930720567703, + 0.006590786390006542, + 0.013580462895333767, + 0.026222120970487595, + 0.027868283912539482, + 0.0547618493437767, + 0.02754025161266327, + 0.0242916252464056, + 0.037309128791093826, + 0.01737605594098568, + 0.02652920037508011, + 0.006199815310537815, + 0.0069944774731993675, + 0.016603771597146988, + 0.005972953513264656, + 0.008002039976418018, + 0.008763648569583893, + 0.005562859121710062, + 0.013280833140015602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.030978208407759666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029816141352057457 + ], + "opposite_logprob_tokens": [ + -3.529816150665283 + ], + "logprob_reward": -0.029816141352057457, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029816141352057457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18134018778800964, + 0.03596958518028259, + 0.020038237795233727, + 0.02019261009991169, + 0.036026306450366974, + 0.035687997937202454, + 0.034011319279670715, + 0.01758783496916294, + 0.5268633961677551, + 0.09228245168924332 + ], + "credit_map": [ + 0.01589025743305683, + 0.020803382620215416, + 0.022797176614403725, + 0.016655314713716507, + 0.037628062069416046, + 0.021731937304139137, + 0.02168702520430088, + 0.024147046729922295, + 0.0030260970816016197, + 0.006010851822793484, + 0.004648651462048292, + 0.0032613712828606367, + 0.0045161000452935696, + 0.004510650411248207, + 0.004679729696363211, + 0.0053161378018558025, + 0.0018624486401677132, + 0.003060584655031562, + 0.0033277643378823996, + 0.0017934497445821762, + 0.0025648775044828653, + 0.0028353959787636995, + 0.00239401962608099, + 0.0021996991708874702, + 0.0025206077843904495, + 0.0022252113558351994, + 0.00433727353811264, + 0.0015485960757359862, + 0.0021545644849538803, + 0.0024921188596636057, + 0.002186643425375223, + 0.00272759678773582, + 0.003209174843505025, + 0.003384720766916871, + 0.008277572691440582, + 0.0024796274956315756, + 0.002833224833011627, + 0.008083191700279713, + 0.0019982389640063047, + 0.005760557483881712, + 0.004250426311045885, + 0.004292397294193506, + 0.005817510187625885, + 0.0028918732423335314, + 0.0049133808352053165, + 0.0058450158685445786, + 0.002653212286531925, + 0.005024183541536331, + 0.00400360906496644, + 0.007449009455740452, + 0.0049828700721263885, + 0.0026539629325270653, + 0.003918901551514864, + 0.005697523709386587, + 0.0023062515538185835, + 0.0029991946648806334, + 0.002104504033923149, + 0.002073105191811919, + 0.0024334387853741646, + 0.0017106480663642287, + 0.0018784395651891828, + 0.0027321679517626762, + 0.001958627952262759, + 0.0026969057507812977, + 0.053857553750276566, + 0.05770649388432503, + 0.11870722472667694, + 0.06580644100904465, + 0.05270717665553093, + 0.08272445201873779, + 0.040266867727041245, + 0.05508723109960556, + 0.008970302529633045, + 0.010459506884217262, + 0.012460487894713879, + 0.010126064531505108, + 0.015606457367539406, + 0.012264762073755264, + 0.009808960370719433, + 0.012585915625095367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.029816141352057457 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009230881696566939 + ], + "opposite_logprob_tokens": [ + -7.000923156738281 + ], + "logprob_reward": -0.0009230881696566939, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009230881696566939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23381786048412323, + 0.08128630369901657, + 0.26179632544517517, + 0.14224262535572052, + 0.05122918635606766, + 0.045861706137657166, + 0.05721115693449974, + 0.04362450912594795, + 0.0365171916782856, + 0.046413131058216095 + ], + "credit_map": [ + 0.02459149993956089, + 0.04528848081827164, + 0.02824094332754612, + 0.01662769541144371, + 0.037582919001579285, + 0.03336770832538605, + 0.022557755932211876, + 0.02556084282696247, + 0.008596551604568958, + 0.017851581797003746, + 0.009894948452711105, + 0.006324701476842165, + 0.010608558543026447, + 0.01047053374350071, + 0.0073984176851809025, + 0.010140999220311642, + 0.034755490720272064, + 0.09126652032136917, + 0.027711108326911926, + 0.012628121301531792, + 0.025100503116846085, + 0.030258605256676674, + 0.015211252495646477, + 0.024864699691534042, + 0.017963699996471405, + 0.035908132791519165, + 0.02448810078203678, + 0.009483874775469303, + 0.01306054275482893, + 0.01991335116326809, + 0.007377194240689278, + 0.014047719538211823, + 0.005234218668192625, + 0.009004824794828892, + 0.011410586535930634, + 0.0037100373301655054, + 0.004644285421818495, + 0.006241440307348967, + 0.002897967817261815, + 0.008085818961262703, + 0.004731394816190004, + 0.004863231908529997, + 0.00686689093708992, + 0.004921708721667528, + 0.005252407398074865, + 0.00762045755982399, + 0.004022624343633652, + 0.007582988124340773, + 0.005361356772482395, + 0.005163777153939009, + 0.01233351044356823, + 0.004761652089655399, + 0.004836613778024912, + 0.011906309984624386, + 0.0047860597260296345, + 0.008061874657869339, + 0.004620726220309734, + 0.0033930751960724592, + 0.008532238192856312, + 0.0037323583383113146, + 0.004267027601599693, + 0.00889934878796339, + 0.003986097406595945, + 0.006193635053932667, + 0.004516688175499439, + 0.004436921793967485, + 0.005027308128774166, + 0.005054808687418699, + 0.005401137750595808, + 0.0042767273262143135, + 0.003719978965818882, + 0.004083613865077496, + 0.0049691712483763695, + 0.005971691571176052, + 0.004556725732982159, + 0.005263675935566425, + 0.009608865715563297, + 0.0062364237383008, + 0.005151319783180952, + 0.004655252676457167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0009230881696566939 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028444230556488 + ], + "opposite_logprob_tokens": [ + -2.3502843379974365 + ], + "logprob_reward": -0.10028444230556488, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028444230556488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1136285737156868, + 0.07697439938783646, + 0.031065838411450386, + 0.07854730635881424, + 0.09666019678115845, + 0.05835824832320213, + 0.07451219856739044, + 0.0330607071518898, + 0.3256787359714508, + 0.11151377111673355 + ], + "credit_map": [ + 0.012170035392045975, + 0.016083505004644394, + 0.018036717548966408, + 0.007801816798746586, + 0.01759560965001583, + 0.0195817518979311, + 0.00956336036324501, + 0.012795766815543175, + 0.006518624257296324, + 0.00953982025384903, + 0.01270886417478323, + 0.00617569126188755, + 0.006644012406468391, + 0.011098633520305157, + 0.013056843541562557, + 0.01123190950602293, + 0.003248836612328887, + 0.0064643207006156445, + 0.004123018588870764, + 0.0020262952893972397, + 0.0028183795511722565, + 0.0038786535151302814, + 0.004825573414564133, + 0.0036807607393711805, + 0.0096516702324152, + 0.020567242056131363, + 0.011324345134198666, + 0.00574398273602128, + 0.008186688646674156, + 0.009051486849784851, + 0.006086023990064859, + 0.00793586578220129, + 0.010841645300388336, + 0.010092368349432945, + 0.015347804874181747, + 0.006581366062164307, + 0.009751064702868462, + 0.02415897697210312, + 0.007166391704231501, + 0.01272058766335249, + 0.006785540841519833, + 0.0057939025573432446, + 0.007631209678947926, + 0.00945632066577673, + 0.006101679056882858, + 0.00692131370306015, + 0.007464261259883642, + 0.008204021491110325, + 0.009863571263849735, + 0.01465881709009409, + 0.009039398282766342, + 0.006729801185429096, + 0.010835332795977592, + 0.009823684580624104, + 0.004522956907749176, + 0.009038632735610008, + 0.003639505710452795, + 0.003690566634759307, + 0.0058329724706709385, + 0.0038831764832139015, + 0.0037258677184581757, + 0.00477784825488925, + 0.0036037745885550976, + 0.003906995989382267, + 0.034102119505405426, + 0.03797391057014465, + 0.07235506176948547, + 0.03802896663546562, + 0.03083699941635132, + 0.05253756418824196, + 0.02309667505323887, + 0.03674744814634323, + 0.008193950168788433, + 0.010733929462730885, + 0.025939125567674637, + 0.010900426656007767, + 0.010592064820230007, + 0.027226470410823822, + 0.0061805471777915955, + 0.011747254058718681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.10028444230556488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a long shot of a simple wooden boat on a broad, calm river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a poet standing on the boat's bow wearing a flowing moon-white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a bright full moon in an inky night sky, casting a silver reflection on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly pan to the right while the boat glides silently across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the distant mountains visible as rolling black silhouettes against the night sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732163019478321 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006732163019478321, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732163019478321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12205683439970016, + 0.05786164104938507, + 0.047489315271377563, + 0.019559785723686218, + 0.31844452023506165, + 0.07848643511533737, + 0.05855373665690422, + 0.061660028994083405, + 0.10691875964403152, + 0.1289689540863037 + ], + "credit_map": [ + 0.011409441009163857, + 0.016894565895199776, + 0.02213873714208603, + 0.008477858267724514, + 0.015636447817087173, + 0.020160410553216934, + 0.012344839982688427, + 0.014994539320468903, + 0.005542234983295202, + 0.014838439412415028, + 0.006399814039468765, + 0.004815774038434029, + 0.005790175870060921, + 0.008772906847298145, + 0.005364410113543272, + 0.006337885744869709, + 0.005659988150000572, + 0.01228956039994955, + 0.0061288392171263695, + 0.0035374269355088472, + 0.004375709220767021, + 0.006885458715260029, + 0.003588874125853181, + 0.0050234561786055565, + 0.0027238945476710796, + 0.0029015287291258574, + 0.002420327626168728, + 0.0015572007978335023, + 0.0021281931549310684, + 0.0042698075994849205, + 0.0015627843094989657, + 0.0019960494246333838, + 0.03316545858979225, + 0.034915562719106674, + 0.050283946096897125, + 0.02525782771408558, + 0.03308546915650368, + 0.0748610720038414, + 0.026940956711769104, + 0.039934225380420685, + 0.007013034541159868, + 0.009507098235189915, + 0.010086542926728725, + 0.0067193759605288506, + 0.008347000926733017, + 0.015772756189107895, + 0.009495329111814499, + 0.01154529582709074, + 0.006705631967633963, + 0.012559205293655396, + 0.008263538591563702, + 0.004561161156743765, + 0.008278114721179008, + 0.008068753406405449, + 0.004591307137161493, + 0.005526023916900158, + 0.006815325003117323, + 0.0062460629269480705, + 0.012473229318857193, + 0.006931477226316929, + 0.008076823316514492, + 0.008142165839672089, + 0.005741968750953674, + 0.007232976146042347, + 0.01123617123812437, + 0.012768405489623547, + 0.022661568596959114, + 0.01262946892529726, + 0.009787785820662975, + 0.018611619248986244, + 0.007475236430764198, + 0.011748497374355793, + 0.010244348086416721, + 0.014511632733047009, + 0.02722552977502346, + 0.01039828546345234, + 0.019056901335716248, + 0.024659274145960808, + 0.010186903178691864, + 0.012686074711382389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.006732163019478321 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_8/index_tensor([7217, 7217])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_8/index_tensor([7217, 7217])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f998b46da7082ba141f74fe7cf2adcf4a82e2cda --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_8/index_tensor([7217, 7217])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "single, static, eye-level, full-length shot immerses the viewer in a scene rendered in Tim Burton's distinctive gothic style, featuring two two-spotted Elaphe bimaculata snakes performing a mesmerizing courtship dance. Their slender bodies, covered in scales that take on a delicate, dry sheen in the moonlight, are rendered against a high-contrast, monochrome palette, their natural two-spotted pattern visible as a complex arrangement of charcoal and ash-grey. They rise vertically from the ground, gracefully twisting and interweaving in a slow, hypnotic rhythm, their movements fluid and seamless. The dance unfolds against the backdrop of a gnarled, leafless tree, its skeletal branches reaching toward the night sky, silhouetted by a massive, unnaturally glowing full moon. This unique light source bathes the scene in stark highlights and deep, inky shadows that stretch and twist with the snakes' movements. A sinister thorny vine wraps tightly around the trunk, and a velvety crimson flower bursts from above, its saturated red providing a solitary, dramatic splash of color in the black and white world, creating an atmosphere that is at once terrifying, enchanting, and strangely romantic.", + "seed": 24601, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_8/index_7217_seed_24601_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08787203638348728, + "vlm_logprob": -0.08787203638348728, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.08787203638348728, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14697280526161194 + ], + "opposite_logprob_tokens": [ + -2.021972894668579 + ], + "logprob_reward": -0.14697280526161194, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14697280526161194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15137739479541779, + 0.09466412663459778, + 0.07580195367336273, + 0.2685360312461853, + 0.16605858504772186, + 0.05252545326948166, + 0.04867371916770935, + 0.05016285181045532, + 0.06048250198364258, + 0.031717389822006226 + ], + "credit_map": [ + 0.019190970808267593, + 0.018084757030010223, + 0.02143067494034767, + 0.008187542669475079, + 0.02145233191549778, + 0.024468902498483658, + 0.01617010496556759, + 0.02239210158586502, + 0.012784512713551521, + 0.009071556851267815, + 0.01366916112601757, + 0.005478923674672842, + 0.012533584609627724, + 0.01867665909230709, + 0.008594026789069176, + 0.013855707831680775, + 0.00505724037066102, + 0.004346202593296766, + 0.011273271404206753, + 0.00426758686080575, + 0.014856814406812191, + 0.017109278589487076, + 0.006906085181981325, + 0.011985477060079575, + 0.03687215596437454, + 0.05029622092843056, + 0.036896344274282455, + 0.020456980913877487, + 0.024887677282094955, + 0.03629761561751366, + 0.01692717894911766, + 0.045901864767074585, + 0.005686047021299601, + 0.006194271147251129, + 0.014492618851363659, + 0.00429402245208621, + 0.015225338749587536, + 0.05549408867955208, + 0.008421631529927254, + 0.05625057965517044, + 0.003767205635085702, + 0.004652094561606646, + 0.005588713567703962, + 0.003202007384970784, + 0.0083101075142622, + 0.010153082199394703, + 0.0069347284734249115, + 0.00991751253604889, + 0.0032863591331988573, + 0.005695991683751345, + 0.005851005669683218, + 0.0025989811401814222, + 0.005651602987200022, + 0.01095663383603096, + 0.005247413646429777, + 0.009385727345943451, + 0.0025268071331083775, + 0.0023412159644067287, + 0.007660737726837397, + 0.0016547663835808635, + 0.0049718949012458324, + 0.010856933891773224, + 0.004573268815875053, + 0.0155772240832448, + 0.0047818743623793125, + 0.005337060429155827, + 0.009881040081381798, + 0.0034258703235536814, + 0.009360438212752342, + 0.014189460314810276, + 0.005681980401277542, + 0.007824784144759178, + 0.003695881459861994, + 0.0034596468321979046, + 0.00506755243986845, + 0.0021212883293628693, + 0.006329674739390612, + 0.004971654154360294, + 0.0023952950723469257, + 0.003676392836496234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.14697280526161194 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036501940339803696 + ], + "opposite_logprob_tokens": [ + -5.628650188446045 + ], + "logprob_reward": -0.0036501940339803696, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036501940339803696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3206964135169983, + 0.16826960444450378, + 0.1442844420671463, + 0.08134045451879501, + 0.10392661392688751, + 0.055662356317043304, + 0.02144257351756096, + 0.03164435178041458, + 0.04140162095427513, + 0.03133153170347214 + ], + "credit_map": [ + 0.03701421990990639, + 0.029542140662670135, + 0.045176319777965546, + 0.02709466591477394, + 0.04538384824991226, + 0.04651707783341408, + 0.048537854105234146, + 0.041430313140153885, + 0.02419823221862316, + 0.015194406732916832, + 0.026612503454089165, + 0.015517900697886944, + 0.018953608348965645, + 0.0268348827958107, + 0.02188859134912491, + 0.019069503992795944, + 0.015210908837616444, + 0.01943192631006241, + 0.0165358018130064, + 0.010934282094240189, + 0.018214788287878036, + 0.026342622935771942, + 0.009158494882285595, + 0.02845562808215618, + 0.01080456655472517, + 0.012134122662246227, + 0.013750654645264149, + 0.007716942112892866, + 0.007819797843694687, + 0.010177474468946457, + 0.007752994075417519, + 0.0111839072778821, + 0.005015852861106396, + 0.009217658080160618, + 0.008760210126638412, + 0.007990692742168903, + 0.008440131321549416, + 0.025851869955658913, + 0.010958483442664146, + 0.027691727504134178, + 0.0028249688912183046, + 0.005040684714913368, + 0.005545896477997303, + 0.0021900543943047523, + 0.0048795281909406185, + 0.014641040936112404, + 0.005142641719430685, + 0.015397543087601662, + 0.002104828367009759, + 0.0018278710776939988, + 0.0031035130377858877, + 0.0008462602272629738, + 0.003985276445746422, + 0.003831708338111639, + 0.0023382692597806454, + 0.00340484781190753, + 0.0027379330713301897, + 0.002563748275861144, + 0.004436124581843615, + 0.0009374581859447062, + 0.004346833564341068, + 0.006383078638464212, + 0.003411875106394291, + 0.006827301345765591, + 0.00270070880651474, + 0.002424504142254591, + 0.006832854822278023, + 0.0015534783015027642, + 0.007225365843623877, + 0.0116355512291193, + 0.0023274971172213554, + 0.006701663602143526, + 0.0038930641021579504, + 0.005387333687394857, + 0.003866561222821474, + 0.0029288846999406815, + 0.004289065953344107, + 0.004732084926217794, + 0.0024578107986599207, + 0.00377672933973372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0036501940339803696 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707598507404327 + ], + "opposite_logprob_tokens": [ + -2.1270759105682373 + ], + "logprob_reward": -0.12707598507404327, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12707598507404327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20323967933654785, + 0.06296750903129578, + 0.14368869364261627, + 0.05550631880760193, + 0.23763789236545563, + 0.05994882434606552, + 0.06836428493261337, + 0.030389003455638885, + 0.12203153222799301, + 0.01622622273862362 + ], + "credit_map": [ + 0.01247017364948988, + 0.00896020233631134, + 0.036620642989873886, + 0.01138520147651434, + 0.045125383883714676, + 0.03536522760987282, + 0.026494089514017105, + 0.026818763464689255, + 0.0034655805211514235, + 0.0036048865877091885, + 0.008472497574985027, + 0.004561377223581076, + 0.011140809394419193, + 0.013873051851987839, + 0.008484414778649807, + 0.009364885278046131, + 0.019765609875321388, + 0.023462912067770958, + 0.02068771980702877, + 0.012546638026833534, + 0.012706262059509754, + 0.02017582580447197, + 0.009097114205360413, + 0.025246605277061462, + 0.005155824590474367, + 0.007388690486550331, + 0.009179678745567799, + 0.005303612444549799, + 0.00631159171462059, + 0.008423902094364166, + 0.005241818726062775, + 0.008501200005412102, + 0.010033831931650639, + 0.018818242475390434, + 0.021122027188539505, + 0.009678306989371777, + 0.012412880547344685, + 0.07668396830558777, + 0.01089224312454462, + 0.0779963955283165, + 0.003689719596877694, + 0.004192301072180271, + 0.005515292286872864, + 0.002806359436362982, + 0.006507332902401686, + 0.012863441370427608, + 0.007900775410234928, + 0.01647360250353813, + 0.005208617076277733, + 0.009225158020853996, + 0.00715186633169651, + 0.00477738119661808, + 0.010045865550637245, + 0.008736418560147285, + 0.011854268610477448, + 0.011364709585905075, + 0.002063776133581996, + 0.002123695332556963, + 0.003993222024291754, + 0.001065090880729258, + 0.004125731065869331, + 0.0069976686500012875, + 0.002197491703554988, + 0.007822328247129917, + 0.01095725130289793, + 0.01652410812675953, + 0.02123713679611683, + 0.007372803520411253, + 0.014542724937200546, + 0.02719317376613617, + 0.011882686987519264, + 0.01232164353132248, + 0.001257282099686563, + 0.002239387482404709, + 0.0021024318411946297, + 0.001120514003559947, + 0.0034528023097664118, + 0.002241076435893774, + 0.0016008198726922274, + 0.002211909042671323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.12707598507404327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013333010720089078 + ], + "opposite_logprob_tokens": [ + -6.626333236694336 + ], + "logprob_reward": -0.0013333010720089078, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013333010720089078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2229411005973816, + 0.27291354537010193, + 0.07317326217889786, + 0.12241408228874207, + 0.07006335258483887, + 0.059721723198890686, + 0.049049459397792816, + 0.03291572630405426, + 0.054462604224681854, + 0.042345140129327774 + ], + "credit_map": [ + 0.01745717227458954, + 0.026318786665797234, + 0.03527350723743439, + 0.016511207446455956, + 0.029520418494939804, + 0.04174058139324188, + 0.02465813048183918, + 0.03146129101514816, + 0.018059566617012024, + 0.0140073262155056, + 0.03689458966255188, + 0.014636739157140255, + 0.03185158595442772, + 0.08130500465631485, + 0.0230798851698637, + 0.05307886376976967, + 0.006972210947424173, + 0.0067850700579583645, + 0.012835288420319557, + 0.0039951191283762455, + 0.008496450260281563, + 0.014692751690745354, + 0.0048157707788050175, + 0.014580600894987583, + 0.016731418669223785, + 0.019572924822568893, + 0.018905000761151314, + 0.010582546703517437, + 0.012400495819747448, + 0.016292085871100426, + 0.010050781071186066, + 0.017878830432891846, + 0.0035075016785413027, + 0.004290659446269274, + 0.007919110357761383, + 0.002374076284468174, + 0.0054499306716024876, + 0.02194053865969181, + 0.0030691269785165787, + 0.021512402221560478, + 0.004540447611361742, + 0.005373033229261637, + 0.00632863212376833, + 0.00334654888138175, + 0.004387048538774252, + 0.013303832150995731, + 0.005193954799324274, + 0.017248230054974556, + 0.0025752380024641752, + 0.0035221430007368326, + 0.005255504511296749, + 0.0024114183615893126, + 0.005582619924098253, + 0.009086348116397858, + 0.008792128413915634, + 0.01182405836880207, + 0.0026343187782913446, + 0.003286059945821762, + 0.005331830121576786, + 0.001760864513926208, + 0.004493333864957094, + 0.005362787749618292, + 0.0042061819694936275, + 0.0058403476141393185, + 0.005428338889032602, + 0.008846580982208252, + 0.008394673466682434, + 0.003542814403772354, + 0.005933015141636133, + 0.008623171597719193, + 0.005655245389789343, + 0.008038769476115704, + 0.004581348970532417, + 0.0055663445964455605, + 0.006652004085481167, + 0.003581971162930131, + 0.005490569397807121, + 0.00652179354801774, + 0.004580605309456587, + 0.0053705028258264065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0013333010720089078 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032789647579193 + ], + "opposite_logprob_tokens": [ + -1.9103279113769531 + ], + "logprob_reward": -0.16032789647579193, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032789647579193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16049210727214813, + 0.08932780474424362, + 0.11752302199602127, + 0.19744932651519775, + 0.24291639029979706, + 0.03262105584144592, + 0.05316001921892166, + 0.023167066276073456, + 0.056819044053554535, + 0.026524167507886887 + ], + "credit_map": [ + 0.015718191862106323, + 0.014557302929461002, + 0.022574948146939278, + 0.008237535133957863, + 0.03382565453648567, + 0.026835603639483452, + 0.017559178173542023, + 0.021183697506785393, + 0.009089385159313679, + 0.006880324799567461, + 0.01282487902790308, + 0.0060882484540343285, + 0.013562473468482494, + 0.01822969689965248, + 0.008850593119859695, + 0.01380220241844654, + 0.01699144020676613, + 0.0171201229095459, + 0.01611889712512493, + 0.010149958543479443, + 0.009766174480319023, + 0.01711561344563961, + 0.009432220831513405, + 0.020828597247600555, + 0.03001343086361885, + 0.035080645233392715, + 0.025945942848920822, + 0.018196960911154747, + 0.018128802999854088, + 0.025609523057937622, + 0.013170624151825905, + 0.03130340948700905, + 0.008570341393351555, + 0.014647318981587887, + 0.020474359393119812, + 0.007443555165082216, + 0.011279282160103321, + 0.08731228113174438, + 0.007364134769886732, + 0.08582509309053421, + 0.0028075582813471556, + 0.0021789956372231245, + 0.005608266219496727, + 0.0013952191220596433, + 0.003516898490488529, + 0.006027155090123415, + 0.0034239483065903187, + 0.007663013879209757, + 0.003867285093292594, + 0.005818601697683334, + 0.004815797321498394, + 0.003049959894269705, + 0.004670289345085621, + 0.012800398282706738, + 0.006033075507730246, + 0.012104609981179237, + 0.00274830125272274, + 0.0022676088847219944, + 0.003086845623329282, + 0.0009793646167963743, + 0.0031898256856948137, + 0.004080366343259811, + 0.0015262182569131255, + 0.005288536194711924, + 0.005361674819141626, + 0.008481447584927082, + 0.00955353956669569, + 0.0036060279235243797, + 0.0071677216328680515, + 0.011759794317185879, + 0.004904976114630699, + 0.005983863025903702, + 0.002958160126581788, + 0.0037815433461219072, + 0.004870297387242317, + 0.001625850098207593, + 0.002598157851025462, + 0.005689541809260845, + 0.001839957432821393, + 0.0031606582924723625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.16032789647579193 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, eye-level, full-length shot immerses the viewer in a scene rendered in Tim Burton's distinctive gothic style, featuring two two-spotted Elaphe bimaculata snakes performing a mesmerizing courtship dance. Their slender bodies, covered in scales that take on a delicate, dry sheen in the moonlight, are rendered against a high-contrast, monochrome palette, their natural two-spotted pattern visible as a complex arrangement of charcoal and ash-grey. They rise vertically from the ground, gracefully twisting and interweaving in a slow, hypnotic rhythm, their movements fluid and seamless. The dance unfolds against the backdrop of a gnarled, leafless tree, its skeletal branches reaching toward the night sky, silhouetted by a massive, unnaturally glowing full moon. This unique light source bathes the scene in stark highlights and deep, inky shadows that stretch and twist with the snakes' movements. A sinister thorny vine wraps tightly around the trunk, and a velvety crimson flower bursts from above, its saturated red providing a solitary, dramatic splash of color in the black and white world, creating an atmosphere that is at once terrifying, enchanting, and strangely romantic.", + "seed": 24602, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_8/index_7217_seed_24602_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0978029135381803, + "vlm_logprob": -0.0978029135381803, + "vlm_yes_no_margin": 3.8999999880790712, + "vlm_token_credit": -0.0978029135381803, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11616238206624985 + ], + "opposite_logprob_tokens": [ + -2.2411623001098633 + ], + "logprob_reward": -0.11616238206624985, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11616238206624985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12122747302055359, + 0.037390343844890594, + 0.0400451235473156, + 0.05543806403875351, + 0.21818453073501587, + 0.2521766722202301, + 0.07207298278808594, + 0.06403154879808426, + 0.04534110426902771, + 0.09409213811159134 + ], + "credit_map": [ + 0.014024805277585983, + 0.009122752584517002, + 0.020096996799111366, + 0.01065207738429308, + 0.02055911161005497, + 0.018073735758662224, + 0.012225763872265816, + 0.01647222973406315, + 0.0037974570877850056, + 0.0034383765887469053, + 0.005333678796887398, + 0.003156270133331418, + 0.004299432039260864, + 0.0072617307305336, + 0.0028924637008458376, + 0.007210934534668922, + 0.0037541736382991076, + 0.002323961816728115, + 0.006188672035932541, + 0.0059556057676672935, + 0.005344532895833254, + 0.00522072147578001, + 0.006701210048049688, + 0.00455624470487237, + 0.005031507462263107, + 0.006300662644207478, + 0.007868015207350254, + 0.007929442450404167, + 0.0066890171729028225, + 0.005971756298094988, + 0.007168755400925875, + 0.008478911593556404, + 0.02529721148312092, + 0.013831695541739464, + 0.02515687420964241, + 0.015843398869037628, + 0.02640056610107422, + 0.040712881833314896, + 0.015422101132571697, + 0.055519819259643555, + 0.03906351327896118, + 0.015209359116852283, + 0.02848012186586857, + 0.019481956958770752, + 0.01985841430723667, + 0.04197472333908081, + 0.021458230912685394, + 0.06665035337209702, + 0.009403313510119915, + 0.005904677323997021, + 0.008503491058945656, + 0.007325287442654371, + 0.009581333957612514, + 0.010192439891397953, + 0.006472203880548477, + 0.014690236188471317, + 0.007878768257796764, + 0.004672892391681671, + 0.009304766543209553, + 0.0048787654377520084, + 0.006049816031008959, + 0.009935881942510605, + 0.004515946377068758, + 0.0167947169393301, + 0.0038055742625147104, + 0.003581265453249216, + 0.007210287265479565, + 0.00806258711963892, + 0.006033692974597216, + 0.005853449460119009, + 0.005141176749020815, + 0.005653069820255041, + 0.009536836296319962, + 0.006359742023050785, + 0.009587232954800129, + 0.006237800698727369, + 0.027226081117987633, + 0.0144179817289114, + 0.006865281146019697, + 0.013861185871064663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.11616238206624985 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001735848723910749 + ], + "opposite_logprob_tokens": [ + -6.376735687255859 + ], + "logprob_reward": -0.001735848723910749, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001735848723910749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20260711014270782, + 0.06525197625160217, + 0.06460542976856232, + 0.06496921181678772, + 0.09377957880496979, + 0.12373580783605576, + 0.10765264183282852, + 0.03178269416093826, + 0.04584898427128792, + 0.1997665911912918 + ], + "credit_map": [ + 0.022121313959360123, + 0.01642569713294506, + 0.04294664040207863, + 0.015217204578220844, + 0.015075687319040298, + 0.033088743686676025, + 0.012571015395224094, + 0.045160792768001556, + 0.006548524834215641, + 0.005904319230467081, + 0.009964230470359325, + 0.006311393808573484, + 0.006632696837186813, + 0.010517808608710766, + 0.0051971059292554855, + 0.014175894670188427, + 0.006691589951515198, + 0.003738128114491701, + 0.009606006555259228, + 0.013675551861524582, + 0.007248470559716225, + 0.008715164847671986, + 0.003614540211856365, + 0.011315979063510895, + 0.0056765517219901085, + 0.00704621197655797, + 0.01172782201319933, + 0.010254067368805408, + 0.005989299155771732, + 0.009213256649672985, + 0.0043996116146445274, + 0.010662385262548923, + 0.00996403954923153, + 0.006744449958205223, + 0.011473960243165493, + 0.009689979255199432, + 0.018669169396162033, + 0.014252597466111183, + 0.009950200095772743, + 0.013035191223025322, + 0.014349590055644512, + 0.011746613308787346, + 0.017168905586004257, + 0.011681417934596539, + 0.013388223946094513, + 0.01815422810614109, + 0.011859666556119919, + 0.025387154892086983, + 0.015333855524659157, + 0.008866795338690281, + 0.01677529886364937, + 0.008425528183579445, + 0.01169112604111433, + 0.016878372058272362, + 0.006017076782882214, + 0.023664595559239388, + 0.003891888540238142, + 0.002434580586850643, + 0.00506185507401824, + 0.0022868963424116373, + 0.004699987825006247, + 0.004558125976473093, + 0.002603117609396577, + 0.006246241740882397, + 0.004895471502095461, + 0.003681708360090852, + 0.006154883652925491, + 0.005427742376923561, + 0.006815293338149786, + 0.006634816061705351, + 0.003948765341192484, + 0.0082903066650033, + 0.018918476998806, + 0.01534038782119751, + 0.023426992818713188, + 0.01810002140700817, + 0.04861538112163544, + 0.027977745980024338, + 0.017875857651233673, + 0.029511718079447746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.001735848723910749 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589197054505348 + ], + "opposite_logprob_tokens": [ + -3.8955891132354736 + ], + "logprob_reward": -0.020589197054505348, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589197054505348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1215643435716629, + 0.10646143555641174, + 0.296276718378067, + 0.1077377200126648, + 0.05103886500000954, + 0.1256357729434967, + 0.07927504181861877, + 0.026524856686592102, + 0.053019966930150986, + 0.032465215772390366 + ], + "credit_map": [ + 0.015866633504629135, + 0.008674564771354198, + 0.019309595227241516, + 0.013786556199193, + 0.016729218885302544, + 0.016044490039348602, + 0.01410670392215252, + 0.017046580091118813, + 0.011101718991994858, + 0.010692378506064415, + 0.013690907508134842, + 0.009875399991869926, + 0.018866179510951042, + 0.0117952274158597, + 0.017612909898161888, + 0.012826718389987946, + 0.03050212934613228, + 0.01524706743657589, + 0.05027677118778229, + 0.04137066751718521, + 0.04468884319067001, + 0.03900342434644699, + 0.032482586801052094, + 0.042705237865448, + 0.010823174379765987, + 0.009859774261713028, + 0.013667586259543896, + 0.015512317419052124, + 0.012552809901535511, + 0.01520705595612526, + 0.015579788945615292, + 0.014535214751958847, + 0.004606358706951141, + 0.003587497863918543, + 0.006227564532309771, + 0.004192851483821869, + 0.004689123015850782, + 0.009752548299729824, + 0.0039146170020103455, + 0.014068303629755974, + 0.014010289683938026, + 0.00785793736577034, + 0.020027829334139824, + 0.01287878304719925, + 0.011260260827839375, + 0.01868523098528385, + 0.012711609713733196, + 0.02820383757352829, + 0.010546854697167873, + 0.005929222330451012, + 0.01570706069469452, + 0.007600428070873022, + 0.008220123127102852, + 0.011186025105416775, + 0.006158871576189995, + 0.013926463201642036, + 0.003219742327928543, + 0.0024271325673907995, + 0.003111955476924777, + 0.002976122312247753, + 0.005023642908781767, + 0.0037566719111055136, + 0.002607608214020729, + 0.003401980735361576, + 0.007216536905616522, + 0.0054153138771653175, + 0.007326439488679171, + 0.008550401777029037, + 0.006143598351627588, + 0.0071283746510744095, + 0.004552542231976986, + 0.00668675871565938, + 0.004401871003210545, + 0.0025339839048683643, + 0.004637142177671194, + 0.0021017659455537796, + 0.005253782495856285, + 0.006132496986538172, + 0.0026411900762468576, + 0.0047629824839532375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.020589197054505348 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019359909929335117 + ], + "opposite_logprob_tokens": [ + -6.251935958862305 + ], + "logprob_reward": -0.0019359909929335117, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019359909929335117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17430828511714935, + 0.22099116444587708, + 0.09450837224721909, + 0.06945719569921494, + 0.10807313770055771, + 0.09233010560274124, + 0.11719396710395813, + 0.034099895507097244, + 0.03749234974384308, + 0.051545511931180954 + ], + "credit_map": [ + 0.021740254014730453, + 0.017632726579904556, + 0.0318484790623188, + 0.022267542779445648, + 0.017847565934062004, + 0.021899156272411346, + 0.014240728691220284, + 0.02683185786008835, + 0.03106292523443699, + 0.03392883390188217, + 0.03172975033521652, + 0.019586004316806793, + 0.026593023911118507, + 0.03231453895568848, + 0.010697594843804836, + 0.03507852926850319, + 0.010157209821045399, + 0.008001964539289474, + 0.015417602844536304, + 0.020145760849118233, + 0.008087541908025742, + 0.009673802182078362, + 0.011955360881984234, + 0.011069145053625107, + 0.008369588293135166, + 0.008654567413032055, + 0.01312277652323246, + 0.009726871736347675, + 0.0055470820516347885, + 0.009852271527051926, + 0.0039022911805659533, + 0.010281750001013279, + 0.011682535521686077, + 0.006383424624800682, + 0.010912489145994186, + 0.008195425383746624, + 0.0133766308426857, + 0.0197724848985672, + 0.008043399080634117, + 0.029706761240959167, + 0.013480978086590767, + 0.006501678377389908, + 0.014887633733451366, + 0.00938540231436491, + 0.0092712277546525, + 0.012162956409156322, + 0.008772972039878368, + 0.017867272719740868, + 0.021760445088148117, + 0.007542673032730818, + 0.019417421892285347, + 0.010683497413992882, + 0.009760186076164246, + 0.018113374710083008, + 0.00630892114713788, + 0.023607458919286728, + 0.003565570805221796, + 0.0034620570950210094, + 0.005364815704524517, + 0.0021477225236594677, + 0.007049701176583767, + 0.004739553201943636, + 0.0021506326738744974, + 0.005619845353066921, + 0.00506218709051609, + 0.003707303898409009, + 0.006375647149980068, + 0.006266758311539888, + 0.0032661915756762028, + 0.005129900295287371, + 0.003499342128634453, + 0.0041850232519209385, + 0.0049528381787240505, + 0.005446244962513447, + 0.00741211324930191, + 0.005099611356854439, + 0.008594130165874958, + 0.006627532187849283, + 0.004301017615944147, + 0.00911202747374773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0019359909929335117 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.348591148853302 + ], + "opposite_logprob_tokens": [ + -1.2235910892486572 + ], + "logprob_reward": -0.348591148853302, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.348591148853302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19725099205970764, + 0.09114905446767807, + 0.10571718215942383, + 0.07963645458221436, + 0.05667915940284729, + 0.14716167747974396, + 0.15569601953029633, + 0.032182514667510986, + 0.05169576406478882, + 0.08283115923404694 + ], + "credit_map": [ + 0.02425427921116352, + 0.011965751647949219, + 0.03867151215672493, + 0.011485190130770206, + 0.018219610676169395, + 0.02966349571943283, + 0.016035892069339752, + 0.04695526137948036, + 0.011310635134577751, + 0.005277915392071009, + 0.014230538159608841, + 0.007562202401459217, + 0.015601850114762783, + 0.012040024623274803, + 0.006480866577476263, + 0.0186450257897377, + 0.011105876415967941, + 0.007512904703617096, + 0.019612610340118408, + 0.021186254918575287, + 0.01033295039087534, + 0.013503302820026875, + 0.00468882080167532, + 0.01777447573840618, + 0.009856004267930984, + 0.004180059302598238, + 0.017098788172006607, + 0.006124912295490503, + 0.012610211037099361, + 0.011017444543540478, + 0.006709916517138481, + 0.01203912403434515, + 0.0057367244735360146, + 0.006155529525130987, + 0.005816725082695484, + 0.007101710885763168, + 0.008465617895126343, + 0.007785279769450426, + 0.004865462891757488, + 0.010752110742032528, + 0.019249051809310913, + 0.010052039287984371, + 0.02035301923751831, + 0.01333613134920597, + 0.014291321858763695, + 0.022441795095801353, + 0.01469485368579626, + 0.03274347633123398, + 0.02727155201137066, + 0.011084993369877338, + 0.022276083007454872, + 0.01375519298017025, + 0.010827691294252872, + 0.024996113032102585, + 0.00833064503967762, + 0.037153761833906174, + 0.005138851702213287, + 0.0031266645528376102, + 0.004333454184234142, + 0.0037414187099784613, + 0.002363400300964713, + 0.005013265181332827, + 0.003369548125192523, + 0.005095910746604204, + 0.005199981387704611, + 0.0033085201866924763, + 0.006918782833963633, + 0.005610936786979437, + 0.007821804843842983, + 0.007434159517288208, + 0.00474521005526185, + 0.010656371712684631, + 0.011282891035079956, + 0.00973911676555872, + 0.011321025900542736, + 0.005887290928512812, + 0.0124219274148345, + 0.013681210577487946, + 0.0053850202821195126, + 0.01311267726123333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.348591148853302 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, eye-level, full-length shot immerses the viewer in a scene rendered in Tim Burton's distinctive gothic style, featuring two two-spotted Elaphe bimaculata snakes performing a mesmerizing courtship dance. Their slender bodies, covered in scales that take on a delicate, dry sheen in the moonlight, are rendered against a high-contrast, monochrome palette, their natural two-spotted pattern visible as a complex arrangement of charcoal and ash-grey. They rise vertically from the ground, gracefully twisting and interweaving in a slow, hypnotic rhythm, their movements fluid and seamless. The dance unfolds against the backdrop of a gnarled, leafless tree, its skeletal branches reaching toward the night sky, silhouetted by a massive, unnaturally glowing full moon. This unique light source bathes the scene in stark highlights and deep, inky shadows that stretch and twist with the snakes' movements. A sinister thorny vine wraps tightly around the trunk, and a velvety crimson flower bursts from above, its saturated red providing a solitary, dramatic splash of color in the black and white world, creating an atmosphere that is at once terrifying, enchanting, and strangely romantic.", + "seed": 24603, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_8/index_7217_seed_24603_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10521463942714035, + "vlm_logprob": -0.10521463942714035, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.10521463942714035, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11894308775663376 + ], + "opposite_logprob_tokens": [ + -2.243942975997925 + ], + "logprob_reward": -0.11894308775663376, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11894308775663376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35392269492149353, + 0.056498557329177856, + 0.21895278990268707, + 0.028631143271923065, + 0.09354370087385178, + 0.09351810812950134, + 0.041266731917858124, + 0.05573638528585434, + 0.027605516836047173, + 0.030324459075927734 + ], + "credit_map": [ + 0.06306964159011841, + 0.02428949996829033, + 0.052775267511606216, + 0.03223632648587227, + 0.0218346007168293, + 0.04909204691648483, + 0.02497774362564087, + 0.08564752340316772, + 0.010714293457567692, + 0.0044126263819634914, + 0.009428883902728558, + 0.004540184047073126, + 0.00423953915014863, + 0.007610850036144257, + 0.003422390203922987, + 0.01212978269904852, + 0.02687126211822033, + 0.029322637245059013, + 0.027233513072133064, + 0.04083321988582611, + 0.020938556641340256, + 0.017705397680401802, + 0.02879941463470459, + 0.02724875882267952, + 0.006198620889335871, + 0.0031190193258225918, + 0.0034337390679866076, + 0.0017695596907287836, + 0.0029539274983108044, + 0.002708697458729148, + 0.0029771344270557165, + 0.005470437463372946, + 0.022990601137280464, + 0.007991466671228409, + 0.013998541980981827, + 0.006336298771202564, + 0.008101685903966427, + 0.009877124801278114, + 0.007864703424274921, + 0.016383253037929535, + 0.011844775639474392, + 0.013340982608497143, + 0.013196642510592937, + 0.010477574542164803, + 0.013521584682166576, + 0.009431684389710426, + 0.011248945258557796, + 0.010455896146595478, + 0.007473450619727373, + 0.0046243625693023205, + 0.005840367637574673, + 0.004292021505534649, + 0.004749272018671036, + 0.005244547501206398, + 0.004589013755321503, + 0.004453686531633139, + 0.009485374204814434, + 0.005568090360611677, + 0.00771610252559185, + 0.005390470381826162, + 0.007603237871080637, + 0.006953353527933359, + 0.005208959337323904, + 0.007810786832123995, + 0.005256179254502058, + 0.0034763736184686422, + 0.0044655464589595795, + 0.0024517385754734278, + 0.0030055600218474865, + 0.0028278634417802095, + 0.0024935060646384954, + 0.003628744278103113, + 0.004682953469455242, + 0.0034048091620206833, + 0.004932056181132793, + 0.0026072002947330475, + 0.0042732092551887035, + 0.0029655154794454575, + 0.0033987299539148808, + 0.004059981554746628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.11894308775663376 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022357723210006952 + ], + "opposite_logprob_tokens": [ + -6.1272358894348145 + ], + "logprob_reward": -0.0022357723210006952, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022357723210006952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28329193592071533, + 0.0628066286444664, + 0.12435905635356903, + 0.02625812031328678, + 0.12042796611785889, + 0.08533536642789841, + 0.05533487722277641, + 0.04938347265124321, + 0.06297298520803452, + 0.12982960045337677 + ], + "credit_map": [ + 0.05801653489470482, + 0.019908318296074867, + 0.04088082164525986, + 0.027578270062804222, + 0.01804220676422119, + 0.03734002262353897, + 0.01791369356215, + 0.06361205875873566, + 0.010727480985224247, + 0.006389192771166563, + 0.007714020553976297, + 0.005683452356606722, + 0.0056082415394485, + 0.008701829239726067, + 0.0047801826149225235, + 0.013202233240008354, + 0.015828823670744896, + 0.01598932594060898, + 0.016707992181181908, + 0.02403814159333706, + 0.011372406035661697, + 0.010233634151518345, + 0.016262724995613098, + 0.013926000334322453, + 0.004006045870482922, + 0.0034336925018578768, + 0.0033987844362854958, + 0.002280103275552392, + 0.003917244728654623, + 0.0027760074008256197, + 0.0024093780666589737, + 0.004036865197122097, + 0.02621307782828808, + 0.013601289130747318, + 0.014780750498175621, + 0.009369644336402416, + 0.010428136214613914, + 0.012084576301276684, + 0.008803049102425575, + 0.025147445499897003, + 0.01742098107933998, + 0.006722635589540005, + 0.014205497689545155, + 0.0065392302349209785, + 0.005994438659399748, + 0.0114719457924366, + 0.006335066165775061, + 0.01664556935429573, + 0.008335216902196407, + 0.0061673447489738464, + 0.007589033804833889, + 0.006098967045545578, + 0.007978902198374271, + 0.006414144765585661, + 0.006812009494751692, + 0.005939256399869919, + 0.008073766715824604, + 0.00392569974064827, + 0.007495643571019173, + 0.005614208523184061, + 0.0035293083637952805, + 0.005878493655472994, + 0.007503003813326359, + 0.007363347336649895, + 0.010676081292331219, + 0.009666740894317627, + 0.006764829624444246, + 0.00859896931797266, + 0.006106824614107609, + 0.008047494105994701, + 0.006016457453370094, + 0.007095590699464083, + 0.017313512042164803, + 0.017188318073749542, + 0.014981593936681747, + 0.008547266945242882, + 0.020394492894411087, + 0.017161663621664047, + 0.016915056854486465, + 0.01732771098613739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.0022357723210006952 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031884163618088 + ], + "opposite_logprob_tokens": [ + -2.3503189086914062 + ], + "logprob_reward": -0.10031884163618088, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031884163618088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3128214478492737, + 0.11471476405858994, + 0.06758429855108261, + 0.03520674258470535, + 0.06813277304172516, + 0.06925781816244125, + 0.03699410334229469, + 0.1948201209306717, + 0.05344301462173462, + 0.047024913132190704 + ], + "credit_map": [ + 0.06117736175656319, + 0.027795689180493355, + 0.03944770619273186, + 0.0177853312343359, + 0.018694011494517326, + 0.0408439114689827, + 0.023898465558886528, + 0.08317898213863373, + 0.017517520114779472, + 0.009881258010864258, + 0.01631024479866028, + 0.010006656870245934, + 0.009854219853878021, + 0.019141899421811104, + 0.007353388238698244, + 0.02464957721531391, + 0.011404765769839287, + 0.01000957190990448, + 0.008199958130717278, + 0.008802050724625587, + 0.006051046773791313, + 0.0065830466337502, + 0.006191427819430828, + 0.010342433117330074, + 0.00705503486096859, + 0.0025145909748971462, + 0.0065062278881669044, + 0.0027231890708208084, + 0.003699458669871092, + 0.004214848857372999, + 0.003187298309057951, + 0.005306093487888575, + 0.013013848103582859, + 0.00608415063470602, + 0.010429814457893372, + 0.004258642438799143, + 0.007616149727255106, + 0.006593108177185059, + 0.007700480055063963, + 0.012436578050255775, + 0.012525228783488274, + 0.007911420427262783, + 0.01280217058956623, + 0.006422876380383968, + 0.005666021257638931, + 0.009344235062599182, + 0.004794950596988201, + 0.009790919721126556, + 0.005572216585278511, + 0.0045515806414186954, + 0.005336295813322067, + 0.0033118054270744324, + 0.003798216814175248, + 0.0050612748600542545, + 0.004323067143559456, + 0.0050396486185491085, + 0.029617352411150932, + 0.018482541665434837, + 0.02453439310193062, + 0.019741442054510117, + 0.029964134097099304, + 0.023418687283992767, + 0.022617684677243233, + 0.026443878188729286, + 0.008187714964151382, + 0.007111857179552317, + 0.007762512657791376, + 0.0048027653247118, + 0.005563357844948769, + 0.0080979960039258, + 0.004980371333658695, + 0.00693643931299448, + 0.005097721237689257, + 0.005200861021876335, + 0.006262642331421375, + 0.00512205483391881, + 0.006140395533293486, + 0.004825703799724579, + 0.007589212618768215, + 0.006786323618143797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.10031884163618088 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023274563252925873 + ], + "opposite_logprob_tokens": [ + -3.7732746601104736 + ], + "logprob_reward": -0.023274563252925873, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023274563252925873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.513378381729126, + 0.17937707901000977, + 0.0325016975402832, + 0.019536178559064865, + 0.052672404795885086, + 0.04497896134853363, + 0.018375452607870102, + 0.04763336852192879, + 0.05591128394007683, + 0.035635221749544144 + ], + "credit_map": [ + 0.09881914407014847, + 0.03601748123764992, + 0.07404043525457382, + 0.055113375186920166, + 0.03681831806898117, + 0.06716328859329224, + 0.029649285599589348, + 0.11575702577829361, + 0.038750164210796356, + 0.015139346942305565, + 0.029247719794511795, + 0.011937608011066914, + 0.011651978828012943, + 0.023013683035969734, + 0.012623006477952003, + 0.03701357543468475, + 0.007310893852263689, + 0.0026131656486541033, + 0.005752671044319868, + 0.0022123761009424925, + 0.0019295428646728396, + 0.0039828624576330185, + 0.001654352992773056, + 0.007045830134302378, + 0.004257280379533768, + 0.0015729611041024327, + 0.0028670080937445164, + 0.0010737788397818804, + 0.0017001113155856729, + 0.001959299435839057, + 0.0018903057789430022, + 0.00421543326228857, + 0.012646365910768509, + 0.003388947807252407, + 0.006092538591474295, + 0.0031283001881092787, + 0.0034411149099469185, + 0.005981098860502243, + 0.004031695891171694, + 0.013962343335151672, + 0.009460872039198875, + 0.004951925482600927, + 0.005553210619837046, + 0.0035073088947683573, + 0.0042139580473303795, + 0.005339825060218573, + 0.004115678835660219, + 0.007836181670427322, + 0.002573103178292513, + 0.001532830181531608, + 0.002511975122615695, + 0.0017118395771831274, + 0.0027514719404280186, + 0.001997640123590827, + 0.002288324758410454, + 0.0030082666780799627, + 0.006039733998477459, + 0.009178598411381245, + 0.0066884104162454605, + 0.003786437911912799, + 0.005112944636493921, + 0.004993997514247894, + 0.004491637460887432, + 0.007341605611145496, + 0.006865226197987795, + 0.005225019995123148, + 0.0064333039335906506, + 0.005636898800730705, + 0.007107524201273918, + 0.011688240803778172, + 0.005946320481598377, + 0.007008751854300499, + 0.004382536746561527, + 0.004064298700541258, + 0.004303504712879658, + 0.002556808525696397, + 0.005745220929384232, + 0.0045487890020012856, + 0.005554755683988333, + 0.004479308612644672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.023274563252925873 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28130093216896057 + ], + "opposite_logprob_tokens": [ + -1.4063009023666382 + ], + "logprob_reward": -0.28130093216896057, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28130093216896057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3071609139442444, + 0.05984865874052048, + 0.09545361250638962, + 0.055120863020420074, + 0.13754738867282867, + 0.06409461051225662, + 0.043572161346673965, + 0.05960289016366005, + 0.10537448525428772, + 0.07222440838813782 + ], + "credit_map": [ + 0.05819939821958542, + 0.030632812529802322, + 0.04104456678032875, + 0.027211926877498627, + 0.029525019228458405, + 0.0353233739733696, + 0.027980074286460876, + 0.057243745774030685, + 0.013754223473370075, + 0.0048362379893660545, + 0.00803677923977375, + 0.0070167360827326775, + 0.005517488811165094, + 0.005868829321116209, + 0.00609468761831522, + 0.008723679929971695, + 0.026012415066361427, + 0.008505151607096195, + 0.01959097571671009, + 0.005323860328644514, + 0.006435910705476999, + 0.009172865189611912, + 0.005972547456622124, + 0.014439892955124378, + 0.011499411426484585, + 0.0073012057691812515, + 0.006780371069908142, + 0.00391922565177083, + 0.004131466615945101, + 0.005930517800152302, + 0.004394278395920992, + 0.011164391413331032, + 0.027259765192866325, + 0.015152276493608952, + 0.02005227841436863, + 0.00980764627456665, + 0.009412233717739582, + 0.015246636234223843, + 0.012687738053500652, + 0.027928819879889488, + 0.009126010350883007, + 0.006449040956795216, + 0.010703025385737419, + 0.008352428674697876, + 0.00651682261377573, + 0.007750892546027899, + 0.006938629783689976, + 0.00825776718556881, + 0.010952235199511051, + 0.004485076759010553, + 0.00634064432233572, + 0.003613437293097377, + 0.0036826778668910265, + 0.0056822639890015125, + 0.004233529791235924, + 0.004582300316542387, + 0.007867049425840378, + 0.010675939731299877, + 0.006595142185688019, + 0.00642382213845849, + 0.008809437043964863, + 0.006779734510928392, + 0.006576643791049719, + 0.005875119008123875, + 0.013770604506134987, + 0.010787262581288815, + 0.012182511389255524, + 0.011207951232790947, + 0.011353506706655025, + 0.019805071875452995, + 0.012298821471631527, + 0.013968762010335922, + 0.008408722467720509, + 0.008575952611863613, + 0.008033574558794498, + 0.005575746297836304, + 0.010774091817438602, + 0.013003703206777573, + 0.010722245089709759, + 0.007130375597625971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.28130093216896057 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, eye-level, full-length shot immerses the viewer in a scene rendered in Tim Burton's distinctive gothic style, featuring two two-spotted Elaphe bimaculata snakes performing a mesmerizing courtship dance. Their slender bodies, covered in scales that take on a delicate, dry sheen in the moonlight, are rendered against a high-contrast, monochrome palette, their natural two-spotted pattern visible as a complex arrangement of charcoal and ash-grey. They rise vertically from the ground, gracefully twisting and interweaving in a slow, hypnotic rhythm, their movements fluid and seamless. The dance unfolds against the backdrop of a gnarled, leafless tree, its skeletal branches reaching toward the night sky, silhouetted by a massive, unnaturally glowing full moon. This unique light source bathes the scene in stark highlights and deep, inky shadows that stretch and twist with the snakes' movements. A sinister thorny vine wraps tightly around the trunk, and a velvety crimson flower bursts from above, its saturated red providing a solitary, dramatic splash of color in the black and white world, creating an atmosphere that is at once terrifying, enchanting, and strangely romantic.", + "seed": 24604, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_8/index_7217_seed_24604_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13532102655153722, + "vlm_logprob": -0.13532102655153722, + "vlm_yes_no_margin": 3.5249999880790712, + "vlm_token_credit": -0.13532102655153722, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18499477207660675 + ], + "opposite_logprob_tokens": [ + -1.8099948167800903 + ], + "logprob_reward": -0.18499477207660675, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18499477207660675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20622952282428741, + 0.08134674280881882, + 0.16247773170471191, + 0.08391761034727097, + 0.0962589681148529, + 0.03130985051393509, + 0.082644023001194, + 0.15008282661437988, + 0.07073178142309189, + 0.03500095382332802 + ], + "credit_map": [ + 0.03908304497599602, + 0.018311632797122, + 0.04179125279188156, + 0.01733994670212269, + 0.016321541741490364, + 0.025914069265127182, + 0.018424805253744125, + 0.02904323861002922, + 0.013924437575042248, + 0.007498838938772678, + 0.016355471685528755, + 0.006046263966709375, + 0.00547341862693429, + 0.012072883546352386, + 0.006821943446993828, + 0.01315348595380783, + 0.0205196775496006, + 0.009599177166819572, + 0.022815074771642685, + 0.01946154795587063, + 0.017805904150009155, + 0.024460799992084503, + 0.027259349822998047, + 0.02055620774626732, + 0.008731482550501823, + 0.00676933815702796, + 0.01107074972242117, + 0.009203154593706131, + 0.007726968266069889, + 0.013750198297202587, + 0.012754113413393497, + 0.013911603949964046, + 0.014845729805529118, + 0.008969257585704327, + 0.019928526133298874, + 0.008820937015116215, + 0.010821023024618626, + 0.009771754965186119, + 0.009480117820203304, + 0.013621628284454346, + 0.004516489338129759, + 0.0035721969325095415, + 0.0044018919579684734, + 0.004827706143260002, + 0.0037992680445313454, + 0.004525728989392519, + 0.002606005407869816, + 0.0030605620704591274, + 0.012188974767923355, + 0.006208606995642185, + 0.014224641025066376, + 0.01128638256341219, + 0.009062062948942184, + 0.009829558432102203, + 0.008210388943552971, + 0.011633402667939663, + 0.021341806277632713, + 0.012569839134812355, + 0.024974720552563667, + 0.017536267638206482, + 0.01602928340435028, + 0.016508355736732483, + 0.02290620096027851, + 0.018216341733932495, + 0.006574262399226427, + 0.006359898019582033, + 0.00941515527665615, + 0.007908985018730164, + 0.00719005661085248, + 0.00843022856861353, + 0.010983497835695744, + 0.013869700953364372, + 0.004736631643027067, + 0.003298126859590411, + 0.00575842522084713, + 0.0030395579524338245, + 0.001896233530715108, + 0.007038915064185858, + 0.002993932692334056, + 0.006239129230380058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.18499477207660675 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000666277133859694 + ], + "opposite_logprob_tokens": [ + -7.37566614151001 + ], + "logprob_reward": -0.000666277133859694, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.000666277133859694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16420578956604004, + 0.09969651699066162, + 0.13460946083068848, + 0.05438326671719551, + 0.0577259324491024, + 0.100611612200737, + 0.05918446555733681, + 0.22503189742565155, + 0.027026904746890068, + 0.07752418518066406 + ], + "credit_map": [ + 0.024239683523774147, + 0.025249605998396873, + 0.03178755193948746, + 0.010785660706460476, + 0.011750512756407261, + 0.027474841102957726, + 0.012869070284068584, + 0.020048869773745537, + 0.016303682699799538, + 0.015163329429924488, + 0.016818108037114143, + 0.006803278811275959, + 0.006942210253328085, + 0.01643764227628708, + 0.008345433510839939, + 0.012882823124527931, + 0.014105264097452164, + 0.009657126851379871, + 0.016810676082968712, + 0.014741909690201283, + 0.014223354868590832, + 0.024770965799689293, + 0.02493642456829548, + 0.01536372397094965, + 0.006455629598349333, + 0.005624570418149233, + 0.00831524096429348, + 0.005308661609888077, + 0.004905883222818375, + 0.008556235581636429, + 0.006396898999810219, + 0.008820141665637493, + 0.007940234616398811, + 0.00971578061580658, + 0.00906017329543829, + 0.0056004938669502735, + 0.005556388292461634, + 0.006602678447961807, + 0.0055633485317230225, + 0.0076868305914103985, + 0.012560315430164337, + 0.007817350327968597, + 0.013701253570616245, + 0.006852114573121071, + 0.008852575905621052, + 0.0300479456782341, + 0.0076479786075651646, + 0.013132075779139996, + 0.007836365140974522, + 0.004225482232868671, + 0.00928453542292118, + 0.008923547342419624, + 0.005524090491235256, + 0.00801762379705906, + 0.007337546907365322, + 0.008035269565880299, + 0.030794966965913773, + 0.04288838058710098, + 0.04216532036662102, + 0.019051652401685715, + 0.01702861301600933, + 0.027470547705888748, + 0.020615698769688606, + 0.025016702711582184, + 0.0042891185730695724, + 0.004324661567807198, + 0.005045648198574781, + 0.0023146974854171276, + 0.0018669294659048319, + 0.003084566444158554, + 0.0019898004829883575, + 0.004111481364816427, + 0.012367118149995804, + 0.009533939883112907, + 0.014831031672656536, + 0.0097352284938097, + 0.004482904449105263, + 0.010656235739588737, + 0.005699320696294308, + 0.010218409821391106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.000666277133859694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031291097402573 + ], + "opposite_logprob_tokens": [ + -2.3503129482269287 + ], + "logprob_reward": -0.10031291097402573, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031291097402573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17003001272678375, + 0.0629335567355156, + 0.49325457215309143, + 0.07832231372594833, + 0.021883171051740646, + 0.03481197729706764, + 0.03633596748113632, + 0.04902021959424019, + 0.0317627415060997, + 0.021645469591021538 + ], + "credit_map": [ + 0.020460017025470734, + 0.014162445440888405, + 0.01963217370212078, + 0.012883917428553104, + 0.01645413599908352, + 0.04090510308742523, + 0.022224241867661476, + 0.023307975381612778, + 0.007168357260525227, + 0.004612849093973637, + 0.008393964730203152, + 0.00747961550951004, + 0.005683593917638063, + 0.00964006595313549, + 0.010632460936903954, + 0.009322650730609894, + 0.049378082156181335, + 0.0353328101336956, + 0.061296507716178894, + 0.04730121046304703, + 0.051057495176792145, + 0.1006777286529541, + 0.09369571506977081, + 0.05451502650976181, + 0.00855102576315403, + 0.006430169567465782, + 0.009018227458000183, + 0.007783078122884035, + 0.008233609609305859, + 0.014400403946638107, + 0.012082244269549847, + 0.011823553591966629, + 0.0031345526222139597, + 0.002417774870991707, + 0.0036710668355226517, + 0.002088032430037856, + 0.00173013168387115, + 0.0031516908202320337, + 0.002577861538156867, + 0.003112059785053134, + 0.004192742053419352, + 0.0020961565896868706, + 0.005309241358190775, + 0.0025024665519595146, + 0.0017418653005734086, + 0.011830129660665989, + 0.0021196892485022545, + 0.005019683390855789, + 0.0048415325582027435, + 0.002643215237185359, + 0.005605854094028473, + 0.005776315927505493, + 0.003346747951582074, + 0.004268985707312822, + 0.004465346224606037, + 0.0053879693150520325, + 0.0065382663160562515, + 0.005852671805769205, + 0.008296970278024673, + 0.004803610499948263, + 0.006666268687695265, + 0.005889071151614189, + 0.005409372970461845, + 0.00556399067863822, + 0.004512219689786434, + 0.0023999838158488274, + 0.005290312226861715, + 0.0036311240401118994, + 0.003077117260545492, + 0.003056521760299802, + 0.004652353934943676, + 0.005143107380717993, + 0.0034365595784038305, + 0.0018421480199322104, + 0.004371360410004854, + 0.002466317033395171, + 0.0014880563830956817, + 0.0026239166036248207, + 0.0021681843791157007, + 0.0032489285804331303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.10031291097402573 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036098097916692495 + ], + "opposite_logprob_tokens": [ + -5.628609657287598 + ], + "logprob_reward": -0.0036098097916692495, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036098097916692495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13726550340652466, + 0.0432933047413826, + 0.13833101093769073, + 0.05103722959756851, + 0.06967081129550934, + 0.1498139351606369, + 0.10271045565605164, + 0.1576814502477646, + 0.02669624052941799, + 0.12350009381771088 + ], + "credit_map": [ + 0.023107241839170456, + 0.01735054701566696, + 0.025347432121634483, + 0.008968857116997242, + 0.009638351388275623, + 0.026004916056990623, + 0.010157198645174503, + 0.01669096015393734, + 0.006835287902504206, + 0.0041642445139586926, + 0.008018952794373035, + 0.0036384472623467445, + 0.004127100575715303, + 0.005749745760113001, + 0.004522227216511965, + 0.006237298250198364, + 0.014109139330685139, + 0.013433265499770641, + 0.01851634681224823, + 0.015093902125954628, + 0.015576663427054882, + 0.022501936182379723, + 0.024826733395457268, + 0.014273026026785374, + 0.006109260953962803, + 0.006371784955263138, + 0.009398138150572777, + 0.0051123108714818954, + 0.003504084888845682, + 0.007300050929188728, + 0.006131226662546396, + 0.0071103754453361034, + 0.01123916544020176, + 0.005401090253144503, + 0.013728437945246696, + 0.006250323727726936, + 0.006830970756709576, + 0.008948945440351963, + 0.007382010109722614, + 0.009889867156744003, + 0.016642065718770027, + 0.009471332654356956, + 0.019773242995142937, + 0.01331335213035345, + 0.011348603293299675, + 0.04088788852095604, + 0.013455778360366821, + 0.024921676144003868, + 0.011938541196286678, + 0.009351437911391258, + 0.01248966995626688, + 0.01066870242357254, + 0.012986451387405396, + 0.014603634364902973, + 0.011273623444139957, + 0.019398394972085953, + 0.022192824631929398, + 0.01353404950350523, + 0.028864935040473938, + 0.018252793699502945, + 0.019897745922207832, + 0.017249541357159615, + 0.020083868876099586, + 0.017605701461434364, + 0.004395594820380211, + 0.0018271530279889703, + 0.004862289410084486, + 0.0029591505881398916, + 0.0022434887941926718, + 0.0026095169596374035, + 0.0034127747640013695, + 0.0043862732127308846, + 0.018022697418928146, + 0.01115353312343359, + 0.02268383279442787, + 0.01854393631219864, + 0.008550974540412426, + 0.017120562493801117, + 0.010592855513095856, + 0.01683169975876808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.0036098097916692495 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two two-spotted Elaphe bimaculata snakes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in a high-contrast monochrome palette with a single saturated red flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the snakes rise vertically from the ground and twist in a slow, hypnotic rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gnarled, leafless tree silhouetted by a massive glowing full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level, full-length view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38702136278152466 + ], + "opposite_logprob_tokens": [ + -1.1370213031768799 + ], + "logprob_reward": -0.38702136278152466, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38702136278152466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24299398064613342, + 0.07233739644289017, + 0.13831660151481628, + 0.16155624389648438, + 0.18868304789066315, + 0.049317482858896255, + 0.025968214496970177, + 0.06323279440402985, + 0.029783541336655617, + 0.027810703963041306 + ], + "credit_map": [ + 0.04264075681567192, + 0.020127786323428154, + 0.04358026757836342, + 0.021740276366472244, + 0.021048197522759438, + 0.032465435564517975, + 0.028274785727262497, + 0.03311648219823837, + 0.01178023125976324, + 0.009828847832977772, + 0.012354499660432339, + 0.007657298818230629, + 0.00574377179145813, + 0.009116851724684238, + 0.00832307431846857, + 0.007532824762165546, + 0.02065245620906353, + 0.013805910013616085, + 0.021099144592881203, + 0.012485706247389317, + 0.010889695957303047, + 0.022986654192209244, + 0.013831762596964836, + 0.022565284743905067, + 0.019534459337592125, + 0.01651904359459877, + 0.033926788717508316, + 0.017886510118842125, + 0.010573034174740314, + 0.022820094600319862, + 0.01772385649383068, + 0.02257245033979416, + 0.03009982779622078, + 0.01609065756201744, + 0.037879060953855515, + 0.019653964787721634, + 0.022301213815808296, + 0.0178094319999218, + 0.020447926595807076, + 0.02440096065402031, + 0.00619917968288064, + 0.003341496456414461, + 0.005610552616417408, + 0.0042463731952011585, + 0.003878738498315215, + 0.014187549240887165, + 0.004788780119270086, + 0.007064811885356903, + 0.0033331881277263165, + 0.004035273566842079, + 0.0037712231278419495, + 0.002558971056714654, + 0.0019618654623627663, + 0.0038289616350084543, + 0.0018748719012364745, + 0.004603859968483448, + 0.006656634155660868, + 0.008931394666433334, + 0.007027774583548307, + 0.005660598166286945, + 0.00905749574303627, + 0.007305232807993889, + 0.010015166364610195, + 0.008578500710427761, + 0.003966546151787043, + 0.0027121829334646463, + 0.004573226906359196, + 0.0030191910918802023, + 0.002547563286498189, + 0.004538679029792547, + 0.003895956091582775, + 0.004530194215476513, + 0.004237976856529713, + 0.002309744479134679, + 0.005122727248817682, + 0.0025323077570647, + 0.002027868526056409, + 0.00463784858584404, + 0.002535142470151186, + 0.0044070882722735405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.38702136278152466 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_9/index_tensor([4731, 4731])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_9/index_tensor([4731, 4731])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8a5e116d8366afc47883400964bdf771bb78f64d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_9/index_tensor([4731, 4731])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide shot, captured from afar, slowly pulls out to reveal a grander panorama. A towering, rugged, snow-capped mountain rises beneath a clear, cold sky. Its thick, untouched snow reflects a dazzling glare in the bright sunlight, casting deep, sharp shadows across the folds of rock. Suddenly, a large section of the snow slope high on the mountain's side breaks away without warning, triggering a massive avalanche. Huge blocks of ice, like shattered glaciers, mix with tons of powdery, dry snow, forming a torrent of white water that cascades down with thunderous force. Massive blocks of snow collide and shatter as they fall, stirring up clouds of snow and fog that obscure the sky. This unstoppable force rushes towards the valley floor at an astonishing speed. The entire scene is filled with a sense of nature's raw, destructive power, creating an atmosphere of tension both majestic and awe-inspiring.", + "seed": 11800, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_9/index_4731_seed_11800_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.18811955963028595, + "vlm_logprob": -0.18811955963028595, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.18811955963028595, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03739487752318382 + ], + "opposite_logprob_tokens": [ + -3.7873947620391846 + ], + "logprob_reward": -0.03739487752318382, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03739487752318382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12618379294872284, + 0.0651530921459198, + 0.06481517851352692, + 0.12633992731571198, + 0.08151695132255554, + 0.03947443887591362, + 0.04131148383021355, + 0.19159336388111115, + 0.0727403312921524, + 0.1908714324235916 + ], + "credit_map": [ + 0.020777558907866478, + 0.021548083052039146, + 0.016570383682847023, + 0.01804954744875431, + 0.01342325285077095, + 0.011741152964532375, + 0.011679062619805336, + 0.012394743040204048, + 0.009681446477770805, + 0.010551695711910725, + 0.0074259620159864426, + 0.01030725333839655, + 0.005817272700369358, + 0.005085428711026907, + 0.01014852523803711, + 0.0061355032958090305, + 0.0071801915764808655, + 0.008619445376098156, + 0.007427581585943699, + 0.009489267133176327, + 0.008396392688155174, + 0.005843073595315218, + 0.011122463271021843, + 0.0067367576994001865, + 0.01210799440741539, + 0.02135358564555645, + 0.018047822639346123, + 0.018773159012198448, + 0.01279981154948473, + 0.011654616333544254, + 0.018732290714979172, + 0.012870650738477707, + 0.006893244571983814, + 0.0092998081818223, + 0.010464821942150593, + 0.011240416206419468, + 0.01140638254582882, + 0.007713015656918287, + 0.013749861158430576, + 0.010749398730695248, + 0.005102637689560652, + 0.00398899894207716, + 0.005482807289808989, + 0.004217101726680994, + 0.004943012725561857, + 0.0045608459040522575, + 0.006221325136721134, + 0.004957710858434439, + 0.004671724978834391, + 0.006269782781600952, + 0.006375827360898256, + 0.003678506938740611, + 0.003805906279012561, + 0.004577552899718285, + 0.006382277701050043, + 0.005549903493374586, + 0.028875786811113358, + 0.039383258670568466, + 0.01713080331683159, + 0.024380475282669067, + 0.028162797912955284, + 0.01698162779211998, + 0.02137552574276924, + 0.015303096733987331, + 0.006384612061083317, + 0.020934613421559334, + 0.007933259941637516, + 0.011057715862989426, + 0.00711078941822052, + 0.00561077194288373, + 0.007908823899924755, + 0.005799741949886084, + 0.007994831539690495, + 0.013880190439522266, + 0.016481982544064522, + 0.06985381990671158, + 0.008424872532486916, + 0.0077670058235526085, + 0.05296654254198074, + 0.013502184301614761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.03739487752318382 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983947038650513 + ], + "opposite_logprob_tokens": [ + -0.5233947038650513 + ], + "logprob_reward": -0.8983947038650513, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983947038650513, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05360710248351097, + 0.055513110011816025, + 0.02751513570547104, + 0.03360335901379585, + 0.2757999300956726, + 0.059541620314121246, + 0.16037549078464508, + 0.1436314433813095, + 0.08728477358818054, + 0.10312805324792862 + ], + "credit_map": [ + 0.006574478466063738, + 0.007400067523121834, + 0.00733156455680728, + 0.006522000301629305, + 0.007064700126647949, + 0.0071484483778476715, + 0.006948502268642187, + 0.004617334809154272, + 0.007852167822420597, + 0.008118218742311, + 0.007530283182859421, + 0.008434426970779896, + 0.0050199744291603565, + 0.005710754543542862, + 0.006239992566406727, + 0.006607286632061005, + 0.0018109114607796073, + 0.003630954073742032, + 0.0036572141107171774, + 0.0036251754499971867, + 0.0029337566811591387, + 0.0019996888004243374, + 0.006979844998568296, + 0.0028775851242244244, + 0.0032856552861630917, + 0.005751848686486483, + 0.004706494975835085, + 0.004157858435064554, + 0.003767244750633836, + 0.0026325888466089964, + 0.005758386105298996, + 0.0035432777367532253, + 0.011377429589629173, + 0.01621178537607193, + 0.019229890778660774, + 0.11661355942487717, + 0.007046869024634361, + 0.010681655257940292, + 0.01942451484501362, + 0.0752141997218132, + 0.0065629975870251656, + 0.01107999961823225, + 0.005950182676315308, + 0.00746932253241539, + 0.00667275907471776, + 0.008340240456163883, + 0.006065932102501392, + 0.007400180213153362, + 0.011640935204923153, + 0.016476662829518318, + 0.009849677793681622, + 0.00917830504477024, + 0.009302451275289059, + 0.009945685975253582, + 0.04769095033407211, + 0.0462908074259758, + 0.02223975583910942, + 0.030912045389413834, + 0.014599831774830818, + 0.01569857820868492, + 0.019054755568504333, + 0.013732531107962132, + 0.015913376584649086, + 0.01148055400699377, + 0.006645485293120146, + 0.01948091760277748, + 0.010513490997254848, + 0.01219208911061287, + 0.00902579352259636, + 0.00864651519805193, + 0.011918234638869762, + 0.008862237446010113, + 0.0066283405758440495, + 0.010452738963067532, + 0.0109799113124609, + 0.032196979969739914, + 0.005829656962305307, + 0.006181149743497372, + 0.01491856761276722, + 0.015940694138407707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.8983947038650513 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005701346672140062 + ], + "opposite_logprob_tokens": [ + -7.500570297241211 + ], + "logprob_reward": -0.0005701346672140062, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005701346672140062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13648441433906555, + 0.2194880098104477, + 0.05429632216691971, + 0.04358598217368126, + 0.12450026720762253, + 0.06582635641098022, + 0.07186225056648254, + 0.14251628518104553, + 0.06383032351732254, + 0.07760989665985107 + ], + "credit_map": [ + 0.020866533741354942, + 0.02051495388150215, + 0.018483152613043785, + 0.015672888606786728, + 0.011832945980131626, + 0.014984345063567162, + 0.017470333725214005, + 0.016659246757626534, + 0.03577001392841339, + 0.04073043540120125, + 0.02381119504570961, + 0.032073065638542175, + 0.016497807577252388, + 0.018976807594299316, + 0.028484538197517395, + 0.023144131526350975, + 0.004622492007911205, + 0.008159924298524857, + 0.008095752447843552, + 0.007957656867802143, + 0.0060590049251914024, + 0.0042432453483343124, + 0.009685982950031757, + 0.00547226145863533, + 0.004392124712467194, + 0.005802066996693611, + 0.005976407788693905, + 0.006169117987155914, + 0.005090599879622459, + 0.0035655044484883547, + 0.008524168282747269, + 0.004065989051014185, + 0.005731108132749796, + 0.007561338599771261, + 0.009955953806638718, + 0.04531582072377205, + 0.006468271836638451, + 0.006258305627852678, + 0.010304059833288193, + 0.032905396074056625, + 0.011893922463059425, + 0.008762149140238762, + 0.008607028052210808, + 0.010262143798172474, + 0.00474183913320303, + 0.007210419047623873, + 0.007440182380378246, + 0.0069086672738194466, + 0.009347476065158844, + 0.008569275960326195, + 0.006761768367141485, + 0.008087093010544777, + 0.005108924582600594, + 0.006158843636512756, + 0.014751107431948185, + 0.013077755458652973, + 0.024128545075654984, + 0.037592437118291855, + 0.01018378883600235, + 0.016758272424340248, + 0.0144323306158185, + 0.011232953518629074, + 0.01676744595170021, + 0.011420493945479393, + 0.0059815216809511185, + 0.011335987597703934, + 0.0068307542242109776, + 0.009239808656275272, + 0.0046013593673706055, + 0.005198562052100897, + 0.014702171087265015, + 0.005940151400864124, + 0.003949033096432686, + 0.006265371572226286, + 0.008548018522560596, + 0.022904546931385994, + 0.005670856684446335, + 0.004943443927913904, + 0.01906379498541355, + 0.00626482255756855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.0005701346672140062 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193043638020754 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.003193043638020754, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193043638020754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048926956951618195, + 0.043193425983190536, + 0.044572487473487854, + 0.047723039984703064, + 0.04952346161007881, + 0.041974037885665894, + 0.10654070973396301, + 0.29701119661331177, + 0.1794232428073883, + 0.14111141860485077 + ], + "credit_map": [ + 0.0054440065287053585, + 0.005648890044540167, + 0.007662342861294746, + 0.007465336471796036, + 0.0044508567079901695, + 0.006924511399120092, + 0.0062774065881967545, + 0.00505360821262002, + 0.007868112064898014, + 0.006892426870763302, + 0.005253689829260111, + 0.004769749939441681, + 0.0031457121949642897, + 0.004953860305249691, + 0.004599909298121929, + 0.005709968041628599, + 0.004910201299935579, + 0.005982234608381987, + 0.006108599714934826, + 0.007621340453624725, + 0.004810771904885769, + 0.0038737680297344923, + 0.0061233495362102985, + 0.005142219830304384, + 0.004825028590857983, + 0.0053930943831801414, + 0.005422302521765232, + 0.005880814976990223, + 0.005979356821626425, + 0.005754940677434206, + 0.009556337259709835, + 0.004911164287477732, + 0.0036319035571068525, + 0.0038852510042488575, + 0.0065643819980323315, + 0.013576801866292953, + 0.0033185684587806463, + 0.003375293454155326, + 0.006480441428720951, + 0.00869082286953926, + 0.004636729136109352, + 0.005680315662175417, + 0.005248746834695339, + 0.005371536128222942, + 0.005882122553884983, + 0.004076039418578148, + 0.006102392449975014, + 0.004976159427314997, + 0.017642557621002197, + 0.01481255516409874, + 0.013241805136203766, + 0.012723534367978573, + 0.009564883075654507, + 0.009766117669641972, + 0.014619911089539528, + 0.01416934747248888, + 0.04633975401520729, + 0.0737118348479271, + 0.025508787482976913, + 0.03442554548382759, + 0.03488077223300934, + 0.024274783208966255, + 0.03389568254351616, + 0.023974036797881126, + 0.016934357583522797, + 0.023144733160734177, + 0.023655764758586884, + 0.028117740526795387, + 0.019106201827526093, + 0.020209355279803276, + 0.033109571784734726, + 0.015145518817007542, + 0.010243233293294907, + 0.017974860966205597, + 0.01270780898630619, + 0.040445346385240555, + 0.010744513012468815, + 0.006914282217621803, + 0.027462918311357498, + 0.014618465676903725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.003193043638020754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010450384579598904 + ], + "opposite_logprob_tokens": [ + -6.876045227050781 + ], + "logprob_reward": -0.0010450384579598904, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010450384579598904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05687406659126282, + 0.023359421640634537, + 0.04920751973986626, + 0.03044932521879673, + 0.07011660933494568, + 0.044922880828380585, + 0.04897686839103699, + 0.4314097464084625, + 0.12350840121507645, + 0.12117515504360199 + ], + "credit_map": [ + 0.004982086829841137, + 0.006773558910936117, + 0.007752883713692427, + 0.00601149769499898, + 0.006201034877449274, + 0.009309818036854267, + 0.008145966567099094, + 0.007697217632085085, + 0.003959943540394306, + 0.0038701018784195185, + 0.002166426507756114, + 0.0029597494285553694, + 0.0018137877341359854, + 0.00268536782823503, + 0.002929527312517166, + 0.0029745171777904034, + 0.006102594081312418, + 0.00714171351864934, + 0.007746686227619648, + 0.0077341762371361256, + 0.004065110348165035, + 0.003968001343309879, + 0.007024076301604509, + 0.005425162613391876, + 0.0032196459360420704, + 0.003946823999285698, + 0.0032742416951805353, + 0.0041100503876805305, + 0.00370286894030869, + 0.004180542193353176, + 0.004918309394270182, + 0.0030968429055064917, + 0.003885174635797739, + 0.004598782863467932, + 0.008368880487978458, + 0.022571707144379616, + 0.003543264465406537, + 0.004055873025208712, + 0.007574609015136957, + 0.015518318861722946, + 0.00625755125656724, + 0.004900561645627022, + 0.00529265683144331, + 0.006072510499507189, + 0.0037779593840241432, + 0.006152465473860502, + 0.0073081934824585915, + 0.005160984583199024, + 0.007451103068888187, + 0.005238386802375317, + 0.006174948066473007, + 0.00570880388841033, + 0.004288730211555958, + 0.005087296478450298, + 0.007345851510763168, + 0.007681745104491711, + 0.07267041504383087, + 0.10840839892625809, + 0.03449888527393341, + 0.0456773079931736, + 0.051232367753982544, + 0.035586144775152206, + 0.04822595790028572, + 0.0351102389395237, + 0.01040811836719513, + 0.013941803947091103, + 0.013483602553606033, + 0.022226762026548386, + 0.011463075876235962, + 0.011802134104073048, + 0.028040487319231033, + 0.012142427265644073, + 0.01121088583022356, + 0.015335091389715672, + 0.016319245100021362, + 0.02566896751523018, + 0.011721102520823479, + 0.007767403498291969, + 0.02130930870771408, + 0.011843143031001091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.0010450384579598904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide shot, captured from afar, slowly pulls out to reveal a grander panorama. A towering, rugged, snow-capped mountain rises beneath a clear, cold sky. Its thick, untouched snow reflects a dazzling glare in the bright sunlight, casting deep, sharp shadows across the folds of rock. Suddenly, a large section of the snow slope high on the mountain's side breaks away without warning, triggering a massive avalanche. Huge blocks of ice, like shattered glaciers, mix with tons of powdery, dry snow, forming a torrent of white water that cascades down with thunderous force. Massive blocks of snow collide and shatter as they fall, stirring up clouds of snow and fog that obscure the sky. This unstoppable force rushes towards the valley floor at an astonishing speed. The entire scene is filled with a sense of nature's raw, destructive power, creating an atmosphere of tension both majestic and awe-inspiring.", + "seed": 11801, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_9/index_4731_seed_11801_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21718825742718764, + "vlm_logprob": -0.21718825742718764, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.21718825742718764, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02652115747332573 + ], + "opposite_logprob_tokens": [ + -3.9015212059020996 + ], + "logprob_reward": -0.02652115747332573, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02652115747332573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26189306378364563, + 0.08317260444164276, + 0.14190910756587982, + 0.09927360713481903, + 0.07791821658611298, + 0.04508371278643608, + 0.09847026318311691, + 0.11179543286561966, + 0.028669806197285652, + 0.05181419476866722 + ], + "credit_map": [ + 0.049750905483961105, + 0.036053504794836044, + 0.02755410224199295, + 0.029154684394598007, + 0.019959989935159683, + 0.028050502762198448, + 0.03183712437748909, + 0.03953223675489426, + 0.011501062661409378, + 0.01328662782907486, + 0.008997813798487186, + 0.009141828864812851, + 0.012148676440119743, + 0.00939245242625475, + 0.009398892521858215, + 0.009305253624916077, + 0.01733105070888996, + 0.02862611599266529, + 0.01134707685559988, + 0.02001960575580597, + 0.0177167821675539, + 0.01269539538770914, + 0.017549650743603706, + 0.016623441129922867, + 0.010469290427863598, + 0.016841111704707146, + 0.01092119887471199, + 0.019475771114230156, + 0.010996202938258648, + 0.009735156781971455, + 0.011640498414635658, + 0.00919437501579523, + 0.006582866422832012, + 0.010028334334492683, + 0.008938883431255817, + 0.011881581507623196, + 0.010801048949360847, + 0.00906930211931467, + 0.011287220753729343, + 0.009328985586762428, + 0.0057242754846811295, + 0.003198159858584404, + 0.012629927136003971, + 0.004793714266270399, + 0.0039390986785292625, + 0.004956572316586971, + 0.006271235644817352, + 0.0035707298666238785, + 0.007516822777688503, + 0.005252423696219921, + 0.0067100669257342815, + 0.009727093391120434, + 0.005059553775936365, + 0.005124060902744532, + 0.03205479308962822, + 0.027025442570447922, + 0.008087963797152042, + 0.011866316199302673, + 0.0212198905646801, + 0.018651043996214867, + 0.00936372485011816, + 0.00738444272428751, + 0.018383976072072983, + 0.0168380718678236, + 0.0027846896555274725, + 0.0027070646174252033, + 0.0036663368809968233, + 0.005370419006794691, + 0.002212188672274351, + 0.0021645422093570232, + 0.004062121734023094, + 0.005702443420886993, + 0.004032619297504425, + 0.00922087300568819, + 0.006379734259098768, + 0.009638221934437752, + 0.0038897208869457245, + 0.0045034694485366344, + 0.006365047302097082, + 0.00778450770303607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.02652115747332573 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539579391479492 + ], + "opposite_logprob_tokens": [ + -0.4289579391479492 + ], + "logprob_reward": -1.0539579391479492, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539579391479492, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1536521166563034, + 0.050078798085451126, + 0.13592855632305145, + 0.13001514971256256, + 0.15789395570755005, + 0.08890340477228165, + 0.059846144169569016, + 0.06105748564004898, + 0.1077854335308075, + 0.054838914424180984 + ], + "credit_map": [ + 0.022268882021307945, + 0.030601628124713898, + 0.015490548685193062, + 0.01229336392134428, + 0.017298607155680656, + 0.010987701825797558, + 0.017653100192546844, + 0.027058297768235207, + 0.005508370231837034, + 0.005303438752889633, + 0.005397769156843424, + 0.007840504869818687, + 0.006639193277806044, + 0.007928070612251759, + 0.006007777526974678, + 0.005453673657029867, + 0.018301723524928093, + 0.026357833296060562, + 0.016603868454694748, + 0.016168244183063507, + 0.020054016262292862, + 0.010160685516893864, + 0.015369139611721039, + 0.012913066893815994, + 0.014621746726334095, + 0.016673117876052856, + 0.019016582518815994, + 0.02474554441869259, + 0.013298526406288147, + 0.012504680082201958, + 0.01446282584220171, + 0.014692115597426891, + 0.02040492184460163, + 0.011012780480086803, + 0.008760969154536724, + 0.020147792994976044, + 0.01234105508774519, + 0.010882798582315445, + 0.030411694198846817, + 0.04393193870782852, + 0.009079583920538425, + 0.00862764474004507, + 0.014048569835722446, + 0.008427509106695652, + 0.010241829790174961, + 0.006966826040297747, + 0.009965426288545132, + 0.0215460117906332, + 0.0059751784428954124, + 0.006014026701450348, + 0.009471035562455654, + 0.010263997130095959, + 0.005069553852081299, + 0.0036088244523853064, + 0.011203236877918243, + 0.008240291848778725, + 0.00666598416864872, + 0.007073062006384134, + 0.00874315109103918, + 0.01086357980966568, + 0.005659050308167934, + 0.005270756781101227, + 0.009079167619347572, + 0.0077027324587106705, + 0.008945773355662823, + 0.016233213245868683, + 0.015686802566051483, + 0.016136378049850464, + 0.007632959168404341, + 0.0112485745921731, + 0.011693890206515789, + 0.0202078428119421, + 0.004866852890700102, + 0.007433128077536821, + 0.007121119182556868, + 0.010899766348302364, + 0.0034376836847513914, + 0.004929914604872465, + 0.006632647942751646, + 0.009517804719507694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -1.0539579391479492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004450284468475729 + ], + "opposite_logprob_tokens": [ + -7.7504448890686035 + ], + "logprob_reward": -0.0004450284468475729, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004450284468475729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26999831199645996, + 0.11811074614524841, + 0.09009147435426712, + 0.09520693868398666, + 0.030845005065202713, + 0.07619911432266235, + 0.047497209161520004, + 0.09778115898370743, + 0.07901826500892639, + 0.09525179862976074 + ], + "credit_map": [ + 0.05747075751423836, + 0.0412244014441967, + 0.021778544411063194, + 0.029135987162590027, + 0.02022540383040905, + 0.025993511080741882, + 0.03095386177301407, + 0.04321581870317459, + 0.0201732125133276, + 0.030056601390242577, + 0.008336243219673634, + 0.014568007551133633, + 0.00896359235048294, + 0.011127177625894547, + 0.01211122889071703, + 0.012774672359228134, + 0.011080561205744743, + 0.016897937282919884, + 0.01062034908682108, + 0.010593516752123833, + 0.012001379393041134, + 0.008883490227162838, + 0.010087305679917336, + 0.009926926344633102, + 0.010490959510207176, + 0.017578095197677612, + 0.011580966413021088, + 0.013731847517192364, + 0.012167596258223057, + 0.009174540638923645, + 0.010487379506230354, + 0.009995546191930771, + 0.005733222234994173, + 0.003073639702051878, + 0.0020742921624332666, + 0.003580364165827632, + 0.0031201138626784086, + 0.002078958787024021, + 0.003972247242927551, + 0.007212162483483553, + 0.014868729747831821, + 0.005959100089967251, + 0.00620658602565527, + 0.009661578573286533, + 0.007590976543724537, + 0.006872270721942186, + 0.013773687183856964, + 0.011266179382801056, + 0.005396885797381401, + 0.0040155984461307526, + 0.005297319497913122, + 0.007097558118402958, + 0.0038058492355048656, + 0.003328760154545307, + 0.00941886380314827, + 0.009136371314525604, + 0.008512158878147602, + 0.008531083352863789, + 0.010767476633191109, + 0.013761875219643116, + 0.007834417745471, + 0.006127324886620045, + 0.020114600658416748, + 0.02213222347199917, + 0.006518966518342495, + 0.010038204491138458, + 0.010112649761140347, + 0.01330555696040392, + 0.005374545697122812, + 0.008148622699081898, + 0.010260159149765968, + 0.01525955181568861, + 0.010538330301642418, + 0.00872827973216772, + 0.008636352606117725, + 0.013590087182819843, + 0.009392015635967255, + 0.01141531951725483, + 0.012546686455607414, + 0.020404724404215813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0004450284468475729 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409490754827857 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.00409490754827857, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409490754827857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049129631370306015, + 0.046953652054071426, + 0.069907046854496, + 0.11764000356197357, + 0.05960347503423691, + 0.05479101464152336, + 0.10447412729263306, + 0.23452128469944, + 0.11601526290178299, + 0.14696453511714935 + ], + "credit_map": [ + 0.007471929769963026, + 0.008098862133920193, + 0.004827872850000858, + 0.005154806189239025, + 0.006328579504042864, + 0.005508022848516703, + 0.005277086049318314, + 0.006462472025305033, + 0.006037405226379633, + 0.007803862914443016, + 0.004597880877554417, + 0.008248860947787762, + 0.005493452772498131, + 0.006614676211029291, + 0.0039051780477166176, + 0.004252335987985134, + 0.007549700792878866, + 0.009522094391286373, + 0.009086930193006992, + 0.011534600518643856, + 0.007304221857339144, + 0.009233220480382442, + 0.007314675021916628, + 0.008361593820154667, + 0.013646661303937435, + 0.017557401210069656, + 0.01504887267947197, + 0.015298943966627121, + 0.018292410299181938, + 0.012062707915902138, + 0.011911197565495968, + 0.013821807689964771, + 0.010642713867127895, + 0.004554960876703262, + 0.0037907534278929234, + 0.00543157709762454, + 0.0029685823246836662, + 0.0031177399214357138, + 0.009576682932674885, + 0.019520463421940804, + 0.009973074309527874, + 0.004039769060909748, + 0.004488710314035416, + 0.007014500442892313, + 0.00383611093275249, + 0.005114115308970213, + 0.007770248223096132, + 0.012554485350847244, + 0.006769711151719093, + 0.006396142300218344, + 0.00599515438079834, + 0.009355513378977776, + 0.004816056229174137, + 0.004334212746471167, + 0.03574499487876892, + 0.031062332913279533, + 0.013755124993622303, + 0.01934638060629368, + 0.029259145259857178, + 0.03618044778704643, + 0.015594839118421078, + 0.014995916746556759, + 0.05211739242076874, + 0.05327202379703522, + 0.010896683670580387, + 0.013811982236802578, + 0.015970923006534576, + 0.016381099820137024, + 0.007390320301055908, + 0.011595776304602623, + 0.01585453748703003, + 0.024113932624459267, + 0.015267896465957165, + 0.016154183074831963, + 0.015484950505197048, + 0.019857104867696762, + 0.013868517242372036, + 0.018352847546339035, + 0.01806640625, + 0.029912639409303665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.00409490754827857 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009222545195370913 + ], + "opposite_logprob_tokens": [ + -7.000922203063965 + ], + "logprob_reward": -0.0009222545195370913, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009222545195370913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07248933613300323, + 0.04124761372804642, + 0.08788108825683594, + 0.12286930531263351, + 0.16748905181884766, + 0.03594491258263588, + 0.038842905312776566, + 0.23771196603775024, + 0.1092543974518776, + 0.08626938611268997 + ], + "credit_map": [ + 0.012083600275218487, + 0.00956897996366024, + 0.006495283916592598, + 0.007127856370061636, + 0.00816596020013094, + 0.00745284091681242, + 0.008599125780165195, + 0.012995690107345581, + 0.006050961557775736, + 0.006979879457503557, + 0.003230730537325144, + 0.006768197752535343, + 0.006517479661852121, + 0.005368881858885288, + 0.0031516519375145435, + 0.0031798318959772587, + 0.011536113917827606, + 0.01925954595208168, + 0.010024720802903175, + 0.010196710005402565, + 0.012035751715302467, + 0.007556918542832136, + 0.009465387091040611, + 0.007805942557752132, + 0.014202936552464962, + 0.020450981333851814, + 0.014481979422271252, + 0.016703827306628227, + 0.022238953039050102, + 0.010424491949379444, + 0.013107351958751678, + 0.011258778162300587, + 0.026862258091568947, + 0.016724547371268272, + 0.010336312465369701, + 0.020725026726722717, + 0.020473334938287735, + 0.016634954139590263, + 0.0342717207968235, + 0.021460888907313347, + 0.005085192620754242, + 0.003238878445699811, + 0.0042014652863144875, + 0.004862469621002674, + 0.0036449134349823, + 0.0053249248303473, + 0.004853252787142992, + 0.0047338190488517284, + 0.004311143886297941, + 0.0042501408606767654, + 0.005012904293835163, + 0.006892082281410694, + 0.0033623019699007273, + 0.004038895945996046, + 0.006495519541203976, + 0.004479916766285896, + 0.016722457483410835, + 0.02164570428431034, + 0.02264687605202198, + 0.03689676523208618, + 0.017031067982316017, + 0.015195195563137531, + 0.054531313478946686, + 0.05304257571697235, + 0.010999834164977074, + 0.012069305405020714, + 0.014324123971164227, + 0.020553426817059517, + 0.008844779804348946, + 0.009541471488773823, + 0.016906864941120148, + 0.016014590859413147, + 0.009430940262973309, + 0.007667230442166328, + 0.009207462891936302, + 0.013083712197840214, + 0.007503781467676163, + 0.010401740670204163, + 0.011392704211175442, + 0.017581813037395477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.0009222545195370913 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide shot, captured from afar, slowly pulls out to reveal a grander panorama. A towering, rugged, snow-capped mountain rises beneath a clear, cold sky. Its thick, untouched snow reflects a dazzling glare in the bright sunlight, casting deep, sharp shadows across the folds of rock. Suddenly, a large section of the snow slope high on the mountain's side breaks away without warning, triggering a massive avalanche. Huge blocks of ice, like shattered glaciers, mix with tons of powdery, dry snow, forming a torrent of white water that cascades down with thunderous force. Massive blocks of snow collide and shatter as they fall, stirring up clouds of snow and fog that obscure the sky. This unstoppable force rushes towards the valley floor at an astonishing speed. The entire scene is filled with a sense of nature's raw, destructive power, creating an atmosphere of tension both majestic and awe-inspiring.", + "seed": 11802, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_9/index_4731_seed_11802_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10511996492277831, + "vlm_logprob": -0.10511996492277831, + "vlm_yes_no_margin": 4.149999952316284, + "vlm_token_credit": -0.10511996492277831, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02919784002006054 + ], + "opposite_logprob_tokens": [ + -3.779197931289673 + ], + "logprob_reward": -0.02919784002006054, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02919784002006054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5049146413803101, + 0.03509040176868439, + 0.013595850206911564, + 0.010946226306259632, + 0.25414973497390747, + 0.03600991144776344, + 0.06032334640622139, + 0.04528132826089859, + 0.026093684136867523, + 0.013594867661595345 + ], + "credit_map": [ + 0.010049999691545963, + 0.008521631360054016, + 0.011448487639427185, + 0.009210611693561077, + 0.20002180337905884, + 0.20505160093307495, + 0.018377436324954033, + 0.04223308712244034, + 0.001556531758978963, + 0.0027451710775494576, + 0.0045045227743685246, + 0.005201986990869045, + 0.005349196959286928, + 0.0040063015185296535, + 0.007947205565869808, + 0.003779485821723938, + 0.0010005487129092216, + 0.0017382053192704916, + 0.0014074656646698713, + 0.001574280089698732, + 0.0021005687303841114, + 0.0022185349371284246, + 0.0015288767172023654, + 0.0020273705013096333, + 0.0010134514886885881, + 0.0014221169985830784, + 0.0015680844662711024, + 0.0014219044242054224, + 0.0015332783805206418, + 0.001079654903151095, + 0.0013597606448456645, + 0.0015479751164093614, + 0.020573841407895088, + 0.016203071922063828, + 0.018968379124999046, + 0.03004561923444271, + 0.01768527925014496, + 0.04412832483649254, + 0.04408419877290726, + 0.06246102228760719, + 0.00199018488638103, + 0.002273675287142396, + 0.0017530296463519335, + 0.0025810447987169027, + 0.0017088240711018443, + 0.0021808429155498743, + 0.01110460888594389, + 0.012417702935636044, + 0.0031082730274647474, + 0.0031615793704986572, + 0.00309879332780838, + 0.002113129012286663, + 0.0033465789165347815, + 0.003911397885531187, + 0.021030042320489883, + 0.020553557202219963, + 0.0020045293495059013, + 0.002107868669554591, + 0.0025404817424714565, + 0.002490633400157094, + 0.004045021720230579, + 0.007646626327186823, + 0.01612163707613945, + 0.008324528113007545, + 0.0021880753338336945, + 0.0021310942247509956, + 0.0025480843614786863, + 0.002067611552774906, + 0.002668979810550809, + 0.003358821151778102, + 0.005030808504670858, + 0.006100207567214966, + 0.0010176494251936674, + 0.0016307724872604012, + 0.0013690633932128549, + 0.0018382795387879014, + 0.0016270261257886887, + 0.002729636849835515, + 0.001237891148775816, + 0.0021445483434945345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.02919784002006054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47430551052093506 + ], + "opposite_logprob_tokens": [ + -0.9743055105209351 + ], + "logprob_reward": -0.47430551052093506, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47430551052093506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5698349475860596, + 0.08591802418231964, + 0.00813298113644123, + 0.04079296439886093, + 0.05012653395533562, + 0.04965369403362274, + 0.02371377684175968, + 0.050752196460962296, + 0.04372211918234825, + 0.07735270261764526 + ], + "credit_map": [ + 0.01170000247657299, + 0.01014635618776083, + 0.012769533321261406, + 0.009653554297983646, + 0.23870070278644562, + 0.21813806891441345, + 0.020414646714925766, + 0.04831212759017944, + 0.005267740692943335, + 0.007223180029541254, + 0.009351124987006187, + 0.011381814256310463, + 0.010006626136600971, + 0.010797839611768723, + 0.018209068104624748, + 0.013680639676749706, + 0.0005943273426964879, + 0.0008503178833052516, + 0.0011728063691407442, + 0.0009814182994887233, + 0.0009669858263805509, + 0.0013513609301298857, + 0.001203283085487783, + 0.0010124820983037353, + 0.0012003399897366762, + 0.001812704373151064, + 0.0018154154531657696, + 0.001715026912279427, + 0.002542185364291072, + 0.0022685190197080374, + 0.014510742388665676, + 0.01492803543806076, + 0.004407365340739489, + 0.0033124368637800217, + 0.003896754700690508, + 0.005894667934626341, + 0.003512816270813346, + 0.008147257380187511, + 0.008562118746340275, + 0.012393121607601643, + 0.0035596690140664577, + 0.002086360240355134, + 0.002004201989620924, + 0.0025423364713788033, + 0.0020687966607511044, + 0.0021889880299568176, + 0.01740119978785515, + 0.017802143469452858, + 0.002976388670504093, + 0.002274395665153861, + 0.0023340682964771986, + 0.00156666140537709, + 0.0022586053237318993, + 0.002013284480199218, + 0.0050024897791445255, + 0.0052878838032484055, + 0.0023335968144237995, + 0.002258918946608901, + 0.0025203621480613947, + 0.0021951620001345873, + 0.0036845565773546696, + 0.010104457847774029, + 0.019576232880353928, + 0.008078915067017078, + 0.0025280711706727743, + 0.0033507721964269876, + 0.003797945100814104, + 0.00391386216506362, + 0.007308416999876499, + 0.008830957114696503, + 0.009431104175746441, + 0.004560993984341621, + 0.0069716330617666245, + 0.009856490418314934, + 0.012322675436735153, + 0.010887747630476952, + 0.0105774886906147, + 0.011270743794739246, + 0.006577163469046354, + 0.008888759650290012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.47430551052093506 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002203538315370679 + ], + "opposite_logprob_tokens": [ + -6.127203464508057 + ], + "logprob_reward": -0.002203538315370679, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002203538315370679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4332740008831024, + 0.1294337809085846, + 0.044278569519519806, + 0.041608281433582306, + 0.1331111192703247, + 0.024679476395249367, + 0.032072123140096664, + 0.04917890951037407, + 0.10339007526636124, + 0.008973627351224422 + ], + "credit_map": [ + 0.008847287856042385, + 0.007541601546108723, + 0.009800883010029793, + 0.0068581970408558846, + 0.17970234155654907, + 0.16961078345775604, + 0.015268987976014614, + 0.03564390540122986, + 0.009792394004762173, + 0.009091776795685291, + 0.013752532191574574, + 0.010186217725276947, + 0.015372922644019127, + 0.013391518034040928, + 0.039604805409908295, + 0.018241601064801216, + 0.0035514605697244406, + 0.0037460471503436565, + 0.00690001854673028, + 0.004741949960589409, + 0.004869985859841108, + 0.006867777090519667, + 0.006848933175206184, + 0.006752397865056992, + 0.0016588610596954823, + 0.001546113402582705, + 0.00247769383713603, + 0.0021896602120250463, + 0.002837838139384985, + 0.0030546041671186686, + 0.01337479893118143, + 0.014468716457486153, + 0.009916538372635841, + 0.009988254867494106, + 0.013031932525336742, + 0.015386631712317467, + 0.011868514120578766, + 0.01946103200316429, + 0.02283688634634018, + 0.03062131442129612, + 0.0009944532066583633, + 0.0019623758271336555, + 0.0008986270404420793, + 0.001327416393905878, + 0.0012503803009167314, + 0.0014928276650607586, + 0.0085237892344594, + 0.008229606784880161, + 0.0031887146178632975, + 0.0016188127920031548, + 0.0019300708081573248, + 0.00131130195222795, + 0.0015695812180638313, + 0.0022427851799875498, + 0.010588189586997032, + 0.00962266605347395, + 0.0017050672322511673, + 0.0023458716459572315, + 0.0017206198535859585, + 0.002081241924315691, + 0.005297759082168341, + 0.01784553937613964, + 0.013574209064245224, + 0.004608600400388241, + 0.004369806032627821, + 0.00611083721742034, + 0.006947988644242287, + 0.008672882802784443, + 0.019574252888560295, + 0.022844554856419563, + 0.023161672055721283, + 0.011708082631230354, + 0.0006500307354144752, + 0.0008441169047728181, + 0.0006553844432346523, + 0.0011760513298213482, + 0.0009768817108124495, + 0.002098189201205969, + 0.0009549466776661575, + 0.0016180268721655011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.002203538315370679 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179578706622124 + ], + "opposite_logprob_tokens": [ + -4.018179416656494 + ], + "logprob_reward": -0.018179578706622124, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018179578706622124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20678934454917908, + 0.03591771051287651, + 0.03483551740646362, + 0.3203023672103882, + 0.19291025400161743, + 0.0294059868901968, + 0.058164291083812714, + 0.06326586753129959, + 0.03683991730213165, + 0.02156866528093815 + ], + "credit_map": [ + 0.00474777864292264, + 0.005209102761000395, + 0.01108707394450903, + 0.00635719345882535, + 0.08215360343456268, + 0.06314718723297119, + 0.01434686966240406, + 0.019740549847483635, + 0.00196782941929996, + 0.0022324775345623493, + 0.003851019311696291, + 0.003589814994484186, + 0.0043877470307052135, + 0.005615995265543461, + 0.0064873057417571545, + 0.007785526569932699, + 0.002487989142537117, + 0.004071372095495462, + 0.003132520243525505, + 0.0036953347735106945, + 0.005035176407545805, + 0.007388948928564787, + 0.003851792775094509, + 0.005172388628125191, + 0.008749357424676418, + 0.008470302447676659, + 0.012472805567085743, + 0.012251893989741802, + 0.018803900107741356, + 0.02003389783203602, + 0.11564698070287704, + 0.1238732561469078, + 0.015726638957858086, + 0.013656350784003735, + 0.01664799079298973, + 0.02130597084760666, + 0.015390141867101192, + 0.03074195794761181, + 0.032857127487659454, + 0.04658409580588341, + 0.004157742019742727, + 0.00234862114302814, + 0.002505662152543664, + 0.0030535038094967604, + 0.002167758299037814, + 0.002241267589852214, + 0.006536779925227165, + 0.006394657772034407, + 0.002679862780496478, + 0.0028869626112282276, + 0.002026431728154421, + 0.0019241294357925653, + 0.0029251333326101303, + 0.0035636390093714, + 0.021336635574698448, + 0.020821498706936836, + 0.0019075297750532627, + 0.002448043553158641, + 0.0026823042426258326, + 0.0026229352224618196, + 0.004141729790717363, + 0.023163912817835808, + 0.02340245619416237, + 0.002896965481340885, + 0.001903684576973319, + 0.0023202416487038136, + 0.0022700082045048475, + 0.002444853074848652, + 0.005680727772414684, + 0.009958767332136631, + 0.005480035208165646, + 0.006781601347029209, + 0.0019533056765794754, + 0.0033008938189595938, + 0.0024580524768680334, + 0.002488983329385519, + 0.002350222785025835, + 0.0038591763004660606, + 0.001768125337548554, + 0.0033899086993187666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.018179578706622124 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017133570509031415 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017133570509031415, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017133570509031415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45704159140586853, + 0.04294424504041672, + 0.02397022768855095, + 0.10364113003015518, + 0.06959661096334457, + 0.020711852237582207, + 0.06766755878925323, + 0.11578287184238434, + 0.07715683430433273, + 0.021487034857273102 + ], + "credit_map": [ + 0.009686190634965897, + 0.00879183690994978, + 0.011507600545883179, + 0.00872041191905737, + 0.19768938422203064, + 0.14289146661758423, + 0.0173521526157856, + 0.060402531176805496, + 0.0030043034348636866, + 0.004242038819938898, + 0.005501243751496077, + 0.004593594465404749, + 0.0071500567719340324, + 0.007055748254060745, + 0.004711474757641554, + 0.006685786880552769, + 0.0017236418789252639, + 0.0023852831218391657, + 0.003019843716174364, + 0.0028100844938308, + 0.0025538115296512842, + 0.004611053038388491, + 0.003641747636720538, + 0.003224761690944433, + 0.003583131358027458, + 0.003176417900249362, + 0.004910129122436047, + 0.004271849058568478, + 0.006351931020617485, + 0.007105727214366198, + 0.03557519242167473, + 0.038666754961013794, + 0.005494769662618637, + 0.006050129886716604, + 0.009161775931715965, + 0.008747223764657974, + 0.006183495279401541, + 0.008172513917088509, + 0.01109952013939619, + 0.014687186107039452, + 0.0034966643434017897, + 0.002266485011205077, + 0.0023620198480784893, + 0.0027568854857236147, + 0.0022220583632588387, + 0.002179235452786088, + 0.0027429971378296614, + 0.002685506595298648, + 0.0064643011428415775, + 0.0042224000208079815, + 0.003547778818756342, + 0.002658104756847024, + 0.0037130911368876696, + 0.00509678153321147, + 0.02123480662703514, + 0.020730290561914444, + 0.0035693906247615814, + 0.004519691225141287, + 0.00518083618953824, + 0.004419215489178896, + 0.006538317538797855, + 0.03822685778141022, + 0.045427221804857254, + 0.007901346310973167, + 0.004136618226766586, + 0.00488484650850296, + 0.006242212373763323, + 0.00661502406001091, + 0.011505190283060074, + 0.014526637271046638, + 0.015458828769624233, + 0.013787475414574146, + 0.0016300370916724205, + 0.0032511046156287193, + 0.0028258664533495903, + 0.0027735591866075993, + 0.0029893743339926004, + 0.0031288210302591324, + 0.002701970050111413, + 0.002186301164329052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0017133570509031415 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide shot, captured from afar, slowly pulls out to reveal a grander panorama. A towering, rugged, snow-capped mountain rises beneath a clear, cold sky. Its thick, untouched snow reflects a dazzling glare in the bright sunlight, casting deep, sharp shadows across the folds of rock. Suddenly, a large section of the snow slope high on the mountain's side breaks away without warning, triggering a massive avalanche. Huge blocks of ice, like shattered glaciers, mix with tons of powdery, dry snow, forming a torrent of white water that cascades down with thunderous force. Massive blocks of snow collide and shatter as they fall, stirring up clouds of snow and fog that obscure the sky. This unstoppable force rushes towards the valley floor at an astonishing speed. The entire scene is filled with a sense of nature's raw, destructive power, creating an atmosphere of tension both majestic and awe-inspiring.", + "seed": 11803, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000119/group_9/index_4731_seed_11803_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.21708093978231774, + "vlm_logprob": -0.21708093978231774, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.21708093978231774, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025203140452504158 + ], + "opposite_logprob_tokens": [ + -3.900203227996826 + ], + "logprob_reward": -0.025203140452504158, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.025203140452504158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17589038610458374, + 0.1048443391919136, + 0.08909512311220169, + 0.045490555465221405, + 0.19794438779354095, + 0.12400201708078384, + 0.045525889843702316, + 0.03519940748810768, + 0.10029274970293045, + 0.08171511441469193 + ], + "credit_map": [ + 0.017037581652402878, + 0.019257185980677605, + 0.01560946460813284, + 0.04053313657641411, + 0.02009161375463009, + 0.028556734323501587, + 0.015062513761222363, + 0.01974213495850563, + 0.008811921812593937, + 0.012516248971223831, + 0.009236101992428303, + 0.021234530955553055, + 0.012029165402054787, + 0.014251680113375187, + 0.011694017797708511, + 0.01507066935300827, + 0.009269081987440586, + 0.013153638690710068, + 0.007930205203592777, + 0.014873355627059937, + 0.009979554452002048, + 0.010808958671987057, + 0.012192578986287117, + 0.0108877494931221, + 0.002835081424564123, + 0.005943662486970425, + 0.00684162974357605, + 0.007806094828993082, + 0.005017068237066269, + 0.006504538934677839, + 0.00551185104995966, + 0.005030625965446234, + 0.011869411915540695, + 0.017155548557639122, + 0.03395865857601166, + 0.02519044652581215, + 0.03879334032535553, + 0.021663803607225418, + 0.027723759412765503, + 0.021589409559965134, + 0.006469969637691975, + 0.011171330697834492, + 0.014596298336982727, + 0.01740400306880474, + 0.02633090130984783, + 0.021556807681918144, + 0.01206202432513237, + 0.014410682953894138, + 0.0035886813420802355, + 0.004268938675522804, + 0.006762182340025902, + 0.006158944685012102, + 0.007547649089246988, + 0.005996872670948505, + 0.005370246712118387, + 0.005832379218190908, + 0.002211236394941807, + 0.0033724105451256037, + 0.00382604100741446, + 0.005179654341191053, + 0.008434430696070194, + 0.0038565904833376408, + 0.004542794544249773, + 0.0037762518040835857, + 0.006048371084034443, + 0.008390801027417183, + 0.01374876406043768, + 0.0095472177490592, + 0.012323679402470589, + 0.02188439853489399, + 0.01821027882397175, + 0.01013924553990364, + 0.00805103313177824, + 0.00702121714130044, + 0.011061049997806549, + 0.009977540001273155, + 0.007419727277010679, + 0.01049580704420805, + 0.010764907114207745, + 0.01692383363842964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.025203140452504158 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.053951621055603 + ], + "opposite_logprob_tokens": [ + -0.42895159125328064 + ], + "logprob_reward": -1.053951621055603, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.053951621055603, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0918598398566246, + 0.051309093832969666, + 0.07663523405790329, + 0.132742777466774, + 0.07478704303503036, + 0.13502933084964752, + 0.05307648330926895, + 0.0951620414853096, + 0.10037481039762497, + 0.18902337551116943 + ], + "credit_map": [ + 0.012855390086770058, + 0.011319284327328205, + 0.0067810374312102795, + 0.016856851056218147, + 0.009877271018922329, + 0.010346535593271255, + 0.013677486218512058, + 0.010145981796085835, + 0.007169835269451141, + 0.005493264179676771, + 0.004902490880340338, + 0.010174818336963654, + 0.00491890637204051, + 0.005466809030622244, + 0.007717906031757593, + 0.005465065129101276, + 0.00685033667832613, + 0.014014971442520618, + 0.005820735357701778, + 0.012791676446795464, + 0.00947810709476471, + 0.009490413591265678, + 0.010106912814080715, + 0.008082075044512749, + 0.01027406845241785, + 0.01680700294673443, + 0.02600177191197872, + 0.017153413966298103, + 0.020521128550171852, + 0.013235296122729778, + 0.014438973739743233, + 0.014311122708022594, + 0.004816286265850067, + 0.008149603381752968, + 0.011471636593341827, + 0.009097808040678501, + 0.013734741136431694, + 0.008706926368176937, + 0.010282889008522034, + 0.00852715689688921, + 0.00745023600757122, + 0.013274874538183212, + 0.021295595914125443, + 0.02135392278432846, + 0.022293739020824432, + 0.016248201951384544, + 0.017584649845957756, + 0.015528119169175625, + 0.003402936039492488, + 0.006548404693603516, + 0.008719517849385738, + 0.007968062534928322, + 0.009307809174060822, + 0.004946738947182894, + 0.005102158058434725, + 0.007080854382365942, + 0.005539072677493095, + 0.008692735806107521, + 0.019769713282585144, + 0.012619570828974247, + 0.014753085561096668, + 0.00952822994440794, + 0.013979385606944561, + 0.010280255228281021, + 0.006032801233232021, + 0.007484689354896545, + 0.01202896237373352, + 0.011667565442621708, + 0.017033148556947708, + 0.024166110903024673, + 0.012845097109675407, + 0.009116430766880512, + 0.018300870433449745, + 0.01677098125219345, + 0.03232240676879883, + 0.025473667308688164, + 0.020913178101181984, + 0.02586403861641884, + 0.02438821829855442, + 0.024990001693367958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -1.053951621055603 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008182753226719797 + ], + "opposite_logprob_tokens": [ + -7.125818252563477 + ], + "logprob_reward": -0.0008182753226719797, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008182753226719797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21709217131137848, + 0.09808043390512466, + 0.0969265028834343, + 0.0593114048242569, + 0.11040254682302475, + 0.05733678862452507, + 0.046400848776102066, + 0.03489850088953972, + 0.13036899268627167, + 0.1491817682981491 + ], + "credit_map": [ + 0.018313366919755936, + 0.021873919293284416, + 0.023475665599107742, + 0.0515570230782032, + 0.02304663509130478, + 0.033465299755334854, + 0.01996452920138836, + 0.025395741686224937, + 0.007917416281998158, + 0.010713459923863411, + 0.014353246428072453, + 0.01573719084262848, + 0.010679618455469608, + 0.015657363459467888, + 0.011076311580836773, + 0.011945835314691067, + 0.007064618170261383, + 0.010148411616683006, + 0.013925965875387192, + 0.014190632849931717, + 0.012186793610453606, + 0.016081783920526505, + 0.01253457646816969, + 0.01079372875392437, + 0.005434670019894838, + 0.004556421656161547, + 0.008307360112667084, + 0.009378012269735336, + 0.01049110759049654, + 0.007194647565484047, + 0.006576623767614365, + 0.007372559979557991, + 0.010620695538818836, + 0.017465637996792793, + 0.012422493658959866, + 0.014430729672312737, + 0.01278710924088955, + 0.016271039843559265, + 0.01421442162245512, + 0.012190433219075203, + 0.0033421122934669256, + 0.005681995768100023, + 0.010255624540150166, + 0.009590685367584229, + 0.005475278478115797, + 0.009409650228917599, + 0.006656982004642487, + 0.006924464832991362, + 0.00386556307785213, + 0.005199266131967306, + 0.0062791830860078335, + 0.005829814355820417, + 0.005706531926989555, + 0.008170312270522118, + 0.005217444151639938, + 0.006132736336439848, + 0.002174400258809328, + 0.0029773914720863104, + 0.0034128413535654545, + 0.004920024424791336, + 0.00935506820678711, + 0.0042519159615039825, + 0.004326510243117809, + 0.0034803536254912615, + 0.008483752608299255, + 0.011071380227804184, + 0.01754673942923546, + 0.01329617016017437, + 0.01502292975783348, + 0.021015023812651634, + 0.02721843309700489, + 0.01671457663178444, + 0.011849252507090569, + 0.008651954121887684, + 0.021346205845475197, + 0.016025833785533905, + 0.008351911790668964, + 0.01846284791827202, + 0.0225656870752573, + 0.041928090155124664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0008182753226719797 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409455131739378 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.00409455131739378, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409455131739378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06666354835033417, + 0.03518601879477501, + 0.07533527910709381, + 0.08841823041439056, + 0.13331842422485352, + 0.26657310128211975, + 0.14225229620933533, + 0.05323135107755661, + 0.08717121183872223, + 0.051850564777851105 + ], + "credit_map": [ + 0.007365988567471504, + 0.007093448657542467, + 0.006103853695094585, + 0.015407864935696125, + 0.007666620425879955, + 0.007108842488378286, + 0.008214841596782207, + 0.007702088914811611, + 0.002658925950527191, + 0.0038476891350001097, + 0.004682497121393681, + 0.006640211213380098, + 0.003931649494916201, + 0.004748988896608353, + 0.0047374023124575615, + 0.003938655834645033, + 0.006614679470658302, + 0.009871558286249638, + 0.008246568962931633, + 0.010796291753649712, + 0.008975477889180183, + 0.01071266457438469, + 0.011295843869447708, + 0.008822191506624222, + 0.005539608187973499, + 0.007315714377909899, + 0.013757344335317612, + 0.016377178952097893, + 0.013951780274510384, + 0.010351382195949554, + 0.011531822383403778, + 0.009593402035534382, + 0.007994502782821655, + 0.011160372756421566, + 0.01619231328368187, + 0.02235398441553116, + 0.025086862966418266, + 0.021430009976029396, + 0.01473584957420826, + 0.014364529401063919, + 0.015493921935558319, + 0.026829035952687263, + 0.04133226349949837, + 0.04114513471722603, + 0.047473419457674026, + 0.031520456075668335, + 0.03170664608478546, + 0.0310722216963768, + 0.008642502129077911, + 0.01579597033560276, + 0.021952293813228607, + 0.020386114716529846, + 0.02397840842604637, + 0.01490748580545187, + 0.017587769776582718, + 0.019001740962266922, + 0.003184522269293666, + 0.005610515829175711, + 0.008527504280209541, + 0.008121677674353123, + 0.008442838676273823, + 0.006335345562547445, + 0.006666991859674454, + 0.006341955624520779, + 0.0050673107616603374, + 0.0075678639113903046, + 0.01254255510866642, + 0.007497852202504873, + 0.009862036444246769, + 0.017795933410525322, + 0.017144499346613884, + 0.009693159721791744, + 0.003195455763489008, + 0.003567708656191826, + 0.00588911771774292, + 0.005714982282370329, + 0.004132634028792381, + 0.006328942719846964, + 0.00913955643773079, + 0.013882169499993324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.00409455131739378 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a wide shot captured from afar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pull out to reveal a grander panorama?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a towering, rugged, snow-capped mountain visible beneath a clear, cold sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive avalanche occur where a large section of the snow slope breaks away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do huge blocks of ice and powdery snow cascade down the mountain at high speed?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013371107634156942 + ], + "opposite_logprob_tokens": [ + -6.626337051391602 + ], + "logprob_reward": -0.0013371107634156942, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013371107634156942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09822316467761993, + 0.03730277344584465, + 0.06055684760212898, + 0.10565242916345596, + 0.11579019576311111, + 0.17466449737548828, + 0.16392816603183746, + 0.09373298287391663, + 0.09325312077999115, + 0.056895773857831955 + ], + "credit_map": [ + 0.01164516992866993, + 0.01077420637011528, + 0.0088926050812006, + 0.023074619472026825, + 0.009989652782678604, + 0.012045491486787796, + 0.01094785425812006, + 0.01085357740521431, + 0.0028874152339994907, + 0.0038229140918701887, + 0.0038514689076691866, + 0.01016948465257883, + 0.004148397594690323, + 0.004348096903413534, + 0.00464034266769886, + 0.0034346566535532475, + 0.0056367493234574795, + 0.007709484081715345, + 0.005591538269072771, + 0.008003363385796547, + 0.00966640841215849, + 0.008082959800958633, + 0.009404167532920837, + 0.006462185177952051, + 0.007168053649365902, + 0.008499659597873688, + 0.016645606607198715, + 0.020069623365998268, + 0.016370870172977448, + 0.01201078575104475, + 0.01305309683084488, + 0.011834735982120037, + 0.006660168059170246, + 0.010158490389585495, + 0.014788191765546799, + 0.01809508353471756, + 0.019101006910204887, + 0.020052608102560043, + 0.01426555123180151, + 0.012669096700847149, + 0.010319310240447521, + 0.017632557079195976, + 0.029149556532502174, + 0.028446834534406662, + 0.028078319504857063, + 0.019590191543102264, + 0.02105749025940895, + 0.020390240475535393, + 0.011650036089122295, + 0.019885806366801262, + 0.027959303930401802, + 0.02517213113605976, + 0.02208760753273964, + 0.016131365671753883, + 0.019302645698189735, + 0.021739279851317406, + 0.004963153041899204, + 0.010255029425024986, + 0.018862804397940636, + 0.011262119747698307, + 0.013829699717462063, + 0.011111280880868435, + 0.012261305935680866, + 0.01118759997189045, + 0.00512254424393177, + 0.00884705688804388, + 0.014273387379944324, + 0.00831997487694025, + 0.008824515156447887, + 0.016839150339365005, + 0.019703853875398636, + 0.011322645470499992, + 0.0042146421037614346, + 0.0046263025142252445, + 0.008082758635282516, + 0.00724163930863142, + 0.005192595534026623, + 0.007463148329406977, + 0.009423570707440376, + 0.010651122778654099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.0013371107634156942 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_0/index_tensor([4937, 4937])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_0/index_tensor([4937, 4937])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3872794d0c91104bb0fb42a5b292e372518cf315 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_0/index_tensor([4937, 4937])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, static shot, framed vertically, captures a fully blossomed lotus flower, floating steadily and calmly on the surface of crystal-clear, mirror-like water. The petals, layered and gracefully unfolding, take on a velvety texture. Their delicate pink gradually transitions to a soft, milky white from the tips to the center, nestled within a golden pod. Soft, golden morning sunlight slants through the mist, precisely illuminating the distinct veins of the petals and making the crystal-clear water droplets clinging to them sparkle like tiny diamonds. As the petals sway gently, a plump droplet, unbalanced by gravity, begins to slide cautiously down the petal's graceful curve, leaving a wet, sparkling trail before silently dripping back into the water, stirring up a circle of tiny, rapidly spreading ripples. Through the clear water, you can vaguely see the green aquatic plants swaying underwater. The whole picture is extremely rich in details, with delicate light and shadow, creating an extremely tranquil, pure and Zen atmosphere.", + "seed": 20629, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_0/index_4937_seed_20629_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20401085847988726, + "vlm_logprob": -0.20401085847988726, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.20401085847988726, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007668462116271257 + ], + "opposite_logprob_tokens": [ + -6.382668495178223 + ], + "logprob_reward": -0.007668462116271257, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.007668462116271257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18738006055355072, + 0.16150817275047302, + 0.06322693824768066, + 0.044092483818531036, + 0.08107659220695496, + 0.05160430073738098, + 0.11049417406320572, + 0.18591536581516266, + 0.06539898365736008, + 0.04930300638079643 + ], + "credit_map": [ + 0.026611920446157455, + 0.021926328539848328, + 0.02499500662088394, + 0.017343807965517044, + 0.03046487085521221, + 0.024903835728764534, + 0.013734200969338417, + 0.027400091290473938, + 0.03380066156387329, + 0.012946403585374355, + 0.023891031742095947, + 0.006825743243098259, + 0.011893197894096375, + 0.03610232099890709, + 0.007849362678825855, + 0.028199436143040657, + 0.014245666563510895, + 0.003664591582491994, + 0.008007233962416649, + 0.0036563544999808073, + 0.005045667756348848, + 0.014908844605088234, + 0.003878960618749261, + 0.009819624945521355, + 0.00843554362654686, + 0.004801597446203232, + 0.0064705898985266685, + 0.003785930573940277, + 0.0063203261233866215, + 0.004947599954903126, + 0.004213877022266388, + 0.0051170168444514275, + 0.01884060725569725, + 0.005299371201545, + 0.016084497794508934, + 0.005300926044583321, + 0.004809155594557524, + 0.012807749211788177, + 0.004527316428720951, + 0.013406968675553799, + 0.012372525408864021, + 0.004164683632552624, + 0.00799007061868906, + 0.003064163262024522, + 0.004752406384795904, + 0.006155935116112232, + 0.0038501194212585688, + 0.00925439689308405, + 0.021168813109397888, + 0.012634078040719032, + 0.01290918793529272, + 0.006979851517826319, + 0.014116066507995129, + 0.014524082653224468, + 0.008148571476340294, + 0.020013511180877686, + 0.023267896845936775, + 0.006929405499249697, + 0.011842610314488411, + 0.004470831248909235, + 0.00821932777762413, + 0.0650949776172638, + 0.005003432277590036, + 0.06108687445521355, + 0.013301127590239048, + 0.004124477971345186, + 0.011588011868298054, + 0.004104174207895994, + 0.005562777630984783, + 0.012020367197692394, + 0.004148299340158701, + 0.01054974365979433, + 0.010042240843176842, + 0.003514552488923073, + 0.007858119904994965, + 0.003388327080756426, + 0.005885246209800243, + 0.006341388449072838, + 0.004678848199546337, + 0.007594283204525709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.007668462116271257 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.474301278591156 + ], + "opposite_logprob_tokens": [ + -0.974301278591156 + ], + "logprob_reward": -0.474301278591156, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.474301278591156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08855770528316498, + 0.20811547338962555, + 0.24405129253864288, + 0.024008678272366524, + 0.04649697616696358, + 0.08402746170759201, + 0.12213816493749619, + 0.1286325603723526, + 0.0211233738809824, + 0.03284834325313568 + ], + "credit_map": [ + 0.01863090507686138, + 0.006711368449032307, + 0.013163314200937748, + 0.004832136910408735, + 0.008925944566726685, + 0.016350744292140007, + 0.005375376436859369, + 0.014567910693585873, + 0.04800194129347801, + 0.013379829935729504, + 0.034370891749858856, + 0.008286072872579098, + 0.01235341839492321, + 0.04413063824176788, + 0.009402104653418064, + 0.03819053992629051, + 0.056073036044836044, + 0.015098911710083485, + 0.027165791019797325, + 0.00996341835707426, + 0.019029537215828896, + 0.06407462805509567, + 0.012727039866149426, + 0.03991890698671341, + 0.004599441774189472, + 0.0015667421976104379, + 0.0022575869224965572, + 0.0013732427032664418, + 0.002588625531643629, + 0.005482642911374569, + 0.0016696203965693712, + 0.0044707744382321835, + 0.010589527897536755, + 0.0037948833778500557, + 0.00863898079842329, + 0.0024097019340842962, + 0.004517401568591595, + 0.006876260507851839, + 0.002851598896086216, + 0.00681861350312829, + 0.018002323806285858, + 0.0042759720236063, + 0.020635513588786125, + 0.0051473150961101055, + 0.00369975040666759, + 0.01054892223328352, + 0.006120406091213226, + 0.01559725683182478, + 0.02435656078159809, + 0.010735074058175087, + 0.01759343408048153, + 0.006418136414140463, + 0.010126128792762756, + 0.02262767218053341, + 0.007214282173663378, + 0.023066863417625427, + 0.021069442853331566, + 0.009109510108828545, + 0.014102225191891193, + 0.007117020431905985, + 0.008704383857548237, + 0.02643335610628128, + 0.008079581893980503, + 0.034017033874988556, + 0.00425755837932229, + 0.0012580535840243101, + 0.004010474309325218, + 0.0011935234069824219, + 0.0014560080599039793, + 0.00412174966186285, + 0.0016024995129555464, + 0.0032235044054687023, + 0.0047546494752168655, + 0.005055236630141735, + 0.003735494567081332, + 0.0023197527043521404, + 0.005844288971275091, + 0.003082573413848877, + 0.0029292607214301825, + 0.005127084907144308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.474301278591156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294778436422348 + ], + "opposite_logprob_tokens": [ + -2.237947702407837 + ], + "logprob_reward": -0.11294778436422348, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294778436422348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04602058231830597, + 0.18087905645370483, + 0.3005290925502777, + 0.026874922215938568, + 0.06499391049146652, + 0.10220081359148026, + 0.078792043030262, + 0.13557590544223785, + 0.02504766546189785, + 0.03908601775765419 + ], + "credit_map": [ + 0.008601648733019829, + 0.003922416362911463, + 0.009203042834997177, + 0.0029302325565367937, + 0.004315992817282677, + 0.0067322333343327045, + 0.0034617697820067406, + 0.006853245664387941, + 0.041628096252679825, + 0.009531556628644466, + 0.03162071853876114, + 0.008643603883683681, + 0.009677289985120296, + 0.03905065357685089, + 0.008835784159600735, + 0.0318913571536541, + 0.06594426184892654, + 0.019251761958003044, + 0.04137389734387398, + 0.012820680625736713, + 0.020859522745013237, + 0.07395677268505096, + 0.01637488789856434, + 0.04994731768965721, + 0.0062601217068731785, + 0.0016534783644601703, + 0.006367120426148176, + 0.0015423642471432686, + 0.0015715681947767735, + 0.00426066480576992, + 0.0015577307203784585, + 0.0036618737503886223, + 0.015222407877445221, + 0.004603634588420391, + 0.01366532500833273, + 0.0033122459426522255, + 0.004363230429589748, + 0.009824709966778755, + 0.0036842154804617167, + 0.01031813956797123, + 0.022658463567495346, + 0.004559851251542568, + 0.025104893371462822, + 0.006253999192267656, + 0.0034274253994226456, + 0.01207900233566761, + 0.007282446138560772, + 0.020834730938076973, + 0.014108199626207352, + 0.00639687804505229, + 0.011054436676204205, + 0.004920321051031351, + 0.005399363115429878, + 0.01208682544529438, + 0.006015043240040541, + 0.018810978159308434, + 0.0297156423330307, + 0.008046410977840424, + 0.016752030700445175, + 0.006008656229823828, + 0.005416010972112417, + 0.028837090358138084, + 0.007139624562114477, + 0.03366043046116829, + 0.004395289346575737, + 0.001982560148462653, + 0.003593666246160865, + 0.0017182830488309264, + 0.0024651752319186926, + 0.004059572704136372, + 0.0015053407987579703, + 0.005327780265361071, + 0.008842757903039455, + 0.0026337371673434973, + 0.009366300888359547, + 0.0022522411309182644, + 0.0020153564400970936, + 0.005102954339236021, + 0.0023463652469217777, + 0.006526306737214327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.11294778436422348 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811153024435043 + ], + "opposite_logprob_tokens": [ + -3.288111448287964 + ], + "logprob_reward": -0.03811153024435043, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811153024435043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11566343903541565, + 0.05751395225524902, + 0.05476272478699684, + 0.039872296154499054, + 0.04227733239531517, + 0.0902390405535698, + 0.20801398158073425, + 0.35764771699905396, + 0.01516712922602892, + 0.018842335790395737 + ], + "credit_map": [ + 0.027428383007645607, + 0.009259240701794624, + 0.013691486790776253, + 0.005357804708182812, + 0.007435017731040716, + 0.026021188125014305, + 0.005711828358471394, + 0.02075849659740925, + 0.012747454456984997, + 0.0037199289072304964, + 0.010681701824069023, + 0.0029738638550043106, + 0.0032901985105127096, + 0.01137804426252842, + 0.0031493206042796373, + 0.009573438204824924, + 0.012062028050422668, + 0.003281878773123026, + 0.005973308812826872, + 0.002187651814892888, + 0.0044394563883543015, + 0.014728379435837269, + 0.002870929427444935, + 0.0092190932482481, + 0.007076446432620287, + 0.003303971141576767, + 0.0064135706052184105, + 0.002474036067724228, + 0.004077111836522818, + 0.007780618034303188, + 0.002951863221824169, + 0.00579468347132206, + 0.008206949569284916, + 0.002842175541445613, + 0.01129732932895422, + 0.0024703610688447952, + 0.0033436892554163933, + 0.0059571885503828526, + 0.0028469508979469538, + 0.005312686320394278, + 0.014991297386586666, + 0.0062756771221756935, + 0.02231815829873085, + 0.0054237013682723045, + 0.004792015999555588, + 0.010473782196640968, + 0.0060556926764547825, + 0.019908713176846504, + 0.0446670837700367, + 0.014125630259513855, + 0.02845495194196701, + 0.012584460899233818, + 0.010960564017295837, + 0.03766798973083496, + 0.014992866665124893, + 0.04456043988466263, + 0.05028950795531273, + 0.013530688360333443, + 0.03148084506392479, + 0.011606195941567421, + 0.012968573719263077, + 0.11337622255086899, + 0.014286919496953487, + 0.11010876297950745, + 0.0026002393569797277, + 0.00119013839866966, + 0.001951094251126051, + 0.0012715809280052781, + 0.001816099276766181, + 0.002279518870636821, + 0.0011617642594501376, + 0.0028966935351490974, + 0.0026910314336419106, + 0.0030089092906564474, + 0.0015707914717495441, + 0.0013743200106546283, + 0.0038054368924349546, + 0.0019876165315508842, + 0.0018019239651039243, + 0.002602306893095374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.03811153024435043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38702523708343506 + ], + "opposite_logprob_tokens": [ + -1.137025237083435 + ], + "logprob_reward": -0.38702523708343506, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38702523708343506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11120130866765976, + 0.1226772740483284, + 0.08951922506093979, + 0.0441078320145607, + 0.07180998474359512, + 0.09693355858325958, + 0.056199196726083755, + 0.2617778778076172, + 0.03682830557227135, + 0.10894547402858734 + ], + "credit_map": [ + 0.02452275902032852, + 0.008868684060871601, + 0.018010608851909637, + 0.004944788292050362, + 0.007333827205002308, + 0.02355664037168026, + 0.0045534465461969376, + 0.01941055618226528, + 0.033297549933195114, + 0.006553456652909517, + 0.022350242361426353, + 0.006432240828871727, + 0.007508879527449608, + 0.021519698202610016, + 0.005660701543092728, + 0.01935451477766037, + 0.01979382336139679, + 0.005173414945602417, + 0.014523460529744625, + 0.003924949560314417, + 0.00498297018930316, + 0.020963462069630623, + 0.005177851766347885, + 0.014979296363890171, + 0.007479654625058174, + 0.0032392877619713545, + 0.004422508645802736, + 0.002218889305368066, + 0.003644571639597416, + 0.011952108703553677, + 0.002664210507646203, + 0.008486599661409855, + 0.01797574572265148, + 0.0049166809767484665, + 0.016329510137438774, + 0.0032675755210220814, + 0.004719546530395746, + 0.01075578574091196, + 0.004775060806423426, + 0.009070080704987049, + 0.01596556045114994, + 0.004590719472616911, + 0.02729412540793419, + 0.005041936878114939, + 0.004230409394949675, + 0.011874094605445862, + 0.007931632921099663, + 0.02000507339835167, + 0.01020648330450058, + 0.00585183734074235, + 0.007555169053375721, + 0.004354353062808514, + 0.005780376959592104, + 0.006709040608257055, + 0.004546757787466049, + 0.011195180006325245, + 0.03045598231256008, + 0.010098936036229134, + 0.018450433388352394, + 0.006632088217884302, + 0.009478705003857613, + 0.09419754892587662, + 0.007121336180716753, + 0.08534286171197891, + 0.00716075673699379, + 0.0019470517290756106, + 0.0058501833118498325, + 0.0022199873346835375, + 0.0022497265599668026, + 0.008471013978123665, + 0.0026185845490545034, + 0.006311001256108284, + 0.02516940049827099, + 0.0053380802273750305, + 0.02425624057650566, + 0.005481128115206957, + 0.005583824589848518, + 0.014900697395205498, + 0.006172672379761934, + 0.02204342558979988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.38702523708343506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot, framed vertically, captures a fully blossomed lotus flower, floating steadily and calmly on the surface of crystal-clear, mirror-like water. The petals, layered and gracefully unfolding, take on a velvety texture. Their delicate pink gradually transitions to a soft, milky white from the tips to the center, nestled within a golden pod. Soft, golden morning sunlight slants through the mist, precisely illuminating the distinct veins of the petals and making the crystal-clear water droplets clinging to them sparkle like tiny diamonds. As the petals sway gently, a plump droplet, unbalanced by gravity, begins to slide cautiously down the petal's graceful curve, leaving a wet, sparkling trail before silently dripping back into the water, stirring up a circle of tiny, rapidly spreading ripples. Through the clear water, you can vaguely see the green aquatic plants swaying underwater. The whole picture is extremely rich in details, with delicate light and shadow, creating an extremely tranquil, pure and Zen atmosphere.", + "seed": 20630, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_0/index_4937_seed_20630_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09854688355699182, + "vlm_logprob": -0.09854688355699182, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.09854688355699182, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005276088137179613 + ], + "opposite_logprob_tokens": [ + -6.755276203155518 + ], + "logprob_reward": -0.005276088137179613, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.005276088137179613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10048195719718933, + 0.12228599190711975, + 0.0926469936966896, + 0.03587684407830238, + 0.040817584842443466, + 0.11204952001571655, + 0.06687571108341217, + 0.1413404941558838, + 0.12182886898517609, + 0.16579608619213104 + ], + "credit_map": [ + 0.01133174542337656, + 0.015208523720502853, + 0.012555685825645924, + 0.011099768802523613, + 0.01281801238656044, + 0.016749581322073936, + 0.010459959506988525, + 0.010258672758936882, + 0.01757875271141529, + 0.01795811392366886, + 0.018688691779971123, + 0.009621412493288517, + 0.017342787235975266, + 0.019528327509760857, + 0.011533462442457676, + 0.010034441947937012, + 0.012593028135597706, + 0.007773535791784525, + 0.013870236463844776, + 0.007957923226058483, + 0.011052023619413376, + 0.020931605249643326, + 0.009306908585131168, + 0.009161729365587234, + 0.005003360100090504, + 0.004776161629706621, + 0.005317933391779661, + 0.004117692820727825, + 0.004047175403684378, + 0.006036240141838789, + 0.003727132920175791, + 0.0028511430136859417, + 0.004798708483576775, + 0.004571487661451101, + 0.004817933309823275, + 0.005484732333570719, + 0.00540993083268404, + 0.005043070297688246, + 0.005013109650462866, + 0.0056786080822348595, + 0.020823320373892784, + 0.01049103680998087, + 0.014187654480338097, + 0.006807411555200815, + 0.008659850805997849, + 0.024456042796373367, + 0.006444548722356558, + 0.02017965167760849, + 0.01138146873563528, + 0.005659629590809345, + 0.00952358078211546, + 0.005893934518098831, + 0.0057908413000404835, + 0.008490540087223053, + 0.006248808465898037, + 0.01388690248131752, + 0.02780584804713726, + 0.013946009799838066, + 0.019845033064484596, + 0.012525700032711029, + 0.01356272678822279, + 0.020006632432341576, + 0.010847142897546291, + 0.02280140481889248, + 0.020413951948285103, + 0.010497898794710636, + 0.017277393490076065, + 0.008725486695766449, + 0.012461403384804726, + 0.024055372923612595, + 0.008477208204567432, + 0.019920146092772484, + 0.025638477876782417, + 0.014456472359597683, + 0.02378472127020359, + 0.012876302003860474, + 0.014825012534856796, + 0.03252033144235611, + 0.013219390995800495, + 0.028475353494286537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.005276088137179613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34865036606788635 + ], + "opposite_logprob_tokens": [ + -1.223650336265564 + ], + "logprob_reward": -0.34865036606788635, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34865036606788635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10311822593212128, + 0.06285687536001205, + 0.04257234185934067, + 0.05090607702732086, + 0.059416718780994415, + 0.23870036005973816, + 0.06608345359563828, + 0.2180439531803131, + 0.09142079949378967, + 0.06688126176595688 + ], + "credit_map": [ + 0.01773538440465927, + 0.013389043509960175, + 0.012868101708590984, + 0.009153412654995918, + 0.010316492058336735, + 0.019351251423358917, + 0.009369405917823315, + 0.010935133323073387, + 0.007707869168370962, + 0.010983013547956944, + 0.008112283423542976, + 0.006157496944069862, + 0.013509470038115978, + 0.007567891385406256, + 0.004889005329459906, + 0.003929843194782734, + 0.006543186493217945, + 0.003794497111812234, + 0.005974696483463049, + 0.00572693906724453, + 0.005202299449592829, + 0.0058001186698675156, + 0.0046350532211363316, + 0.004895546939224005, + 0.007625643629580736, + 0.004361956380307674, + 0.007405448239296675, + 0.004237515851855278, + 0.005769279319792986, + 0.010407153517007828, + 0.005529696121811867, + 0.005569386761635542, + 0.010515118017792702, + 0.004335199482738972, + 0.00792552437633276, + 0.006174448877573013, + 0.004814631305634975, + 0.011883800849318504, + 0.004953921772539616, + 0.008814076893031597, + 0.030216408893465996, + 0.023043889552354813, + 0.03568752855062485, + 0.015524645335972309, + 0.020425215363502502, + 0.05512963607907295, + 0.01744178868830204, + 0.04123125225305557, + 0.008743143640458584, + 0.006033778190612793, + 0.008767163380980492, + 0.007753465790301561, + 0.0085143418982625, + 0.009299339726567268, + 0.006849709898233414, + 0.010122501291334629, + 0.04749363660812378, + 0.017601167783141136, + 0.033350154757499695, + 0.018492072820663452, + 0.015474469400942326, + 0.03386470302939415, + 0.017082108184695244, + 0.0346856489777565, + 0.015145550481975079, + 0.009086030535399914, + 0.014865772798657417, + 0.005978719796985388, + 0.011146021075546741, + 0.016797613352537155, + 0.005706764291971922, + 0.012694323435425758, + 0.009817148558795452, + 0.006193750072270632, + 0.009371770545840263, + 0.004968639928847551, + 0.006784945726394653, + 0.012493401765823364, + 0.005886052269488573, + 0.011365549638867378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.34865036606788635 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205747276544571 + ], + "opposite_logprob_tokens": [ + -2.8120574951171875 + ], + "logprob_reward": -0.06205747276544571, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205747276544571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07865890860557556, + 0.03408293426036835, + 0.035792529582977295, + 0.044340189546346664, + 0.08295425027608871, + 0.10069819539785385, + 0.2755711078643799, + 0.13390861451625824, + 0.10173559188842773, + 0.11225771903991699 + ], + "credit_map": [ + 0.010892746038734913, + 0.008046682924032211, + 0.00960292387753725, + 0.011728527955710888, + 0.008261846378445625, + 0.010301926173269749, + 0.009130630642175674, + 0.010693628340959549, + 0.005367407575249672, + 0.0031855886336416006, + 0.0055455490946769714, + 0.0025089203845709562, + 0.004077676683664322, + 0.005562107544392347, + 0.0037564041558653116, + 0.004079280886799097, + 0.007865150459110737, + 0.002509075915440917, + 0.00564908841624856, + 0.0026800448540598154, + 0.0030365714337676764, + 0.0086813447996974, + 0.0018383660353720188, + 0.0035328881349414587, + 0.009574785828590393, + 0.0030693558510392904, + 0.0071083917282521725, + 0.0038061589002609253, + 0.0033550008665770292, + 0.010584667325019836, + 0.002470507752150297, + 0.004371316637843847, + 0.012799867428839207, + 0.006011923775076866, + 0.012826205231249332, + 0.007406703196465969, + 0.007524840068072081, + 0.0161276925355196, + 0.0060626971535384655, + 0.014194317162036896, + 0.014457845129072666, + 0.006262481212615967, + 0.012005891650915146, + 0.00730623584240675, + 0.007914717309176922, + 0.016851194202899933, + 0.008279824629426003, + 0.027620013803243637, + 0.03946741670370102, + 0.017617354169487953, + 0.03727603703737259, + 0.026836812496185303, + 0.018858257681131363, + 0.036912910640239716, + 0.022908369079232216, + 0.07569395005702972, + 0.026869304478168488, + 0.01037074625492096, + 0.02045495994389057, + 0.011040580458939075, + 0.010327272117137909, + 0.02046460472047329, + 0.010494092479348183, + 0.02388705499470234, + 0.015468505211174488, + 0.01457857247442007, + 0.016453644260764122, + 0.008362012915313244, + 0.012133555486798286, + 0.013220393098890781, + 0.008396757766604424, + 0.01312214694917202, + 0.01816035434603691, + 0.00725924177095294, + 0.017362521961331367, + 0.008454873226583004, + 0.009182881563901901, + 0.019888393580913544, + 0.009467469528317451, + 0.02248198725283146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.06205747276544571 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033705245703458786 + ], + "opposite_logprob_tokens": [ + -3.408705234527588 + ], + "logprob_reward": -0.033705245703458786, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033705245703458786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04003902152180672, + 0.037658192217350006, + 0.02335529215633869, + 0.02863377518951893, + 0.07887116074562073, + 0.3173864185810089, + 0.06854081898927689, + 0.33516189455986023, + 0.04751753434538841, + 0.02283594012260437 + ], + "credit_map": [ + 0.00644133985042572, + 0.004722567740827799, + 0.0048323823139071465, + 0.005376534536480904, + 0.004720229655504227, + 0.006947740446776152, + 0.00353427417576313, + 0.0034639532677829266, + 0.005633778404444456, + 0.0039011870976537466, + 0.00557398097589612, + 0.0036968551576137543, + 0.004057531710714102, + 0.00776791712269187, + 0.002833503996953368, + 0.004193435423076153, + 0.004083243664354086, + 0.0019884007051587105, + 0.0034783631563186646, + 0.002370347734540701, + 0.0019877294544130564, + 0.004494254942983389, + 0.0020073940977454185, + 0.0029455572366714478, + 0.005075450055301189, + 0.002739288844168186, + 0.00418985215947032, + 0.0026565927546471357, + 0.0029569873586297035, + 0.005872736219316721, + 0.002190124476328492, + 0.0029527447186410427, + 0.0157857034355402, + 0.0060723996721208096, + 0.01288298424333334, + 0.004692613612860441, + 0.006520358379930258, + 0.01738007925450802, + 0.00401743408292532, + 0.011519592255353928, + 0.042671166360378265, + 0.022285861894488335, + 0.042788315564394, + 0.020830361172556877, + 0.02901645004749298, + 0.06409097462892532, + 0.022622769698500633, + 0.07308053970336914, + 0.009472568519413471, + 0.006451744586229324, + 0.00834205374121666, + 0.006792332045733929, + 0.0052756574004888535, + 0.009281618520617485, + 0.006614128593355417, + 0.016310716047883034, + 0.06897766143083572, + 0.026103710755705833, + 0.04922305420041084, + 0.028657324612140656, + 0.024806367233395576, + 0.051881272345781326, + 0.026168815791606903, + 0.05934370309114456, + 0.00703373784199357, + 0.007274176459759474, + 0.007612315472215414, + 0.0037872102111577988, + 0.005219137761741877, + 0.005927686113864183, + 0.004156272858381271, + 0.0065069948323071, + 0.0034020566381514072, + 0.0023689253721386194, + 0.0032414007000625134, + 0.0021585070062428713, + 0.00327374879270792, + 0.0032213919330388308, + 0.001628429745323956, + 0.00354147981852293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.033705245703458786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304524511098862 + ], + "opposite_logprob_tokens": [ + -3.1680452823638916 + ], + "logprob_reward": -0.04304524511098862, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304524511098862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07299342751502991, + 0.023760560899972916, + 0.03567058965563774, + 0.02793418988585472, + 0.10983403772115707, + 0.11837411671876907, + 0.07199761271476746, + 0.1769561767578125, + 0.137962207198143, + 0.22451697289943695 + ], + "credit_map": [ + 0.01389920525252819, + 0.007304316386580467, + 0.009731419384479523, + 0.008770845830440521, + 0.006634791847318411, + 0.011733438819646835, + 0.006003540940582752, + 0.008915875107049942, + 0.003565679769963026, + 0.0029029198922216892, + 0.0037127991672605276, + 0.0017847444396466017, + 0.0027868684846907854, + 0.004658187739551067, + 0.0020083945710211992, + 0.0023409684654325247, + 0.006601185537874699, + 0.0029142778366804123, + 0.0054276203736662865, + 0.002719528740271926, + 0.0036390996538102627, + 0.007797556463629007, + 0.002535763895139098, + 0.004035555757582188, + 0.005428450647741556, + 0.002051068702712655, + 0.004698613192886114, + 0.0019005602225661278, + 0.002171257743611932, + 0.006237825844436884, + 0.0020493182819336653, + 0.0033970950171351433, + 0.01896471530199051, + 0.007415794767439365, + 0.01644754223525524, + 0.010756135918200016, + 0.0100991390645504, + 0.0219353549182415, + 0.007298813201487064, + 0.01691654697060585, + 0.019180044531822205, + 0.010859123431146145, + 0.013803509995341301, + 0.007077008951455355, + 0.011265793815255165, + 0.019809694960713387, + 0.00801283773034811, + 0.028366094455122948, + 0.009367895312607288, + 0.005774478428065777, + 0.009251669049263, + 0.00781325250864029, + 0.008248521946370602, + 0.008600872941315174, + 0.00643931794911623, + 0.01650160737335682, + 0.032087456434965134, + 0.013423158787190914, + 0.025926535949110985, + 0.014269473031163216, + 0.016559606418013573, + 0.027735980227589607, + 0.013387693092226982, + 0.03356628492474556, + 0.02450495958328247, + 0.011160634458065033, + 0.01928974874317646, + 0.009063168428838253, + 0.01291683129966259, + 0.030145902186632156, + 0.009294637478888035, + 0.021586334332823753, + 0.03767779842019081, + 0.014633871614933014, + 0.0330689400434494, + 0.016745880246162415, + 0.020106961950659752, + 0.04170979559421539, + 0.017587844282388687, + 0.04298589378595352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.04304524511098862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot, framed vertically, captures a fully blossomed lotus flower, floating steadily and calmly on the surface of crystal-clear, mirror-like water. The petals, layered and gracefully unfolding, take on a velvety texture. Their delicate pink gradually transitions to a soft, milky white from the tips to the center, nestled within a golden pod. Soft, golden morning sunlight slants through the mist, precisely illuminating the distinct veins of the petals and making the crystal-clear water droplets clinging to them sparkle like tiny diamonds. As the petals sway gently, a plump droplet, unbalanced by gravity, begins to slide cautiously down the petal's graceful curve, leaving a wet, sparkling trail before silently dripping back into the water, stirring up a circle of tiny, rapidly spreading ripples. Through the clear water, you can vaguely see the green aquatic plants swaying underwater. The whole picture is extremely rich in details, with delicate light and shadow, creating an extremely tranquil, pure and Zen atmosphere.", + "seed": 20631, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_0/index_4937_seed_20631_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.36146279973909257, + "vlm_logprob": -0.36146279973909257, + "vlm_yes_no_margin": 1.8, + "vlm_token_credit": -0.36146279973909257, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007288177963346243 + ], + "opposite_logprob_tokens": [ + -6.632287979125977 + ], + "logprob_reward": -0.007288177963346243, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.007288177963346243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09909562021493912, + 0.08577430993318558, + 0.05095516890287399, + 0.04656863585114479, + 0.05460401996970177, + 0.084124356508255, + 0.1522465944290161, + 0.07412294298410416, + 0.2884311079978943, + 0.06407725811004639 + ], + "credit_map": [ + 0.01739116571843624, + 0.012024610303342342, + 0.014212390407919884, + 0.010279044508934021, + 0.010274254716932774, + 0.00860452651977539, + 0.007757043931633234, + 0.018552569672465324, + 0.014033786021173, + 0.008371053263545036, + 0.014828933402895927, + 0.010080614127218723, + 0.00994507223367691, + 0.01168037485331297, + 0.008096533827483654, + 0.008737930096685886, + 0.009376198053359985, + 0.0042014638893306255, + 0.009209507144987583, + 0.00521491514518857, + 0.004566140938550234, + 0.008045490831136703, + 0.005046214442700148, + 0.0052952333353459835, + 0.00563178351148963, + 0.004908638074994087, + 0.00784311629831791, + 0.006058398634195328, + 0.005011423025280237, + 0.005534275434911251, + 0.005587490741163492, + 0.005993502680212259, + 0.010751563124358654, + 0.004145454149693251, + 0.008239613845944405, + 0.008927046321332455, + 0.004108390770852566, + 0.004225973505526781, + 0.0054187895730137825, + 0.008787184022367, + 0.01867559179663658, + 0.009522933512926102, + 0.01274118572473526, + 0.009049489162862301, + 0.010716041550040245, + 0.009166724048554897, + 0.005996359046548605, + 0.008256024681031704, + 0.034802742302417755, + 0.007817481644451618, + 0.024063309654593468, + 0.006951232440769672, + 0.010027933865785599, + 0.018585851415991783, + 0.011855559423565865, + 0.03814246132969856, + 0.012181495316326618, + 0.004883458372205496, + 0.012380249798297882, + 0.004627884831279516, + 0.0075001767836511135, + 0.01005295105278492, + 0.004940202925354242, + 0.01755651645362377, + 0.05108031630516052, + 0.019952470436692238, + 0.056070685386657715, + 0.01728425733745098, + 0.018923718482255936, + 0.04047149419784546, + 0.016314612701535225, + 0.06833353638648987, + 0.008293201215565205, + 0.005339827388525009, + 0.009123447351157665, + 0.00868355855345726, + 0.00590043980628252, + 0.00749571155756712, + 0.008316880092024803, + 0.010924188420176506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.007288177963346243 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38710036873817444 + ], + "opposite_logprob_tokens": [ + -1.137100338935852 + ], + "logprob_reward": -0.38710036873817444, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38710036873817444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11880818754434586, + 0.08336640894412994, + 0.06260892748832703, + 0.032050956040620804, + 0.07752610743045807, + 0.08039070665836334, + 0.2123432606458664, + 0.1136651337146759, + 0.0877075269818306, + 0.13153280317783356 + ], + "credit_map": [ + 0.02741600014269352, + 0.008284780196845531, + 0.020890528336167336, + 0.008626528084278107, + 0.007898791693150997, + 0.01172218844294548, + 0.008849081583321095, + 0.025120286270976067, + 0.010057630017399788, + 0.008167462423443794, + 0.013150849379599094, + 0.0125513207167387, + 0.007510186173021793, + 0.011679372750222683, + 0.008495159447193146, + 0.011754423379898071, + 0.009424902498722076, + 0.005397115368396044, + 0.012736373580992222, + 0.005256244912743568, + 0.005614128895103931, + 0.01220807433128357, + 0.005901942029595375, + 0.006070144008845091, + 0.00535004585981369, + 0.0036547144409269094, + 0.005121935158967972, + 0.0025480277836322784, + 0.00422657560557127, + 0.0040952228009700775, + 0.002872877288609743, + 0.004181555472314358, + 0.01990327052772045, + 0.00501007866114378, + 0.012083329260349274, + 0.009417593479156494, + 0.005521075800061226, + 0.007497574668377638, + 0.007260098587721586, + 0.010833079926669598, + 0.01941925659775734, + 0.005473303608596325, + 0.011619375087320805, + 0.01078482624143362, + 0.006369782146066427, + 0.009790885262191296, + 0.00856039859354496, + 0.008372867479920387, + 0.04608595371246338, + 0.0176259633153677, + 0.036613039672374725, + 0.010718533769249916, + 0.0137754762545228, + 0.028984855860471725, + 0.015139305032789707, + 0.043400123715400696, + 0.011666388250887394, + 0.008338992483913898, + 0.015581239946186543, + 0.01455146074295044, + 0.013716924004256725, + 0.011918852105736732, + 0.018617719411849976, + 0.01927354745566845, + 0.01334718894213438, + 0.007908973842859268, + 0.01554227713495493, + 0.009233003482222557, + 0.007630066946148872, + 0.00935074407607317, + 0.008242943324148655, + 0.016452321782708168, + 0.023052992299199104, + 0.00990445539355278, + 0.0269323717802763, + 0.00861768051981926, + 0.012736854143440723, + 0.01311164628714323, + 0.008834612555801868, + 0.028342176228761673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.38710036873817444 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233404040336609 + ], + "opposite_logprob_tokens": [ + -0.8983404040336609 + ], + "logprob_reward": -0.5233404040336609, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233404040336609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.260570228099823, + 0.1247427836060524, + 0.04790573567152023, + 0.06452297419309616, + 0.040941204875707626, + 0.02546820044517517, + 0.16755647957324982, + 0.10625579953193665, + 0.03286665678024292, + 0.1291699856519699 + ], + "credit_map": [ + 0.033218130469322205, + 0.029818128794431686, + 0.037630595266819, + 0.017667079344391823, + 0.02623300999403, + 0.032293595373630524, + 0.025061804801225662, + 0.05864784121513367, + 0.03021499700844288, + 0.012414396740496159, + 0.02250124141573906, + 0.012197775766253471, + 0.010682638734579086, + 0.01166999340057373, + 0.009191811084747314, + 0.015869922935962677, + 0.00849850568920374, + 0.003968194127082825, + 0.006913976743817329, + 0.004790405742824078, + 0.004342587664723396, + 0.007253326941281557, + 0.005540496204048395, + 0.006598236970603466, + 0.010183573700487614, + 0.005256108473986387, + 0.010556747205555439, + 0.010448182933032513, + 0.005164918024092913, + 0.00643059890717268, + 0.007110011763870716, + 0.009372825734317303, + 0.0072943344712257385, + 0.0037614922039210796, + 0.006537315901368856, + 0.003064287593588233, + 0.004394409712404013, + 0.0047522843815386295, + 0.003699100809171796, + 0.007437974214553833, + 0.003984380047768354, + 0.0021615424193441868, + 0.003761941334232688, + 0.0028528200928121805, + 0.0024164419155567884, + 0.0035146342124789953, + 0.002217856701463461, + 0.00455858139321208, + 0.035277217626571655, + 0.012806971557438374, + 0.027395186945796013, + 0.010620234534144402, + 0.00974646583199501, + 0.021712101995944977, + 0.010119317099452019, + 0.0398789681494236, + 0.018891800194978714, + 0.008555591106414795, + 0.017905661836266518, + 0.006695169024169445, + 0.007448380347341299, + 0.011207054369151592, + 0.007137335371226072, + 0.028414800763130188, + 0.004361344967037439, + 0.003842140082269907, + 0.005485875066369772, + 0.003429036820307374, + 0.003103015711531043, + 0.0035660648718476295, + 0.003483963431790471, + 0.005595215130597353, + 0.02897401712834835, + 0.010067066177725792, + 0.02333204634487629, + 0.007426961790770292, + 0.012468350119888783, + 0.014233172871172428, + 0.009987005032598972, + 0.022681357339024544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.5233404040336609 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31345275044441223 + ], + "opposite_logprob_tokens": [ + -1.3134527206420898 + ], + "logprob_reward": -0.31345275044441223, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31345275044441223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16556337475776672, + 0.08180367201566696, + 0.06916919350624084, + 0.029270542785525322, + 0.06207292154431343, + 0.06209367886185646, + 0.0689237043261528, + 0.08127150684595108, + 0.11897817254066467, + 0.26085323095321655 + ], + "credit_map": [ + 0.026697993278503418, + 0.016749894246459007, + 0.025079842656850815, + 0.016877256333827972, + 0.015614137053489685, + 0.017117325216531754, + 0.015018380247056484, + 0.03240854665637016, + 0.01248088851571083, + 0.008301019668579102, + 0.01153944805264473, + 0.011616192758083344, + 0.0077255466021597385, + 0.01072164811193943, + 0.008364838548004627, + 0.011054088361561298, + 0.011382904835045338, + 0.0069870734587311745, + 0.010710455477237701, + 0.006323450710624456, + 0.006517244502902031, + 0.010197147727012634, + 0.009199592284858227, + 0.00785132311284542, + 0.004064115695655346, + 0.002321499167010188, + 0.004754709545522928, + 0.0047269235365092754, + 0.0027722411323338747, + 0.0040079327300190926, + 0.0035723410546779633, + 0.00305077969096601, + 0.009873735718429089, + 0.005456203129142523, + 0.010031918063759804, + 0.007991822436451912, + 0.005742242559790611, + 0.00853040162473917, + 0.006431106943637133, + 0.008015490137040615, + 0.010060225613415241, + 0.004408048931509256, + 0.00912154745310545, + 0.009972264990210533, + 0.006064970511943102, + 0.007507101632654667, + 0.006116860080510378, + 0.008842661045491695, + 0.0132145956158638, + 0.006205318029969931, + 0.010103008709847927, + 0.005236602853983641, + 0.0065158274956047535, + 0.007066615857183933, + 0.006091725081205368, + 0.014490010216832161, + 0.01507940050214529, + 0.005944917909801006, + 0.014001846313476562, + 0.004034921992570162, + 0.005958687514066696, + 0.00942212250083685, + 0.005773306358605623, + 0.02105630375444889, + 0.021223191171884537, + 0.0076976646669209, + 0.0192541666328907, + 0.011506786569952965, + 0.009319346398115158, + 0.01684759370982647, + 0.008322960697114468, + 0.024806465953588486, + 0.05430920422077179, + 0.022298453375697136, + 0.05163533613085747, + 0.02757839672267437, + 0.0164786484092474, + 0.027744850143790245, + 0.020040767267346382, + 0.04076758027076721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.31345275044441223 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761322975158691 + ], + "opposite_logprob_tokens": [ + -0.8261322975158691 + ], + "logprob_reward": -0.5761322975158691, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761322975158691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23227441310882568, + 0.27381375432014465, + 0.06064701825380325, + 0.0770593136548996, + 0.033030714839696884, + 0.038072992116212845, + 0.0964580625295639, + 0.036635175347328186, + 0.043752048164606094, + 0.1082565113902092 + ], + "credit_map": [ + 0.03935210779309273, + 0.01782228983938694, + 0.036764442920684814, + 0.01334435772150755, + 0.017766715958714485, + 0.02702462114393711, + 0.016829337924718857, + 0.06337053328752518, + 0.06653541326522827, + 0.019232043996453285, + 0.05245039984583855, + 0.01878480054438114, + 0.019631268456578255, + 0.027588771656155586, + 0.01659082993865013, + 0.05300022289156914, + 0.01857101544737816, + 0.003939615096896887, + 0.010757937096059322, + 0.006445088889449835, + 0.0036088440101593733, + 0.005695736035704613, + 0.004736623726785183, + 0.006892154924571514, + 0.021497102454304695, + 0.006221161223948002, + 0.011156192980706692, + 0.007937000133097172, + 0.005722225643694401, + 0.006510260049253702, + 0.0059973737224936485, + 0.012018004432320595, + 0.006302942987531424, + 0.0036461518611758947, + 0.00513589708134532, + 0.003355829045176506, + 0.003426947630941868, + 0.003861884353682399, + 0.0025911759585142136, + 0.004709884990006685, + 0.0061452570371329784, + 0.00301352608948946, + 0.005401603411883116, + 0.005394588224589825, + 0.0038740653544664383, + 0.0050497278571128845, + 0.004899515770375729, + 0.004294707905501127, + 0.01754169352352619, + 0.005802136845886707, + 0.01654144749045372, + 0.0063314917497336864, + 0.006641111336648464, + 0.013896340504288673, + 0.00819466169923544, + 0.02150917612016201, + 0.00901650357991457, + 0.0022077385801821947, + 0.0057823313400149345, + 0.004131515976041555, + 0.0032776729203760624, + 0.0025986081454902887, + 0.003468314651399851, + 0.006152492482215166, + 0.006737302988767624, + 0.003876646514981985, + 0.006717321462929249, + 0.0030945169273763895, + 0.003730383003130555, + 0.0071202595718204975, + 0.0030592544935643673, + 0.009416361339390278, + 0.014821978285908699, + 0.0072785415686666965, + 0.018891476094722748, + 0.007897282019257545, + 0.010867644101381302, + 0.01092495210468769, + 0.009012996219098568, + 0.028561638668179512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.5761322975158691 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot, framed vertically, captures a fully blossomed lotus flower, floating steadily and calmly on the surface of crystal-clear, mirror-like water. The petals, layered and gracefully unfolding, take on a velvety texture. Their delicate pink gradually transitions to a soft, milky white from the tips to the center, nestled within a golden pod. Soft, golden morning sunlight slants through the mist, precisely illuminating the distinct veins of the petals and making the crystal-clear water droplets clinging to them sparkle like tiny diamonds. As the petals sway gently, a plump droplet, unbalanced by gravity, begins to slide cautiously down the petal's graceful curve, leaving a wet, sparkling trail before silently dripping back into the water, stirring up a circle of tiny, rapidly spreading ripples. Through the clear water, you can vaguely see the green aquatic plants swaying underwater. The whole picture is extremely rich in details, with delicate light and shadow, creating an extremely tranquil, pure and Zen atmosphere.", + "seed": 20632, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_0/index_4937_seed_20632_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2854973454028368, + "vlm_logprob": -0.2854973454028368, + "vlm_yes_no_margin": 2.075, + "vlm_token_credit": -0.2854973454028368, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00807780958712101 + ], + "opposite_logprob_tokens": [ + -6.633077621459961 + ], + "logprob_reward": -0.00807780958712101, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.00807780958712101, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04173600673675537, + 0.22099193930625916, + 0.07185742259025574, + 0.055811524391174316, + 0.32851454615592957, + 0.14244094491004944, + 0.026996968314051628, + 0.03275897353887558, + 0.03911872208118439, + 0.03977294638752937 + ], + "credit_map": [ + 0.006368661765009165, + 0.0030396536458283663, + 0.0061601330526173115, + 0.004403674975037575, + 0.005014240741729736, + 0.003529023379087448, + 0.005499313119798899, + 0.007721308618783951, + 0.04479812830686569, + 0.011139133013784885, + 0.03424559161067009, + 0.017472758889198303, + 0.009339146316051483, + 0.024716675281524658, + 0.018999533727765083, + 0.06028096377849579, + 0.01315635722130537, + 0.003659568028524518, + 0.011151593178510666, + 0.005970560014247894, + 0.005489136092364788, + 0.007140773348510265, + 0.00762917660176754, + 0.01766026020050049, + 0.009120468981564045, + 0.0028850967064499855, + 0.008814124390482903, + 0.0043158093467354774, + 0.0050072697922587395, + 0.005067811347544193, + 0.006185541860759258, + 0.014415398240089417, + 0.07367793470621109, + 0.020267868414521217, + 0.055260151624679565, + 0.028632596135139465, + 0.01477307453751564, + 0.02931479550898075, + 0.020692095160484314, + 0.08589602261781693, + 0.04456936940550804, + 0.010906865820288658, + 0.024425093084573746, + 0.009338616393506527, + 0.009698714129626751, + 0.008881433866918087, + 0.008379786275327206, + 0.026241067796945572, + 0.005963461939245462, + 0.0025313247460871935, + 0.004216201603412628, + 0.003373371437191963, + 0.0016122939996421337, + 0.0018871229840442538, + 0.0033510474022477865, + 0.0040621450170874596, + 0.008757475763559341, + 0.0024312424939125776, + 0.005168573465198278, + 0.0024883639998733997, + 0.002315286546945572, + 0.0031274654902517796, + 0.0028556054458022118, + 0.0056149628944695, + 0.008477994240820408, + 0.0028734966181218624, + 0.006952834315598011, + 0.00257542310282588, + 0.003904547542333603, + 0.0037794194649904966, + 0.003912047483026981, + 0.006642959080636501, + 0.009771570563316345, + 0.002874047728255391, + 0.0056341346353292465, + 0.003174793440848589, + 0.002846524817869067, + 0.003230032976716757, + 0.004417589865624905, + 0.007824251428246498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.00807780958712101 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486589193344116 + ], + "opposite_logprob_tokens": [ + -1.2236589193344116 + ], + "logprob_reward": -0.3486589193344116, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486589193344116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03376958519220352, + 0.23746201395988464, + 0.06707116216421127, + 0.02396586164832115, + 0.34595513343811035, + 0.04160473495721817, + 0.1857130378484726, + 0.029033996164798737, + 0.023114684969186783, + 0.01230987161397934 + ], + "credit_map": [ + 0.006069882772862911, + 0.002364154439419508, + 0.005581367760896683, + 0.0020537537056952715, + 0.003978810738772154, + 0.0035530359018594027, + 0.0033871519844979048, + 0.00678142998367548, + 0.04559779167175293, + 0.011786584742367268, + 0.03885657712817192, + 0.01616620272397995, + 0.010744813829660416, + 0.02699405886232853, + 0.01984560489654541, + 0.06747037917375565, + 0.012776254676282406, + 0.0032205432653427124, + 0.009982425719499588, + 0.006564892828464508, + 0.004352858755737543, + 0.006386853288859129, + 0.007502145133912563, + 0.016285184770822525, + 0.008508210070431232, + 0.0013234695652499795, + 0.0037911853287369013, + 0.0022190590389072895, + 0.0012833760119974613, + 0.0016457369783893228, + 0.0017706615617498755, + 0.0034241625107824802, + 0.0723033919930458, + 0.022858107462525368, + 0.05879521369934082, + 0.028575830161571503, + 0.015945380553603172, + 0.03042239509522915, + 0.023212891072034836, + 0.09384191036224365, + 0.006144993472844362, + 0.003968340344727039, + 0.006673609372228384, + 0.0036789041478186846, + 0.005598809570074081, + 0.0038346515502780676, + 0.004478872753679752, + 0.007226549554616213, + 0.0347055047750473, + 0.017337646335363388, + 0.027009528130292892, + 0.02035178616642952, + 0.017486680299043655, + 0.01991000771522522, + 0.018884992226958275, + 0.030026879161596298, + 0.00852868054062128, + 0.0025387934874743223, + 0.004745419602841139, + 0.001524346531368792, + 0.0020975740626454353, + 0.002853610785678029, + 0.0020873206667602062, + 0.004658249672502279, + 0.004790745209902525, + 0.0015477546257898211, + 0.004319928586483002, + 0.0015734145417809486, + 0.002137334318831563, + 0.002474604407325387, + 0.0023495762143284082, + 0.003921326715499163, + 0.003352021099999547, + 0.0008559526759199798, + 0.0016392198158428073, + 0.0008379881037399173, + 0.0010954588651657104, + 0.0009716578060761094, + 0.000987189239822328, + 0.0025703837163746357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.3486589193344116 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039064526557922 + ], + "opposite_logprob_tokens": [ + -1.9103906154632568 + ], + "logprob_reward": -0.16039064526557922, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039064526557922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11484294384717941, + 0.17686296999454498, + 0.04328425973653793, + 0.14604778587818146, + 0.22345469892024994, + 0.16301776468753815, + 0.0414886474609375, + 0.05636367201805115, + 0.015661396086215973, + 0.01897589862346649 + ], + "credit_map": [ + 0.019974196329712868, + 0.007254703436046839, + 0.01844196952879429, + 0.017921587452292442, + 0.011400006711483002, + 0.010290281847119331, + 0.013262848369777203, + 0.016297347843647003, + 0.032170072197914124, + 0.008555857464671135, + 0.0275868009775877, + 0.011491755954921246, + 0.00891085620969534, + 0.020785748958587646, + 0.016168341040611267, + 0.05119352415204048, + 0.011556738056242466, + 0.003153054276481271, + 0.0063911546021699905, + 0.003197090234607458, + 0.005960430018603802, + 0.003672076854854822, + 0.0038800539914518595, + 0.005473664030432701, + 0.02408226951956749, + 0.007680912967771292, + 0.02159797213971615, + 0.01086224801838398, + 0.009028084576129913, + 0.01520448550581932, + 0.015187692828476429, + 0.04240412637591362, + 0.051063422113657, + 0.013260774314403534, + 0.038650993257761, + 0.0180065780878067, + 0.009396122768521309, + 0.020078446716070175, + 0.014383522793650627, + 0.05861484259366989, + 0.04357319697737694, + 0.011262331157922745, + 0.02826782874763012, + 0.011588864959776402, + 0.0128621906042099, + 0.012190653011202812, + 0.011854207143187523, + 0.03141848370432854, + 0.012633158825337887, + 0.0030445081647485495, + 0.0064999982714653015, + 0.0033535975962877274, + 0.003060680814087391, + 0.004230504389852285, + 0.0033532497473061085, + 0.005312950350344181, + 0.012787088751792908, + 0.005103317555040121, + 0.008982649073004723, + 0.003975167404860258, + 0.0048275175504386425, + 0.005156958010047674, + 0.006007703952491283, + 0.009523271583020687, + 0.0029274853877723217, + 0.00151573377661407, + 0.002599874045699835, + 0.0018844034057110548, + 0.0011442368850111961, + 0.001450766110792756, + 0.001887288992293179, + 0.0022516087628901005, + 0.0037563899531960487, + 0.0012917299754917622, + 0.003131198463961482, + 0.0011873227776959538, + 0.0010172007605433464, + 0.0015694715548306704, + 0.00223362073302269, + 0.0047889635898172855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.16039064526557922 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870651423931122 + ], + "opposite_logprob_tokens": [ + -1.1370651721954346 + ], + "logprob_reward": -0.3870651423931122, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870651423931122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1565379947423935, + 0.06788342446088791, + 0.12892524898052216, + 0.06004505977034569, + 0.24618397653102875, + 0.11490565538406372, + 0.029994254931807518, + 0.12449505925178528, + 0.052821893244981766, + 0.018207449465990067 + ], + "credit_map": [ + 0.026545794680714607, + 0.007494215387851, + 0.024301977828145027, + 0.012458872981369495, + 0.01427993830293417, + 0.015348896384239197, + 0.01798480749130249, + 0.03812350332736969, + 0.009301970712840557, + 0.0034453431144356728, + 0.011487413197755814, + 0.004280059598386288, + 0.005581318400800228, + 0.008400061167776585, + 0.0064417277462780476, + 0.018945535644888878, + 0.025018231943249702, + 0.006526728626340628, + 0.018029388040304184, + 0.011615738272666931, + 0.008317756466567516, + 0.01264350675046444, + 0.012198285199701786, + 0.03457561880350113, + 0.008961936458945274, + 0.0027909569907933474, + 0.009556272067129612, + 0.004211326129734516, + 0.0032641084399074316, + 0.006495973095297813, + 0.006669061724096537, + 0.01809542439877987, + 0.05448736995458603, + 0.015118942596018314, + 0.042498502880334854, + 0.018121089786291122, + 0.010680178180336952, + 0.020294856280088425, + 0.016172314062714577, + 0.06881070882081985, + 0.034292902797460556, + 0.009130855090916157, + 0.019402163103222847, + 0.008123917505145073, + 0.007910547778010368, + 0.008404777385294437, + 0.00799099076539278, + 0.019649503752589226, + 0.0060340664349496365, + 0.003193591721355915, + 0.004268035292625427, + 0.0029296334832906723, + 0.004502696450799704, + 0.0022131099831312895, + 0.003738281549885869, + 0.0031148395501077175, + 0.03659752756357193, + 0.010198220610618591, + 0.018232854083180428, + 0.009347223676741123, + 0.009921368211507797, + 0.00982475932687521, + 0.01272535603493452, + 0.017647748813033104, + 0.0077818529680371284, + 0.006098087411373854, + 0.007561182137578726, + 0.00545354001224041, + 0.005738239735364914, + 0.006447922438383102, + 0.005753726698458195, + 0.007987340912222862, + 0.0023923565167933702, + 0.0021593356505036354, + 0.0023370070848613977, + 0.002218075329437852, + 0.0019850144162774086, + 0.002145326929166913, + 0.0022195966448634863, + 0.0027507354971021414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.3870651423931122 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a fully blossomed lotus flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and is the video framed vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lotus flower floating on crystal-clear, mirror-like water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a plump water droplet slide down a petal and create ripples in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are green aquatic plants visible swaying underwater through the clear water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.52329421043396 + ], + "opposite_logprob_tokens": [ + -0.89829421043396 + ], + "logprob_reward": -0.52329421043396, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.52329421043396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09785331785678864, + 0.36654484272003174, + 0.09104864299297333, + 0.034304700791835785, + 0.1401054561138153, + 0.05376748740673065, + 0.0458231046795845, + 0.07082296907901764, + 0.014043688774108887, + 0.0856858491897583 + ], + "credit_map": [ + 0.01738862320780754, + 0.003983433824032545, + 0.01599959284067154, + 0.0064395759254693985, + 0.005283431615680456, + 0.00786895677447319, + 0.01102849468588829, + 0.0298612043261528, + 0.0776328295469284, + 0.01785353012382984, + 0.05898723751306534, + 0.03006679005920887, + 0.015265547670423985, + 0.03965712711215019, + 0.028942028060555458, + 0.09813973307609558, + 0.019803939387202263, + 0.004544680006802082, + 0.014184779487550259, + 0.00626608869060874, + 0.004761094693094492, + 0.007830136455595493, + 0.010113416239619255, + 0.02354450151324272, + 0.005582035519182682, + 0.0017924251733347774, + 0.005900283809751272, + 0.004394577816128731, + 0.001726037124171853, + 0.0024366758298128843, + 0.0036327317357063293, + 0.008839931339025497, + 0.027761666104197502, + 0.00910915993154049, + 0.024544980376958847, + 0.01008264534175396, + 0.006736159790307283, + 0.012577986344695091, + 0.01010322105139494, + 0.03918962925672531, + 0.014741172082722187, + 0.0034344904124736786, + 0.008860202506184578, + 0.003878355026245117, + 0.0035485373809933662, + 0.003405526513233781, + 0.003587293904274702, + 0.01231190375983715, + 0.014399331994354725, + 0.003769797505810857, + 0.006990175694227219, + 0.0033960435539484024, + 0.003424315480515361, + 0.004241370130330324, + 0.003929350525140762, + 0.005672713275998831, + 0.017474720254540443, + 0.005058633629232645, + 0.01257516909390688, + 0.005039437208324671, + 0.00479522068053484, + 0.006473139394074678, + 0.006798139773309231, + 0.012608499266207218, + 0.002113525988534093, + 0.0010909998090937734, + 0.0024477909319102764, + 0.0016924010124057531, + 0.0011598662240430713, + 0.001593513647094369, + 0.002110825851559639, + 0.0018347641453146935, + 0.024054566398262978, + 0.006031878292560577, + 0.01156858541071415, + 0.004884324502199888, + 0.004864409565925598, + 0.007207332644611597, + 0.0069498540833592415, + 0.020124895498156548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.52329421043396 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_1/index_tensor([7260, 7260])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_1/index_tensor([7260, 7260])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..73d9c275804aed7f97336f64fa50c606fb8dade7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_1/index_tensor([7260, 7260])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "grand, static close-up shot, shot from a slightly upward angle, captures the transcendent moment when the end of the dancing auroral band meets the summit of a solitary snowy mountain. Like a flowing cosmic silk, the translucent curtain of vibrant emerald green, edged with wisps of magenta and lavender, sweeps gently, almost caressingly, across the peak's highest point with a slow, graceful movement. The summit itself is a blade-like rock, blanketed in a thick, untouched layer of powdery snow, its surface delicately undulating and textured by the cold wind. As the aurora's light streams past, the snow grains are momentarily illuminated, reflecting a crystalline glow and casting fleeting, dreamy colors across the pristine white. Behind them is the inky night sky, dotted with countless clearly visible, twinkling stars like cold stars. The whole picture, dominated by cold colors, is full of otherworldly tranquility and grandeur, creating an awe-inspiring, pure and cold cosmic atmosphere.", + "seed": 18685, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_1/index_7260_seed_18685_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2117087810765952, + "vlm_logprob": -0.2117087810765952, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.2117087810765952, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07296378165483475 + ], + "opposite_logprob_tokens": [ + -2.8229637145996094 + ], + "logprob_reward": -0.07296378165483475, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07296378165483475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4197706878185272, + 0.06409508734941483, + 0.030128981918096542, + 0.037385664880275726, + 0.08555322885513306, + 0.042122963815927505, + 0.11613395810127258, + 0.050469353795051575, + 0.08138494193553925, + 0.07295503467321396 + ], + "credit_map": [ + 0.0460767038166523, + 0.07917418330907822, + 0.041142795234918594, + 0.04144187271595001, + 0.04524051770567894, + 0.059315573424100876, + 0.062054481357336044, + 0.04532460495829582, + 0.007190414238721132, + 0.01722891628742218, + 0.009034653194248676, + 0.004492254462093115, + 0.007428532466292381, + 0.005768746603280306, + 0.0073270355351269245, + 0.00562454666942358, + 0.0026664684992283583, + 0.0051669650711119175, + 0.002835965948179364, + 0.0029889470897614956, + 0.0035162067506462336, + 0.005964822601526976, + 0.0035152321215718985, + 0.003474376630038023, + 0.005519516300410032, + 0.00817992351949215, + 0.0022941993083804846, + 0.0037356424145400524, + 0.004407145548611879, + 0.004098344128578901, + 0.005194627679884434, + 0.003956271335482597, + 0.010329457931220531, + 0.01683712564408779, + 0.005346908234059811, + 0.005204319022595882, + 0.010266177356243134, + 0.015781162306666374, + 0.014969767071306705, + 0.006818314548581839, + 0.005682973191142082, + 0.01001633983105421, + 0.003719101659953594, + 0.0022213179618120193, + 0.004201594274491072, + 0.00794090423732996, + 0.004746542312204838, + 0.003594196168705821, + 0.012914672493934631, + 0.0220414437353611, + 0.009069523774087429, + 0.009676719084382057, + 0.014971609227359295, + 0.016354555264115334, + 0.015417114831507206, + 0.01568833366036415, + 0.004523433744907379, + 0.006102468818426132, + 0.005251679103821516, + 0.0035297453869134188, + 0.006194002460688353, + 0.010418778285384178, + 0.008647274225950241, + 0.0058019752614200115, + 0.008253303356468678, + 0.015701958909630775, + 0.006703877821564674, + 0.006808426231145859, + 0.011181003414094448, + 0.014285994693636894, + 0.010136359371244907, + 0.008314026519656181, + 0.008500722236931324, + 0.013553797267377377, + 0.007590628694742918, + 0.008264303207397461, + 0.00659620575606823, + 0.010716177523136139, + 0.008771765977144241, + 0.008961440995335579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.07296378165483475 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029817644506692886 + ], + "opposite_logprob_tokens": [ + -3.529817581176758 + ], + "logprob_reward": -0.029817644506692886, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029817644506692886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36848199367523193, + 0.04565592482686043, + 0.08588062226772308, + 0.04917249083518982, + 0.049813199788331985, + 0.03930596634745598, + 0.08108891546726227, + 0.11528496444225311, + 0.029206030070781708, + 0.13610991835594177 + ], + "credit_map": [ + 0.040071748197078705, + 0.04329899698495865, + 0.05748952925205231, + 0.038567207753658295, + 0.0374443419277668, + 0.0552472360432148, + 0.049454234540462494, + 0.0469086691737175, + 0.006537808105349541, + 0.005443450529128313, + 0.0068201967515051365, + 0.004540897440165281, + 0.004695592448115349, + 0.0054151201620697975, + 0.005955642554908991, + 0.0062472098506987095, + 0.011583288200199604, + 0.012009085156023502, + 0.008263049647212029, + 0.008377371355891228, + 0.011950384825468063, + 0.01171355415135622, + 0.010355870239436626, + 0.011628003790974617, + 0.00685189850628376, + 0.010893142782151699, + 0.004551164340227842, + 0.003658698871731758, + 0.004408746026456356, + 0.007542790379375219, + 0.006937053520232439, + 0.004328995011746883, + 0.006581554654985666, + 0.007285271305590868, + 0.003958112094551325, + 0.00418785959482193, + 0.006504057440906763, + 0.008945774286985397, + 0.006537885405123234, + 0.005812677554786205, + 0.006771388929337263, + 0.007117241155356169, + 0.003097941167652607, + 0.0030661760829389095, + 0.004194046836346388, + 0.004443803336471319, + 0.006372427102178335, + 0.004242934752255678, + 0.009115567430853844, + 0.015830427408218384, + 0.007001587655395269, + 0.007402271032333374, + 0.009625964798033237, + 0.01268570963293314, + 0.009643814526498318, + 0.009783558547496796, + 0.012913621962070465, + 0.009201103821396828, + 0.014209258370101452, + 0.008417771197855473, + 0.01092544011771679, + 0.02398362010717392, + 0.018513837829232216, + 0.01712029054760933, + 0.0026865359395742416, + 0.004184136167168617, + 0.0030940319411456585, + 0.0015805416041985154, + 0.005211683455854654, + 0.00559384236112237, + 0.00423657288774848, + 0.0026186818722635508, + 0.022949593141674995, + 0.022814560681581497, + 0.01030468288809061, + 0.011313528753817081, + 0.009910030290484428, + 0.01716369390487671, + 0.02031736448407173, + 0.021336456760764122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.029817644506692886 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002490278100594878 + ], + "opposite_logprob_tokens": [ + -6.002490043640137 + ], + "logprob_reward": -0.002490278100594878, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002490278100594878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5741537809371948, + 0.03482496738433838, + 0.030955197289586067, + 0.028326911851763725, + 0.08096076548099518, + 0.04646636173129082, + 0.03439698740839958, + 0.03633074462413788, + 0.11373341083526611, + 0.01985093764960766 + ], + "credit_map": [ + 0.0656997412443161, + 0.07308651506900787, + 0.08609098941087723, + 0.0693468526005745, + 0.04755542427301407, + 0.0677846148610115, + 0.07912171632051468, + 0.08546790480613708, + 0.0039322893135249615, + 0.006959307938814163, + 0.004039469640702009, + 0.0031433263793587685, + 0.0032987059094011784, + 0.003838498378172517, + 0.004489860497415066, + 0.005123508162796497, + 0.004794934298843145, + 0.005912827793508768, + 0.0020921886898577213, + 0.0023903590627014637, + 0.00443315552547574, + 0.006235223729163408, + 0.003025797661393881, + 0.0020707128569483757, + 0.0038766590878367424, + 0.006640378385782242, + 0.0024447734467685223, + 0.0017593071097508073, + 0.002287032315507531, + 0.004209079779684544, + 0.004386011511087418, + 0.0027236726600676775, + 0.0129181994125247, + 0.026929005980491638, + 0.002287433948367834, + 0.003754580859094858, + 0.0056521473452448845, + 0.005494079552590847, + 0.018039507791399956, + 0.0058858054690063, + 0.006561898160725832, + 0.004470811691135168, + 0.0026154632214456797, + 0.00339053338393569, + 0.004239148925989866, + 0.006908205803483725, + 0.012806665152311325, + 0.005473634228110313, + 0.0038980012759566307, + 0.006496993824839592, + 0.0025006288196891546, + 0.002589432755485177, + 0.004919884726405144, + 0.0040415870025753975, + 0.005252051632851362, + 0.004698406904935837, + 0.0029971518088132143, + 0.005364502780139446, + 0.003171902848407626, + 0.002058811718598008, + 0.004102342762053013, + 0.006505904253572226, + 0.006992954295128584, + 0.005137174390256405, + 0.01185792125761509, + 0.01882762648165226, + 0.010651187039911747, + 0.008369933813810349, + 0.014918670989573002, + 0.020071569830179214, + 0.01639474555850029, + 0.012641754932701588, + 0.002265165327116847, + 0.0026064806152135134, + 0.002647864166647196, + 0.0020533502101898193, + 0.0015759635716676712, + 0.0026472052559256554, + 0.0028036723379045725, + 0.003251234535127878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.002490278100594878 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983279466629028 + ], + "opposite_logprob_tokens": [ + -0.5233279466629028 + ], + "logprob_reward": -0.8983279466629028, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983279466629028, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2749640643596649, + 0.04260030761361122, + 0.257272332906723, + 0.060036227107048035, + 0.18708981573581696, + 0.05502953380346298, + 0.03254649415612221, + 0.035842929035425186, + 0.04230222478508949, + 0.01231605838984251 + ], + "credit_map": [ + 0.026295039802789688, + 0.039984822273254395, + 0.036870911717414856, + 0.03405853733420372, + 0.03103320673108101, + 0.031157679855823517, + 0.040263473987579346, + 0.035300374031066895, + 0.0041935136541724205, + 0.009009651839733124, + 0.004389510024338961, + 0.0030366159044206142, + 0.006278499495238066, + 0.0056719849817454815, + 0.005938563495874405, + 0.004081964027136564, + 0.0405438207089901, + 0.048348318785429, + 0.018500078469514847, + 0.01961495913565159, + 0.046755239367485046, + 0.04017450660467148, + 0.02506043016910553, + 0.018274977803230286, + 0.00897730141878128, + 0.017359010875225067, + 0.004283824935555458, + 0.0028901442419737577, + 0.009700221009552479, + 0.00716549763455987, + 0.005790989845991135, + 0.0038692273665219545, + 0.031439874321222305, + 0.06606075912714005, + 0.005052892025560141, + 0.008207832463085651, + 0.012359999120235443, + 0.01026249397546053, + 0.040429990738630295, + 0.01327595766633749, + 0.010007145814597607, + 0.007868398912250996, + 0.0038957169745117426, + 0.0029880560468882322, + 0.0056199622340500355, + 0.008905595168471336, + 0.010397219099104404, + 0.005347433965653181, + 0.0038261637091636658, + 0.004517749417573214, + 0.003201482119038701, + 0.002929611364379525, + 0.003785771317780018, + 0.006064600311219692, + 0.003925612661987543, + 0.0042954981327056885, + 0.004783076234161854, + 0.00776786869391799, + 0.0028764361049979925, + 0.002346661640331149, + 0.0037063444033265114, + 0.004748228471726179, + 0.005137151572853327, + 0.00447715912014246, + 0.003950267098844051, + 0.006072807591408491, + 0.005141242872923613, + 0.0023363111540675163, + 0.0058112433180212975, + 0.008764035068452358, + 0.0061529274098575115, + 0.004073385614901781, + 0.0017143161967396736, + 0.0016988047864288092, + 0.0014169950736686587, + 0.0015576075529679656, + 0.0009258107747882605, + 0.0019933206494897604, + 0.001328280777670443, + 0.0016809209482744336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.8983279466629028 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494425445795059 + ], + "opposite_logprob_tokens": [ + -2.9299442768096924 + ], + "logprob_reward": -0.05494425445795059, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494425445795059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12373091280460358, + 0.04250975325703621, + 0.10712465643882751, + 0.02069338783621788, + 0.23069408535957336, + 0.1381584107875824, + 0.08690348267555237, + 0.06709548830986023, + 0.16235536336898804, + 0.020734453573822975 + ], + "credit_map": [ + 0.013256354257464409, + 0.017039554193615913, + 0.01717967726290226, + 0.016229920089244843, + 0.010581376031041145, + 0.013675170950591564, + 0.01802860014140606, + 0.017740268260240555, + 0.005164519418030977, + 0.007937940768897533, + 0.004697158467024565, + 0.0037521368358284235, + 0.003753264434635639, + 0.004889166913926601, + 0.005553401540964842, + 0.0067621697671711445, + 0.016161827370524406, + 0.020317845046520233, + 0.007765521761029959, + 0.00870610773563385, + 0.015781641006469727, + 0.023024801164865494, + 0.007916011847555637, + 0.007450919132679701, + 0.002688867039978504, + 0.0033578197471797466, + 0.0018848632462322712, + 0.0014174457173794508, + 0.0018990696407854557, + 0.0034746993333101273, + 0.0023890577722340822, + 0.0035815685987472534, + 0.037671610713005066, + 0.0788942202925682, + 0.006430131383240223, + 0.010327312164008617, + 0.015955476090312004, + 0.015413117595016956, + 0.04992612823843956, + 0.016076091676950455, + 0.022496627643704414, + 0.01585693284869194, + 0.008100568316876888, + 0.008729974739253521, + 0.013612835668027401, + 0.019961712881922722, + 0.03428938612341881, + 0.015110380947589874, + 0.010050103068351746, + 0.016014212742447853, + 0.006594348233193159, + 0.007366244215518236, + 0.01133773848414421, + 0.011402218602597713, + 0.012110436335206032, + 0.012028184719383717, + 0.005355822388082743, + 0.009338793344795704, + 0.006824633106589317, + 0.004365185741335154, + 0.007276942487806082, + 0.014274626038968563, + 0.010797222144901752, + 0.008862268179655075, + 0.01737322099506855, + 0.028818972408771515, + 0.014642945490777493, + 0.01254382822662592, + 0.020261602476239204, + 0.02905344031751156, + 0.021587911993265152, + 0.01807345077395439, + 0.002907276852056384, + 0.003109778743237257, + 0.002576486673206091, + 0.0023451014421880245, + 0.002287916839122772, + 0.0027787750586867332, + 0.002255957340821624, + 0.002473162952810526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.05494425445795059 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "grand, static close-up shot, shot from a slightly upward angle, captures the transcendent moment when the end of the dancing auroral band meets the summit of a solitary snowy mountain. Like a flowing cosmic silk, the translucent curtain of vibrant emerald green, edged with wisps of magenta and lavender, sweeps gently, almost caressingly, across the peak's highest point with a slow, graceful movement. The summit itself is a blade-like rock, blanketed in a thick, untouched layer of powdery snow, its surface delicately undulating and textured by the cold wind. As the aurora's light streams past, the snow grains are momentarily illuminated, reflecting a crystalline glow and casting fleeting, dreamy colors across the pristine white. Behind them is the inky night sky, dotted with countless clearly visible, twinkling stars like cold stars. The whole picture, dominated by cold colors, is full of otherworldly tranquility and grandeur, creating an awe-inspiring, pure and cold cosmic atmosphere.", + "seed": 18686, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_1/index_7260_seed_18686_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14330528918653726, + "vlm_logprob": -0.14330528918653726, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.14330528918653726, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.44231942296028137 + ], + "opposite_logprob_tokens": [ + -1.067319393157959 + ], + "logprob_reward": -0.44231942296028137, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.44231942296028137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13215915858745575, + 0.03407564014196396, + 0.063329316675663, + 0.22119927406311035, + 0.14732253551483154, + 0.16874665021896362, + 0.08773404359817505, + 0.060836415737867355, + 0.02470467798411846, + 0.059892334043979645 + ], + "credit_map": [ + 0.015245369635522366, + 0.01750396005809307, + 0.01350423600524664, + 0.009622332639992237, + 0.025243721902370453, + 0.035110361874103546, + 0.00907021202147007, + 0.006858955603092909, + 0.003673379309475422, + 0.007373449858278036, + 0.004374604672193527, + 0.0029098871164023876, + 0.00462310342118144, + 0.002976128598675132, + 0.003975441679358482, + 0.004169646184891462, + 0.007559108082205057, + 0.009007008746266365, + 0.009391017258167267, + 0.00946041289716959, + 0.008170010522007942, + 0.005484468769282103, + 0.006524549797177315, + 0.007732733618468046, + 0.0329059362411499, + 0.043261390179395676, + 0.03182603046298027, + 0.01730833202600479, + 0.023835284635424614, + 0.0446850061416626, + 0.012132879346609116, + 0.015244406647980213, + 0.027025490999221802, + 0.016072923317551613, + 0.013925697654485703, + 0.009483993984758854, + 0.03976258262991905, + 0.015659859403967857, + 0.015049335546791553, + 0.010342643596231937, + 0.011642147786915302, + 0.038040149956941605, + 0.01579079031944275, + 0.009758670814335346, + 0.047055669128894806, + 0.02594514936208725, + 0.01053109671920538, + 0.009982951916754246, + 0.006712724454700947, + 0.030130064114928246, + 0.003342231037095189, + 0.003920435905456543, + 0.027171703055500984, + 0.008868302218616009, + 0.003765635658055544, + 0.0038229404017329216, + 0.007601052522659302, + 0.012384560890495777, + 0.007831855677068233, + 0.005796608980745077, + 0.009599948301911354, + 0.007353748194873333, + 0.005593650043010712, + 0.004674984607845545, + 0.003710254328325391, + 0.0031802088487893343, + 0.0023543071001768112, + 0.0016851795371621847, + 0.0036063247825950384, + 0.005201099906116724, + 0.00248415837995708, + 0.0024831434711813927, + 0.013127189129590988, + 0.015206195414066315, + 0.009525015018880367, + 0.005086726509034634, + 0.004080186132341623, + 0.0032183132134377956, + 0.005874256137758493, + 0.0037744520232081413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.44231942296028137 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088181287050247 + ], + "opposite_logprob_tokens": [ + -4.5110883712768555 + ], + "logprob_reward": -0.011088181287050247, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088181287050247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1545807123184204, + 0.06943132728338242, + 0.03785526752471924, + 0.059808213263750076, + 0.20912007987499237, + 0.047876160591840744, + 0.14014960825443268, + 0.16858665645122528, + 0.06541258841753006, + 0.047179386019706726 + ], + "credit_map": [ + 0.011249658651649952, + 0.02194155380129814, + 0.012057182379066944, + 0.008519342169165611, + 0.044788848608732224, + 0.03602317348122597, + 0.011780879460275173, + 0.008220067247748375, + 0.010363761335611343, + 0.01420971192419529, + 0.007532812189310789, + 0.006608044262975454, + 0.008853279985487461, + 0.006161499302834272, + 0.008825208060443401, + 0.006877008359879255, + 0.0046255202032625675, + 0.006766362115740776, + 0.004325085319578648, + 0.004246207885444164, + 0.004613697994500399, + 0.003044083947315812, + 0.0043805562891066074, + 0.00585375539958477, + 0.006213178392499685, + 0.009999626316130161, + 0.009600256569683552, + 0.005175115540623665, + 0.0090560931712389, + 0.010087334550917149, + 0.0044761947356164455, + 0.005200413521379232, + 0.041456110775470734, + 0.028484124690294266, + 0.014767690561711788, + 0.009851189330220222, + 0.06441476941108704, + 0.02039150334894657, + 0.016792526468634605, + 0.012962175533175468, + 0.0025898870080709457, + 0.008723504841327667, + 0.004722287878394127, + 0.0055422717705369, + 0.012184027582406998, + 0.00525816110894084, + 0.005134002771228552, + 0.003722023917362094, + 0.006784258875995874, + 0.05489613860845566, + 0.00335184158757329, + 0.00412721699103713, + 0.05477235093712807, + 0.007067233324050903, + 0.0040883952751755714, + 0.005062175914645195, + 0.021401403471827507, + 0.05886520817875862, + 0.009424551390111446, + 0.005106532946228981, + 0.013824095949530602, + 0.049536675214767456, + 0.005123356357216835, + 0.005304848775267601, + 0.01588885858654976, + 0.007533881347626448, + 0.007760294247418642, + 0.004153419751673937, + 0.007489532232284546, + 0.011898993514478207, + 0.005627602804452181, + 0.0050600143149495125, + 0.009275288321077824, + 0.013608664274215698, + 0.004334538709372282, + 0.0038961933460086584, + 0.004638780374079943, + 0.0030704988166689873, + 0.005670314189046621, + 0.002685111714527011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.011088181287050247 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002491348423063755 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.002491348423063755, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002491348423063755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11163174360990524, + 0.06221003085374832, + 0.05072773993015289, + 0.06580853462219238, + 0.3940492868423462, + 0.12395910173654556, + 0.06853073835372925, + 0.04660199582576752, + 0.04307488352060318, + 0.033405911177396774 + ], + "credit_map": [ + 0.010737083852291107, + 0.016476590186357498, + 0.009698339737951756, + 0.006247317884117365, + 0.02293701469898224, + 0.025294817984104156, + 0.013171583414077759, + 0.007068993523716927, + 0.012515194714069366, + 0.016528356820344925, + 0.006702513433992863, + 0.0035039884969592094, + 0.007292739115655422, + 0.005409437231719494, + 0.004605412483215332, + 0.005652397871017456, + 0.006613723002374172, + 0.005343414843082428, + 0.006840210407972336, + 0.007502995431423187, + 0.004247850272804499, + 0.0041407011449337006, + 0.010358893312513828, + 0.0056799533776938915, + 0.009144985117018223, + 0.012690963223576546, + 0.006113516166806221, + 0.005474052857607603, + 0.009009389206767082, + 0.015701327472925186, + 0.004338067024946213, + 0.0033362414687871933, + 0.07259537279605865, + 0.04887251928448677, + 0.02875181846320629, + 0.018441587686538696, + 0.12054505944252014, + 0.044015735387802124, + 0.03515946492552757, + 0.02566777914762497, + 0.00594304921105504, + 0.03489919751882553, + 0.0065871044062078, + 0.0065060933120548725, + 0.044305913150310516, + 0.015393483452498913, + 0.005163648631423712, + 0.0051606204360723495, + 0.004678563680499792, + 0.026090754196047783, + 0.0020281996112316847, + 0.0018655955791473389, + 0.024168428033590317, + 0.005270869005471468, + 0.002324808156117797, + 0.0021035291720181704, + 0.006085898727178574, + 0.015875842422246933, + 0.0038089791778475046, + 0.0024376350920647383, + 0.006841681431978941, + 0.006912398152053356, + 0.002653560135513544, + 0.001986004179343581, + 0.005893816705793142, + 0.004689977969974279, + 0.004862385336309671, + 0.0032897177152335644, + 0.006561745423823595, + 0.007138695102185011, + 0.006247092504054308, + 0.004391455557197332, + 0.0063413833267986774, + 0.0037633723113685846, + 0.007383846677839756, + 0.003914203029125929, + 0.0024259118363261223, + 0.002659132471308112, + 0.0041756038554012775, + 0.0027424609288573265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.002491348423063755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034261882305145 + ], + "opposite_logprob_tokens": [ + -1.9103425741195679 + ], + "logprob_reward": -0.16034261882305145, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034261882305145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16342966258525848, + 0.07758476585149765, + 0.11363234370946884, + 0.06989319622516632, + 0.15545998513698578, + 0.12327103316783905, + 0.1172320544719696, + 0.0596950426697731, + 0.037619173526763916, + 0.08218274265527725 + ], + "credit_map": [ + 0.019400080665946007, + 0.034212540835142136, + 0.019144585356116295, + 0.015668636187911034, + 0.022483043372631073, + 0.020013047382235527, + 0.02113310433924198, + 0.011374621652066708, + 0.017349399626255035, + 0.018318742513656616, + 0.008160117082297802, + 0.004817435052245855, + 0.009276110678911209, + 0.007003180682659149, + 0.006509450729936361, + 0.006150326691567898, + 0.01246571633964777, + 0.016209207475185394, + 0.014832993969321251, + 0.01853369176387787, + 0.011910851113498211, + 0.008647819049656391, + 0.019153034314513206, + 0.011879032477736473, + 0.011053548194468021, + 0.014451978728175163, + 0.005909824278205633, + 0.005846045445650816, + 0.007534258998930454, + 0.017188705503940582, + 0.004231634549796581, + 0.0036771991290152073, + 0.014742456376552582, + 0.027335286140441895, + 0.012165436521172523, + 0.009373397566378117, + 0.05090024694800377, + 0.015548854134976864, + 0.014297730289399624, + 0.011096586473286152, + 0.0070955026894807816, + 0.028154263272881508, + 0.008696814067661762, + 0.007854417897760868, + 0.036175694316625595, + 0.019854076206684113, + 0.007252923678606749, + 0.008187341503798962, + 0.007355057168751955, + 0.04647330939769745, + 0.0027180693577975035, + 0.0032148114405572414, + 0.04586881026625633, + 0.005467213224619627, + 0.002950859721750021, + 0.0031839259900152683, + 0.007780909538269043, + 0.013253501616418362, + 0.004347315523773432, + 0.005301443859934807, + 0.010459869168698788, + 0.010336458683013916, + 0.0050904275849461555, + 0.0031251185573637486, + 0.004799291491508484, + 0.00849861279129982, + 0.0034513925202190876, + 0.002485499018803239, + 0.005724817980080843, + 0.006149455439299345, + 0.003939356189221144, + 0.0025707499589771032, + 0.020467406138777733, + 0.02484998293220997, + 0.010747281834483147, + 0.005928205791860819, + 0.005291110370308161, + 0.004408277105540037, + 0.006745524238795042, + 0.003744958434253931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.16034261882305145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1002848744392395 + ], + "opposite_logprob_tokens": [ + -2.3502848148345947 + ], + "logprob_reward": -0.1002848744392395, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002848744392395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10796121507883072, + 0.0564720518887043, + 0.06441166251897812, + 0.13434098660945892, + 0.18007132411003113, + 0.13338494300842285, + 0.07966338843107224, + 0.15383130311965942, + 0.051119741052389145, + 0.03874336928129196 + ], + "credit_map": [ + 0.011714857071638107, + 0.019255757331848145, + 0.008901412598788738, + 0.006677308119833469, + 0.024813229218125343, + 0.02196485549211502, + 0.00878880824893713, + 0.005844986066222191, + 0.00862917210906744, + 0.012064655311405659, + 0.008385812863707542, + 0.00519960280507803, + 0.006853602826595306, + 0.004726348910480738, + 0.00596968550235033, + 0.004643171560019255, + 0.008619219996035099, + 0.011678686365485191, + 0.008037804625928402, + 0.007123443763703108, + 0.008529906161129475, + 0.005491943098604679, + 0.00744678033515811, + 0.0074838800355792046, + 0.016190141439437866, + 0.029066070914268494, + 0.01045954693108797, + 0.011829295195639133, + 0.016780944541096687, + 0.0323142409324646, + 0.010183684527873993, + 0.007517063058912754, + 0.02590440958738327, + 0.04458538442850113, + 0.0148716876283288, + 0.009896489791572094, + 0.039463967084884644, + 0.02253449708223343, + 0.012524578720331192, + 0.010290302336215973, + 0.00660152779892087, + 0.03391455486416817, + 0.013023406267166138, + 0.011660627089440823, + 0.03467150777578354, + 0.01136747095733881, + 0.014306825585663319, + 0.007839029654860497, + 0.005722847767174244, + 0.02514827996492386, + 0.003728423733264208, + 0.003964979667216539, + 0.02490857243537903, + 0.006147935055196285, + 0.0051638344302773476, + 0.004878519102931023, + 0.01955058053135872, + 0.05407673120498657, + 0.008778986521065235, + 0.005849168635904789, + 0.015444322489202023, + 0.03662508726119995, + 0.00774613069370389, + 0.005760285072028637, + 0.010570487007498741, + 0.006117436569184065, + 0.0044067311100661755, + 0.004216067027300596, + 0.005775600206106901, + 0.010550598613917828, + 0.005332829896360636, + 0.004149989690631628, + 0.006480033043771982, + 0.006095114629715681, + 0.006260179914534092, + 0.0050707533955574036, + 0.003928617108613253, + 0.0030019127298146486, + 0.005014649126678705, + 0.0028921090997755527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.1002848744392395 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "grand, static close-up shot, shot from a slightly upward angle, captures the transcendent moment when the end of the dancing auroral band meets the summit of a solitary snowy mountain. Like a flowing cosmic silk, the translucent curtain of vibrant emerald green, edged with wisps of magenta and lavender, sweeps gently, almost caressingly, across the peak's highest point with a slow, graceful movement. The summit itself is a blade-like rock, blanketed in a thick, untouched layer of powdery snow, its surface delicately undulating and textured by the cold wind. As the aurora's light streams past, the snow grains are momentarily illuminated, reflecting a crystalline glow and casting fleeting, dreamy colors across the pristine white. Behind them is the inky night sky, dotted with countless clearly visible, twinkling stars like cold stars. The whole picture, dominated by cold colors, is full of otherworldly tranquility and grandeur, creating an awe-inspiring, pure and cold cosmic atmosphere.", + "seed": 18687, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_1/index_7260_seed_18687_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1622143366606906, + "vlm_logprob": -0.1622143366606906, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.1622143366606906, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044693198055028915 + ], + "opposite_logprob_tokens": [ + -3.4196932315826416 + ], + "logprob_reward": -0.044693198055028915, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.044693198055028915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25237393379211426, + 0.22450736165046692, + 0.07397855818271637, + 0.08101663738489151, + 0.04080034792423248, + 0.03414129838347435, + 0.16442713141441345, + 0.028084706515073776, + 0.05130545422434807, + 0.04936454817652702 + ], + "credit_map": [ + 0.02867153286933899, + 0.04788098856806755, + 0.019543875008821487, + 0.019028576090931892, + 0.031188784167170525, + 0.03277039900422096, + 0.035899415612220764, + 0.03739036247134209, + 0.020404350012540817, + 0.07084440439939499, + 0.018661515787243843, + 0.015802035108208656, + 0.016270102933049202, + 0.045392170548439026, + 0.01903434284031391, + 0.018098441883921623, + 0.0131791727617383, + 0.012186036445200443, + 0.007255417760461569, + 0.0047926330007612705, + 0.005921203643083572, + 0.009677940979599953, + 0.006297415122389793, + 0.014668733812868595, + 0.013750688172876835, + 0.012331264093518257, + 0.00623764842748642, + 0.004616463091224432, + 0.009551690891385078, + 0.010404692962765694, + 0.00729519035667181, + 0.016829002648591995, + 0.005162063054740429, + 0.003295136382803321, + 0.0026941094547510147, + 0.0030034014489501715, + 0.005851157009601593, + 0.0073939040303230286, + 0.0035890531726181507, + 0.0098115224391222, + 0.004619317594915628, + 0.0040320721454918385, + 0.003097214736044407, + 0.001842451049014926, + 0.0054428172297775745, + 0.005241762846708298, + 0.0035197397228330374, + 0.006345919333398342, + 0.021135998889803886, + 0.04575825482606888, + 0.012052119709551334, + 0.010885793715715408, + 0.01555594801902771, + 0.021933335810899734, + 0.011352167464792728, + 0.025753488764166832, + 0.0035722863394767046, + 0.005766913760453463, + 0.001856583054177463, + 0.0019839475862681866, + 0.004152114037424326, + 0.004593613091856241, + 0.003630826948210597, + 0.002528419252485037, + 0.008128590881824493, + 0.0039033405482769012, + 0.0032477162312716246, + 0.002621058141812682, + 0.007738994900137186, + 0.012553883716464043, + 0.0053933653980493546, + 0.007718500681221485, + 0.008695919066667557, + 0.0047581614926457405, + 0.004211928229779005, + 0.003006360726431012, + 0.008714207448065281, + 0.007187057752162218, + 0.0054799350909888744, + 0.007310975342988968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.044693198055028915 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062063299119472504 + ], + "opposite_logprob_tokens": [ + -2.812063217163086 + ], + "logprob_reward": -0.062063299119472504, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062063299119472504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3315954804420471, + 0.07692248374223709, + 0.03737124800682068, + 0.03484981507062912, + 0.08471456170082092, + 0.03344172611832619, + 0.20588289201259613, + 0.0440533272922039, + 0.031297095119953156, + 0.1198713630437851 + ], + "credit_map": [ + 0.03548869490623474, + 0.05747338384389877, + 0.025473838672041893, + 0.029212726280093193, + 0.04157881438732147, + 0.04045587033033371, + 0.04187564551830292, + 0.06003650650382042, + 0.009912583976984024, + 0.013881287537515163, + 0.006056420039385557, + 0.006413057912141085, + 0.006116453092545271, + 0.017876964062452316, + 0.007626961916685104, + 0.00903876218944788, + 0.004461588803678751, + 0.005820905789732933, + 0.0026922710239887238, + 0.0032775045838207006, + 0.0032312150578945875, + 0.004748894367367029, + 0.006069500930607319, + 0.007069365121424198, + 0.005111852195113897, + 0.005819403566420078, + 0.002100360346958041, + 0.00253717927262187, + 0.003922082483768463, + 0.003932387568056583, + 0.004793436266481876, + 0.006633110344409943, + 0.00935178343206644, + 0.009610265493392944, + 0.005644989665597677, + 0.007625435013324022, + 0.012176966294646263, + 0.01547857653349638, + 0.008815090171992779, + 0.016011454164981842, + 0.004746028687804937, + 0.0033110398799180984, + 0.0026540944818407297, + 0.0022988207638263702, + 0.004277758300304413, + 0.008083559572696686, + 0.003412862541154027, + 0.0046575628221035, + 0.022944526746869087, + 0.053379397839307785, + 0.01731153577566147, + 0.018405647948384285, + 0.017976989969611168, + 0.01937798783183098, + 0.02003440633416176, + 0.03645238280296326, + 0.004482364282011986, + 0.00756749277934432, + 0.004193541593849659, + 0.0032327964436262846, + 0.006025469861924648, + 0.00814156886190176, + 0.005003899801522493, + 0.005406195763498545, + 0.004657784476876259, + 0.003328824881464243, + 0.002701855730265379, + 0.001866828417405486, + 0.0037073069252073765, + 0.004090052098035812, + 0.005738540086895227, + 0.005205903202295303, + 0.014858990907669067, + 0.009439161978662014, + 0.011842962354421616, + 0.007624411024153233, + 0.01595958322286606, + 0.022870661690831184, + 0.020605964586138725, + 0.016669629141688347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.062063299119472504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015141224721446633 + ], + "opposite_logprob_tokens": [ + -6.501513957977295 + ], + "logprob_reward": -0.0015141224721446633, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015141224721446633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24778062105178833, + 0.06508521735668182, + 0.033521734178066254, + 0.13463209569454193, + 0.10308969020843506, + 0.18328577280044556, + 0.042524222284555435, + 0.04296814277768135, + 0.06359879672527313, + 0.08351368457078934 + ], + "credit_map": [ + 0.02566385082900524, + 0.04685705155134201, + 0.018568718805909157, + 0.02052210457623005, + 0.03365352749824524, + 0.03226952999830246, + 0.03354182094335556, + 0.03670401871204376, + 0.007828982546925545, + 0.015679897740483284, + 0.00500608142465353, + 0.00491350656375289, + 0.005421217530965805, + 0.011922189965844154, + 0.006488519255071878, + 0.007824833504855633, + 0.007604310754686594, + 0.00570070231333375, + 0.002608379116281867, + 0.0021589305251836777, + 0.0025921522174030542, + 0.003477024845778942, + 0.00387972011230886, + 0.005500519648194313, + 0.026084011420607567, + 0.02032691054046154, + 0.007348142564296722, + 0.006801905110478401, + 0.017102614045143127, + 0.014168048277497292, + 0.01049366407096386, + 0.032306820154190063, + 0.011917178519070148, + 0.014153320342302322, + 0.007271554321050644, + 0.008982540108263493, + 0.015552736818790436, + 0.018037036061286926, + 0.011884141713380814, + 0.015291194431483746, + 0.022781595587730408, + 0.01850953884422779, + 0.01425684243440628, + 0.012719687074422836, + 0.023169636726379395, + 0.0455656573176384, + 0.021063122898340225, + 0.025219706818461418, + 0.00454166904091835, + 0.005449578166007996, + 0.003795399097725749, + 0.00474697956815362, + 0.004453999921679497, + 0.005989470053464174, + 0.0061920746229588985, + 0.007355055306106806, + 0.00411693612113595, + 0.0054801469668745995, + 0.003308018436655402, + 0.0029245219193398952, + 0.005583433900028467, + 0.0061428127810359, + 0.008619586937129498, + 0.006792689207941294, + 0.009244791232049465, + 0.004971152637153864, + 0.003524609375745058, + 0.00261023105122149, + 0.010868719778954983, + 0.018273653462529182, + 0.005005655810236931, + 0.009099987335503101, + 0.014827551320195198, + 0.00553600350394845, + 0.007969527505338192, + 0.005510446149855852, + 0.012016615830361843, + 0.011953302659094334, + 0.013126342557370663, + 0.012573905289173126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0015141224721446633 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327958106994629 + ], + "opposite_logprob_tokens": [ + -0.7577958106994629 + ], + "logprob_reward": -0.6327958106994629, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327958106994629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3753737211227417, + 0.11446800827980042, + 0.03833374008536339, + 0.1099667102098465, + 0.07004912942647934, + 0.04445734992623329, + 0.05942852422595024, + 0.08507990092039108, + 0.07278522849082947, + 0.03005770966410637 + ], + "credit_map": [ + 0.046922534704208374, + 0.06484758853912354, + 0.023188458755612373, + 0.02462831325829029, + 0.04604537785053253, + 0.05654522031545639, + 0.04276386275887489, + 0.0704323798418045, + 0.013329343870282173, + 0.019385501742362976, + 0.008336985483765602, + 0.008085680194199085, + 0.007619316689670086, + 0.03092794679105282, + 0.009630981832742691, + 0.017152246087789536, + 0.005133454222232103, + 0.008214985020458698, + 0.002549549797549844, + 0.002633672906085849, + 0.003549836575984955, + 0.005519132595509291, + 0.0033754731994122267, + 0.007357637397944927, + 0.021647516638040543, + 0.018756240606307983, + 0.006859424524009228, + 0.005907208193093538, + 0.012123621068894863, + 0.012167086824774742, + 0.007552096154540777, + 0.02495351992547512, + 0.007781330496072769, + 0.009328157640993595, + 0.0054938169196248055, + 0.006372646428644657, + 0.010987788438796997, + 0.011659220792353153, + 0.008152690716087818, + 0.01027347706258297, + 0.0029354833532124758, + 0.004098732490092516, + 0.0037270814646035433, + 0.0031876300927251577, + 0.007918254472315311, + 0.006734669674187899, + 0.00620135897770524, + 0.009654136374592781, + 0.0070687332190573215, + 0.010101541876792908, + 0.0030673251021653414, + 0.0030036100652068853, + 0.0076771024614572525, + 0.014008120633661747, + 0.007110340055078268, + 0.007391749415546656, + 0.008527050726115704, + 0.013425961136817932, + 0.005376828834414482, + 0.005142122972756624, + 0.012622283771634102, + 0.011835523881018162, + 0.013929037377238274, + 0.014221095480024815, + 0.011600126512348652, + 0.00612033111974597, + 0.003721959888935089, + 0.0026664931792765856, + 0.012130665592849255, + 0.021738169714808464, + 0.005338960327208042, + 0.009468521922826767, + 0.0061042397283017635, + 0.002262273570522666, + 0.002273229416459799, + 0.002046911045908928, + 0.004200974479317665, + 0.004173678811639547, + 0.0046938518062233925, + 0.004302550107240677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.6327958106994629 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000525295734406 + ], + "opposite_logprob_tokens": [ + -2.695005178451538 + ], + "logprob_reward": -0.07000525295734406, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000525295734406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22372961044311523, + 0.09481281042098999, + 0.050211351364851, + 0.044246379286050797, + 0.041051946580410004, + 0.2266179472208023, + 0.07681610435247421, + 0.11438950151205063, + 0.0570712648332119, + 0.07105309516191483 + ], + "credit_map": [ + 0.024570608511567116, + 0.0446782149374485, + 0.01556076668202877, + 0.01621527597308159, + 0.028305841609835625, + 0.02870037965476513, + 0.031086578965187073, + 0.03461194038391113, + 0.009919289499521255, + 0.027715975418686867, + 0.006714865565299988, + 0.006470150779932737, + 0.007985924370586872, + 0.016304807737469673, + 0.006560805253684521, + 0.013140994124114513, + 0.009353258647024632, + 0.011616632342338562, + 0.004738228395581245, + 0.002547897631302476, + 0.005463138688355684, + 0.0071131098084151745, + 0.003107312833890319, + 0.0062717716209590435, + 0.007864546962082386, + 0.007589264772832394, + 0.0038441852666437626, + 0.0030018496327102184, + 0.005406402517110109, + 0.005351230036467314, + 0.003971775062382221, + 0.007217124104499817, + 0.0048293909057974815, + 0.00525676179677248, + 0.0034404483158141375, + 0.004525128286331892, + 0.00523037277162075, + 0.006819929461926222, + 0.004913062322884798, + 0.006036852020770311, + 0.022582709789276123, + 0.027147836983203888, + 0.01602546125650406, + 0.017288420349359512, + 0.029419448226690292, + 0.037841007113456726, + 0.0340329073369503, + 0.04228014871478081, + 0.0072422390803694725, + 0.011374477297067642, + 0.005300944671034813, + 0.005369784776121378, + 0.009384452365338802, + 0.015197371132671833, + 0.011942346580326557, + 0.011004487052559853, + 0.011237312108278275, + 0.018790898844599724, + 0.009044461883604527, + 0.00657950434833765, + 0.01761462166905403, + 0.01955122873187065, + 0.014480985701084137, + 0.017090486362576485, + 0.017376022413372993, + 0.007625274825841188, + 0.00416426220908761, + 0.002562246983870864, + 0.00763612100854516, + 0.008211662992835045, + 0.003217638935893774, + 0.006278033368289471, + 0.012355568818747997, + 0.009155807085335255, + 0.005329185631126165, + 0.0043498133309185505, + 0.010090488940477371, + 0.011316826567053795, + 0.00946488045156002, + 0.008990520611405373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.07000525295734406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "grand, static close-up shot, shot from a slightly upward angle, captures the transcendent moment when the end of the dancing auroral band meets the summit of a solitary snowy mountain. Like a flowing cosmic silk, the translucent curtain of vibrant emerald green, edged with wisps of magenta and lavender, sweeps gently, almost caressingly, across the peak's highest point with a slow, graceful movement. The summit itself is a blade-like rock, blanketed in a thick, untouched layer of powdery snow, its surface delicately undulating and textured by the cold wind. As the aurora's light streams past, the snow grains are momentarily illuminated, reflecting a crystalline glow and casting fleeting, dreamy colors across the pristine white. Behind them is the inky night sky, dotted with countless clearly visible, twinkling stars like cold stars. The whole picture, dominated by cold colors, is full of otherworldly tranquility and grandeur, creating an awe-inspiring, pure and cold cosmic atmosphere.", + "seed": 18688, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_1/index_7260_seed_18688_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16670332360081375, + "vlm_logprob": -0.16670332360081375, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.16670332360081375, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08092410117387772 + ], + "opposite_logprob_tokens": [ + -2.7059240341186523 + ], + "logprob_reward": -0.08092410117387772, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08092410117387772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3147737681865692, + 0.12286446988582611, + 0.12552589178085327, + 0.1581800878047943, + 0.026243850588798523, + 0.04380880296230316, + 0.09692700207233429, + 0.018475590273737907, + 0.06678160279989243, + 0.026418928056955338 + ], + "credit_map": [ + 0.030977686867117882, + 0.03529025614261627, + 0.049729399383068085, + 0.04611430689692497, + 0.03779228404164314, + 0.04716993868350983, + 0.03198527544736862, + 0.03571461886167526, + 0.008993534371256828, + 0.024474233388900757, + 0.01112962607294321, + 0.013471625745296478, + 0.014094932936131954, + 0.0294040460139513, + 0.009432537481188774, + 0.011863934807479382, + 0.008903874084353447, + 0.04634290188550949, + 0.00407646968960762, + 0.006251397542655468, + 0.03760714456439018, + 0.011706309393048286, + 0.005283411126583815, + 0.005354390479624271, + 0.007488511968404055, + 0.06024413928389549, + 0.005970871541649103, + 0.008941620588302612, + 0.05713741481304169, + 0.008057928644120693, + 0.005515113472938538, + 0.0048244851641356945, + 0.0018274512840434909, + 0.0052784704603254795, + 0.001539899269118905, + 0.003058485686779022, + 0.00463908864185214, + 0.00417724484577775, + 0.002781408838927746, + 0.00294180354103446, + 0.002695732284337282, + 0.011267066933214664, + 0.0027358238585293293, + 0.00541971530765295, + 0.012602767907083035, + 0.0034554251469671726, + 0.0028284103609621525, + 0.002803863026201725, + 0.007660479284822941, + 0.015570217743515968, + 0.009348508901894093, + 0.01986595429480076, + 0.011135228909552097, + 0.011728142388164997, + 0.012585220858454704, + 0.009033247828483582, + 0.0016722317086532712, + 0.004258612636476755, + 0.0012395858066156507, + 0.0026520858518779278, + 0.0032195430248975754, + 0.0020443531684577465, + 0.001839813427068293, + 0.0015493641840294003, + 0.008343576453626156, + 0.01822284795343876, + 0.0061111231334507465, + 0.0064733633771538734, + 0.01163341011852026, + 0.006704067345708609, + 0.0050492072477936745, + 0.004244003910571337, + 0.0026229966897517443, + 0.0041925846599042416, + 0.0022449889220297337, + 0.003492628689855337, + 0.005577243398874998, + 0.0043406314216554165, + 0.0015378440730273724, + 0.002410010900348425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.08092410117387772 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547831982374191 + ], + "opposite_logprob_tokens": [ + -4.387547969818115 + ], + "logprob_reward": -0.012547831982374191, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547831982374191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15777505934238434, + 0.07721363753080368, + 0.2970561981201172, + 0.15561747550964355, + 0.04298495501279831, + 0.036517184227705, + 0.1473274528980255, + 0.020300626754760742, + 0.03234710544347763, + 0.032860320061445236 + ], + "credit_map": [ + 0.018122784793376923, + 0.021469414234161377, + 0.024241667240858078, + 0.0168800987303257, + 0.020643379539251328, + 0.02663222886621952, + 0.013065186329185963, + 0.01672028750181198, + 0.00621039466932416, + 0.016307314857840538, + 0.006295495200902224, + 0.006777273956686258, + 0.008642606437206268, + 0.020476562902331352, + 0.005391237325966358, + 0.0071127465926110744, + 0.01570802927017212, + 0.12026835978031158, + 0.006119572091847658, + 0.011129775084555149, + 0.11296379566192627, + 0.017141571268439293, + 0.0059526655822992325, + 0.0077724456787109375, + 0.005680188070982695, + 0.06040668860077858, + 0.005000539124011993, + 0.007451784797012806, + 0.059410080313682556, + 0.008918453939259052, + 0.004927275236696005, + 0.0038224586751312017, + 0.0016713276272639632, + 0.015741901472210884, + 0.001909351907670498, + 0.0014670671662315726, + 0.016373980790376663, + 0.002674927469342947, + 0.001599678653292358, + 0.0015467191115021706, + 0.00281084468588233, + 0.008255420252680779, + 0.0022605224512517452, + 0.004533599596470594, + 0.010767543688416481, + 0.0034955309238284826, + 0.0022283548023551702, + 0.0021653666626662016, + 0.015758318826556206, + 0.02819901518523693, + 0.014247083105146885, + 0.02363547682762146, + 0.024040693417191505, + 0.016581008210778236, + 0.013255617581307888, + 0.011610246263444424, + 0.001962424488738179, + 0.0036126484628766775, + 0.0019495566375553608, + 0.003172582946717739, + 0.0025412540417164564, + 0.002087529981508851, + 0.00291104125790298, + 0.002063589170575142, + 0.0037746562156826258, + 0.008624733425676823, + 0.003555119503289461, + 0.0033187419176101685, + 0.004616754595190287, + 0.0037555405870079994, + 0.002409995999187231, + 0.002291562035679817, + 0.0026090447790920734, + 0.0030644582584500313, + 0.0036430538166314363, + 0.00545869767665863, + 0.006228476297110319, + 0.007801069878041744, + 0.002050134353339672, + 0.0020053826738148928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.012547831982374191 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003617530455812812 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.003617530455812812, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003617530455812812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0980750322341919, + 0.14220887422561646, + 0.14710478484630585, + 0.16504645347595215, + 0.057968538254499435, + 0.1379479467868805, + 0.10429046303033829, + 0.033222779631614685, + 0.059980396181344986, + 0.054154712706804276 + ], + "credit_map": [ + 0.01211555115878582, + 0.017023636028170586, + 0.013244817964732647, + 0.012363004498183727, + 0.011201540008187294, + 0.013823175802826881, + 0.008020895533263683, + 0.01028240006417036, + 0.011849026195704937, + 0.034944456070661545, + 0.010509936138987541, + 0.01370968297123909, + 0.01642325520515442, + 0.032095372676849365, + 0.009924395009875298, + 0.012752760201692581, + 0.009999765083193779, + 0.05313403904438019, + 0.004795761778950691, + 0.009607336483895779, + 0.04265745356678963, + 0.015140729956328869, + 0.00498146703466773, + 0.006788236554712057, + 0.007518828380852938, + 0.06049250438809395, + 0.006722445134073496, + 0.011195201426744461, + 0.056856412440538406, + 0.009554494172334671, + 0.006981853861361742, + 0.005724722519516945, + 0.0038329362869262695, + 0.007579020690172911, + 0.0036329750437289476, + 0.010166660882532597, + 0.010095812380313873, + 0.00708615081384778, + 0.009782860055565834, + 0.005792123731225729, + 0.009759854525327682, + 0.04029858857393265, + 0.00691923126578331, + 0.007452521938830614, + 0.04833197966217995, + 0.011975015513598919, + 0.007089856080710888, + 0.006120890378952026, + 0.00613756850361824, + 0.012858203612267971, + 0.00838757585734129, + 0.010389560833573341, + 0.02364470437169075, + 0.028222305700182915, + 0.008229094557464123, + 0.006421452853828669, + 0.0036940027493983507, + 0.006189746782183647, + 0.0035759001038968563, + 0.003708301577717066, + 0.006600960157811642, + 0.003951200749725103, + 0.0029471188317984343, + 0.0025555496104061604, + 0.008272829465568066, + 0.017601052299141884, + 0.005405000876635313, + 0.004814505111426115, + 0.010461985133588314, + 0.006156360264867544, + 0.003934305161237717, + 0.0033343597315251827, + 0.005119225941598415, + 0.00899345800280571, + 0.00574974762275815, + 0.005984046962112188, + 0.0073816897347569466, + 0.014271385036408901, + 0.003159894607961178, + 0.0034952678252011538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.003617530455812812 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933711767196655 + ], + "opposite_logprob_tokens": [ + -0.6933711767196655 + ], + "logprob_reward": -0.6933711767196655, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933711767196655, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.25189822912216187, + 0.1230371743440628, + 0.22385813295841217, + 0.059475429356098175, + 0.02523823454976082, + 0.04131506383419037, + 0.10153061151504517, + 0.03637346997857094, + 0.09579092264175415, + 0.04148270562291145 + ], + "credit_map": [ + 0.02964967116713524, + 0.032547447830438614, + 0.03376680612564087, + 0.03565523028373718, + 0.03343525528907776, + 0.03587134927511215, + 0.025123726576566696, + 0.02584875002503395, + 0.01070637721568346, + 0.028365381062030792, + 0.010329671204090118, + 0.014170642010867596, + 0.012906176969408989, + 0.022899679839611053, + 0.011037467047572136, + 0.012621794827282429, + 0.013963387347757816, + 0.08917290717363358, + 0.004557193256914616, + 0.00814764853566885, + 0.07602250576019287, + 0.018805496394634247, + 0.005729274358600378, + 0.007459717337042093, + 0.004870138131082058, + 0.01542079821228981, + 0.0027144148480147123, + 0.005152333527803421, + 0.018937358632683754, + 0.006136766634881496, + 0.0030452522914856672, + 0.0031983747612684965, + 0.0018934690160676837, + 0.004837071523070335, + 0.0027497955597937107, + 0.0029989343602210283, + 0.0029888490680605173, + 0.0041371663101017475, + 0.00284553412348032, + 0.0027874186635017395, + 0.003616214031353593, + 0.007277330383658409, + 0.0033686368260532618, + 0.010082284919917583, + 0.004203638527542353, + 0.004395083524286747, + 0.004395289346575737, + 0.003976583946496248, + 0.007860002107918262, + 0.01895974576473236, + 0.007140600588172674, + 0.008940544910728931, + 0.022338028997182846, + 0.022138770669698715, + 0.008928827941417694, + 0.0052240993827581406, + 0.004225033335387707, + 0.007742777466773987, + 0.0030268330592662096, + 0.00426014419645071, + 0.006059575825929642, + 0.005262856371700764, + 0.0032704868353903294, + 0.0025257624220103025, + 0.011298009194433689, + 0.024887993931770325, + 0.010121292434632778, + 0.009551623836159706, + 0.017119916155934334, + 0.00998504739254713, + 0.0063988035544753075, + 0.0064282407984137535, + 0.003462801920250058, + 0.006114182993769646, + 0.0052694957703351974, + 0.010552948340773582, + 0.005180405452847481, + 0.0033466711174696684, + 0.004162745084613562, + 0.003393460065126419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.6933711767196655 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene dominated by a solitary snowy mountain with a blade-like rock summit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an auroral band visible with vibrant emerald green colors edged with wisps of magenta and lavender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an inky night sky dotted with countless clearly visible twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the aurora move slowly and gracefully across the mountain peak while illuminating the snow grains?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043055977672338486 + ], + "opposite_logprob_tokens": [ + -3.168056011199951 + ], + "logprob_reward": -0.043055977672338486, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043055977672338486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06655202805995941, + 0.07916293293237686, + 0.13216112554073334, + 0.1355428248643875, + 0.08306349068880081, + 0.22424520552158356, + 0.11153692752122879, + 0.06328780204057693, + 0.060020748525857925, + 0.04442693293094635 + ], + "credit_map": [ + 0.00829396490007639, + 0.009215044789016247, + 0.007944720797240734, + 0.00942365825176239, + 0.01092235092073679, + 0.008265282027423382, + 0.005833218805491924, + 0.006653785239905119, + 0.00697409687563777, + 0.013775945641100407, + 0.005363750737160444, + 0.009889262728393078, + 0.008907320909202099, + 0.020256750285625458, + 0.0066455695778131485, + 0.0073502385057508945, + 0.009322243742644787, + 0.03856037184596062, + 0.006030506454408169, + 0.008099865168333054, + 0.03101545199751854, + 0.023579062893986702, + 0.006193426437675953, + 0.009360198862850666, + 0.005998901557177305, + 0.05019241198897362, + 0.006573530845344067, + 0.007412023842334747, + 0.048771437257528305, + 0.00774223729968071, + 0.004556572996079922, + 0.004295699764043093, + 0.005488396622240543, + 0.01478108111768961, + 0.008460398763418198, + 0.008084396831691265, + 0.017352482303977013, + 0.015995174646377563, + 0.006115628872066736, + 0.006785929668694735, + 0.016316479071974754, + 0.06445375084877014, + 0.010760665871202946, + 0.010351445525884628, + 0.08078836649656296, + 0.020769398659467697, + 0.011248488910496235, + 0.009556623175740242, + 0.012376240454614162, + 0.023390168324112892, + 0.011068410240113735, + 0.015162345953285694, + 0.017666297033429146, + 0.013666300103068352, + 0.01081609446555376, + 0.007391070947051048, + 0.00845541711896658, + 0.011599639430642128, + 0.0059804366901516914, + 0.007467671763151884, + 0.012486235238611698, + 0.007687259465456009, + 0.005139437038451433, + 0.004471704829484224, + 0.008079838007688522, + 0.018097680062055588, + 0.004683911334723234, + 0.005295548588037491, + 0.009882992133498192, + 0.006637705489993095, + 0.00388888968154788, + 0.00345417857170105, + 0.0051552834920585155, + 0.008435498923063278, + 0.004993828479200602, + 0.007082599215209484, + 0.004582816734910011, + 0.0060165077447891235, + 0.003912513144314289, + 0.004247885663062334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.043055977672338486 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_10/index_tensor([8145, 8145])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_10/index_tensor([8145, 8145])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8ce284b3eda90972e8a2c29b2d8b1b3a485e99a1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_10/index_tensor([8145, 8145])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, dynamic aerial chase shot, captured from a slightly elevated rear perspective, captures a divine monkey clad in magnificent golden armor. He rides a swirling cloud, its edges radiating an ethereal golden glow, hurtling at breathtaking speed across the peaks of a vast, endless mountain range, steep as an ink painting. The divine monkey's golden armor is intricately carved with dragon patterns, each plate reflecting a dazzling light in the intense dawn sunlight. Two slender phoenix feathers from the purple-gold crown on his head sway violently backwards in the high-speed air. Through his helmet, his focused, resolute gaze is visible, his golden pupils gleaming with a divine radiance. The mist beneath his feet is not ethereal but a solid chariot, cleaving the air with immense momentum, leaving a clear, turbulent trail in its wake. The camera swoops and rises with him, passing over deep canyons shrouded in morning mist and steep ridges covered with dark green pines. The distant skyline is dyed a brilliant orange-red by the rising sun. The golden sunlight penetrates the clouds, casting magnificent beams of light, rendering the entire scene majestic and mythical, creating an epic atmosphere full of power and freedom.", + "seed": 3850, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_10/index_8145_seed_3850_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0869631052017212, + "vlm_logprob": -0.0869631052017212, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.0869631052017212, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.058521073311567307 + ], + "opposite_logprob_tokens": [ + -2.933521032333374 + ], + "logprob_reward": -0.058521073311567307, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.058521073311567307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04376355931162834, + 0.07620522379875183, + 0.07652689516544342, + 0.08163641393184662, + 0.45947641134262085, + 0.10613738000392914, + 0.050021518021821976, + 0.033569756895303726, + 0.02175881341099739, + 0.0509040430188179 + ], + "credit_map": [ + 0.008277175016701221, + 0.004780337680131197, + 0.004703114274889231, + 0.004177167546004057, + 0.0036073990631848574, + 0.008435402065515518, + 0.0038629802875220776, + 0.005919979885220528, + 0.007817141711711884, + 0.007841299287974834, + 0.006920631974935532, + 0.009430288337171078, + 0.01768679730594158, + 0.00489335972815752, + 0.017062420025467873, + 0.0045532770454883575, + 0.008914207108318806, + 0.006331865210086107, + 0.01151520200073719, + 0.016467273235321045, + 0.009047303348779678, + 0.009204274974763393, + 0.007701986003667116, + 0.007344773970544338, + 0.009077459573745728, + 0.01014243345707655, + 0.008627673611044884, + 0.011994942091405392, + 0.011044378392398357, + 0.008652334101498127, + 0.01510591059923172, + 0.006991281174123287, + 0.050361815840005875, + 0.0214979350566864, + 0.11615440994501114, + 0.025634821504354477, + 0.015291494317352772, + 0.1761491894721985, + 0.014417090453207493, + 0.039969634264707565, + 0.00679530156776309, + 0.01011762022972107, + 0.019707972183823586, + 0.011000584810972214, + 0.00727976206690073, + 0.03706694766879082, + 0.006490772124379873, + 0.007678413297981024, + 0.00628996267914772, + 0.008534461259841919, + 0.004818896763026714, + 0.008084733039140701, + 0.007791273295879364, + 0.004839003086090088, + 0.004527616314589977, + 0.005135569255799055, + 0.004141423385590315, + 0.005375344771891832, + 0.003812742419540882, + 0.0039655049331486225, + 0.004490233957767487, + 0.0032293694093823433, + 0.005114477593451738, + 0.0034406613558530807, + 0.0022792809177190065, + 0.002882166299968958, + 0.0025957399047911167, + 0.0031094569712877274, + 0.002805054187774658, + 0.0026900393422693014, + 0.002529460471123457, + 0.002867614384740591, + 0.004459950607270002, + 0.0064204479567706585, + 0.006014042533934116, + 0.005013902205973864, + 0.007742427755147219, + 0.008575464598834515, + 0.005601225420832634, + 0.007076580543071032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.058521073311567307 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791677817702293 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.009791677817702293, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791677817702293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02090384066104889, + 0.014608017168939114, + 0.059478022158145905, + 0.08228187263011932, + 0.7570245862007141, + 0.01956341043114662, + 0.022525256499648094, + 0.006246702279895544, + 0.005896495655179024, + 0.011471811681985855 + ], + "credit_map": [ + 0.002811454003676772, + 0.0028368872590363026, + 0.0026720829773694277, + 0.0023046170827001333, + 0.0025160254444926977, + 0.00359359965659678, + 0.0017969387117773294, + 0.0023722341284155846, + 0.001511691021732986, + 0.0015068571083247662, + 0.0017466760473325849, + 0.0018866137834265828, + 0.003034762805327773, + 0.0011902704136446118, + 0.002757843816652894, + 0.0009733001352287829, + 0.007791354786604643, + 0.004426456522196531, + 0.009297863580286503, + 0.010802293196320534, + 0.007691475562751293, + 0.006802521180361509, + 0.00541374646127224, + 0.007252308540046215, + 0.009399662725627422, + 0.008726458065211773, + 0.010407301597297192, + 0.012836754322052002, + 0.009718756191432476, + 0.008940336294472218, + 0.01512756384909153, + 0.007125033065676689, + 0.08365187048912048, + 0.036618903279304504, + 0.1947728842496872, + 0.04075136408209801, + 0.02324015647172928, + 0.29273873567581177, + 0.01969776675105095, + 0.06555284559726715, + 0.0017582930158823729, + 0.0019883655477315187, + 0.0036161141470074654, + 0.0020058266818523407, + 0.0012591338017955422, + 0.006270149722695351, + 0.0012833906803280115, + 0.0013821354368701577, + 0.002356327371671796, + 0.004633926786482334, + 0.0023490185849368572, + 0.0033207638189196587, + 0.003542244667187333, + 0.002168703591451049, + 0.0020214186515659094, + 0.002132849534973502, + 0.0007626532460562885, + 0.0009185249800793827, + 0.0007193658384494483, + 0.0008889900636859238, + 0.0007816536817699671, + 0.0005984939634799957, + 0.0008310661651194096, + 0.0007459532935172319, + 0.0007214492652565241, + 0.0007886059465818107, + 0.0007897280156612396, + 0.0006608582916669548, + 0.0008068380411714315, + 0.0010175476782023907, + 0.0005322400247678161, + 0.0005792278097942472, + 0.0017500942340120673, + 0.0023367907851934433, + 0.0015687659615650773, + 0.001369440695270896, + 0.0011611554073169827, + 0.001265582162886858, + 0.0009691007435321808, + 0.0010508811101317406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.009791677817702293 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31343576312065125 + ], + "opposite_logprob_tokens": [ + -1.3134357929229736 + ], + "logprob_reward": -0.31343576312065125, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31343576312065125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03750177472829819, + 0.04025160148739815, + 0.13624878227710724, + 0.02020983397960663, + 0.6725849509239197, + 0.030463017523288727, + 0.016939286142587662, + 0.011248929426074028, + 0.011799801141023636, + 0.0227520614862442 + ], + "credit_map": [ + 0.008038436062633991, + 0.003409998258575797, + 0.005328624974936247, + 0.0028446337673813105, + 0.0030584961641579866, + 0.00900102686136961, + 0.0017319973558187485, + 0.004088554996997118, + 0.004457236733287573, + 0.0035508759319782257, + 0.004287943709641695, + 0.0043070074170827866, + 0.009781571105122566, + 0.003363577648997307, + 0.0071475268341600895, + 0.0033558609429746866, + 0.038704145699739456, + 0.008100476115942001, + 0.01136426068842411, + 0.005983220878988504, + 0.006266588345170021, + 0.01317800022661686, + 0.006375341676175594, + 0.046276748180389404, + 0.003377928165718913, + 0.0033131439704447985, + 0.0020593584049493074, + 0.002830189187079668, + 0.0025558846537023783, + 0.0025117897894233465, + 0.0021958646830171347, + 0.0013656734954565763, + 0.07511302083730698, + 0.03198658302426338, + 0.17227452993392944, + 0.0359271839261055, + 0.02119147591292858, + 0.2602122128009796, + 0.017872091382741928, + 0.05800779536366463, + 0.00391417695209384, + 0.004561254754662514, + 0.006055478472262621, + 0.003933534491807222, + 0.003097354434430599, + 0.0028672292828559875, + 0.0030559981241822243, + 0.002977986354380846, + 0.0019676396623253822, + 0.003949017263948917, + 0.0021258248016238213, + 0.002103432547301054, + 0.0024043270386755466, + 0.0012949536321684718, + 0.0017182978335767984, + 0.0013757911510765553, + 0.001648541772738099, + 0.0020907181315124035, + 0.001311891246587038, + 0.0014987707836553454, + 0.0013281838037073612, + 0.0012058569118380547, + 0.0010893959552049637, + 0.0010755696566775441, + 0.001633320702239871, + 0.003179288934916258, + 0.0012348041636869311, + 0.001366669894196093, + 0.0008428959408774972, + 0.001944911782629788, + 0.0007175629143603146, + 0.0008803458767943084, + 0.002810967853292823, + 0.004732647445052862, + 0.0030857676174491644, + 0.0032939312513917685, + 0.0020813611336052418, + 0.002410710556432605, + 0.0017890334129333496, + 0.002547639422118664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.31343576312065125 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023278873413801193 + ], + "opposite_logprob_tokens": [ + -3.7732789516448975 + ], + "logprob_reward": -0.023278873413801193, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023278873413801193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030401524156332016, + 0.014522980898618698, + 0.44128137826919556, + 0.030889304354786873, + 0.17897123098373413, + 0.056389205157756805, + 0.08854134380817413, + 0.018342874944210052, + 0.032028015702962875, + 0.1086321622133255 + ], + "credit_map": [ + 0.0034208507277071476, + 0.003706190502271056, + 0.0068790502846241, + 0.0031888417433947325, + 0.0030349460430443287, + 0.0033677765168249607, + 0.0025476051960140467, + 0.004256263840943575, + 0.001964671304449439, + 0.0014844618272036314, + 0.0019332718802616, + 0.0026176399551331997, + 0.001965221483260393, + 0.0009368402534164488, + 0.002187576610594988, + 0.0014332965947687626, + 0.06282046437263489, + 0.03435758128762245, + 0.06702794879674911, + 0.06381139159202576, + 0.05465267226099968, + 0.0498192235827446, + 0.035850126296281815, + 0.07294198125600815, + 0.003399269189685583, + 0.004977288190275431, + 0.0025743376463651657, + 0.004239952657371759, + 0.004136753734201193, + 0.00371758290566504, + 0.005876569077372551, + 0.0019675514195114374, + 0.021619398146867752, + 0.009528004564344883, + 0.0453379787504673, + 0.009959998540580273, + 0.0059829759411513805, + 0.06627444922924042, + 0.005579939112067223, + 0.01468847319483757, + 0.005851040128618479, + 0.0060461401008069515, + 0.012549260631203651, + 0.0050038970075547695, + 0.004193573724478483, + 0.014185454696416855, + 0.0040085953660309315, + 0.004551243968307972, + 0.013898076489567757, + 0.020172830671072006, + 0.01039647776633501, + 0.014302663505077362, + 0.006160898599773645, + 0.008169416338205338, + 0.006195066496729851, + 0.009245919063687325, + 0.0028095862362533808, + 0.0033875014632940292, + 0.002091092523187399, + 0.003279798896983266, + 0.0015650085406377912, + 0.0020716518629342318, + 0.0014366833493113518, + 0.0017015521880239248, + 0.007116249762475491, + 0.005771589931100607, + 0.0046468460932374, + 0.0033360389061272144, + 0.0027039353735744953, + 0.0038962543476372957, + 0.002200255636125803, + 0.0023568454198539257, + 0.016991175711154938, + 0.026772411540150642, + 0.015940308570861816, + 0.013732117600739002, + 0.007851897738873959, + 0.009901280514895916, + 0.007771698292344809, + 0.009671280160546303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 9 + ], + "selected_reward": -0.023278873413801193 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978813834488392 + ], + "opposite_logprob_tokens": [ + -3.5297882556915283 + ], + "logprob_reward": -0.02978813834488392, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978813834488392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04128453880548477, + 0.029211223125457764, + 0.08029305189847946, + 0.026989202946424484, + 0.5620644092559814, + 0.08282072097063065, + 0.09624861180782318, + 0.046344954520463943, + 0.015116783790290356, + 0.019626479595899582 + ], + "credit_map": [ + 0.004932804498821497, + 0.005235461518168449, + 0.007379306014627218, + 0.00470846239477396, + 0.004412621725350618, + 0.0050228917971253395, + 0.0038724588230252266, + 0.0057205334305763245, + 0.0031938431784510612, + 0.0031226668506860733, + 0.004845087882131338, + 0.00425740797072649, + 0.002976632211357355, + 0.003940424881875515, + 0.0030321022495627403, + 0.0038430585991591215, + 0.00854745414108038, + 0.006361417006701231, + 0.015654955059289932, + 0.013936611823737621, + 0.009521790780127048, + 0.01150091364979744, + 0.00634218193590641, + 0.008427727967500687, + 0.004835156258195639, + 0.0031639139633625746, + 0.0042184386402368546, + 0.003872388042509556, + 0.00286219734698534, + 0.003139759646728635, + 0.002250197110697627, + 0.0026471514720469713, + 0.06350333243608475, + 0.028429875150322914, + 0.14504578709602356, + 0.030404726043343544, + 0.01788667030632496, + 0.21290583908557892, + 0.0161038339138031, + 0.04778435826301575, + 0.007971101440489292, + 0.010399086400866508, + 0.015096832998096943, + 0.007078977767378092, + 0.00463886559009552, + 0.026308367028832436, + 0.005514331627637148, + 0.005813162308186293, + 0.011927423998713493, + 0.018138229846954346, + 0.009534550830721855, + 0.01686098240315914, + 0.011853012256324291, + 0.009414289146661758, + 0.009651615284383297, + 0.008868509903550148, + 0.006999674718827009, + 0.00826235767453909, + 0.005142827518284321, + 0.008446671068668365, + 0.00430445559322834, + 0.00463648559525609, + 0.004076842218637466, + 0.004475641064345837, + 0.0029657434206455946, + 0.0017856790218502283, + 0.002039580373093486, + 0.001651315251365304, + 0.0015637283213436604, + 0.0020088336896151304, + 0.0013364988844841719, + 0.001765404362231493, + 0.0030354654882103205, + 0.004352012183517218, + 0.0023032568860799074, + 0.0022571245208382607, + 0.0017130388878285885, + 0.003390043508261442, + 0.0011234377743676305, + 0.001452102093026042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.02978813834488392 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, dynamic aerial chase shot, captured from a slightly elevated rear perspective, captures a divine monkey clad in magnificent golden armor. He rides a swirling cloud, its edges radiating an ethereal golden glow, hurtling at breathtaking speed across the peaks of a vast, endless mountain range, steep as an ink painting. The divine monkey's golden armor is intricately carved with dragon patterns, each plate reflecting a dazzling light in the intense dawn sunlight. Two slender phoenix feathers from the purple-gold crown on his head sway violently backwards in the high-speed air. Through his helmet, his focused, resolute gaze is visible, his golden pupils gleaming with a divine radiance. The mist beneath his feet is not ethereal but a solid chariot, cleaving the air with immense momentum, leaving a clear, turbulent trail in its wake. The camera swoops and rises with him, passing over deep canyons shrouded in morning mist and steep ridges covered with dark green pines. The distant skyline is dyed a brilliant orange-red by the rising sun. The golden sunlight penetrates the clouds, casting magnificent beams of light, rendering the entire scene majestic and mythical, creating an epic atmosphere full of power and freedom.", + "seed": 3851, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_10/index_8145_seed_3851_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05702228536829353, + "vlm_logprob": -0.05702228536829353, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.05702228536829353, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05903678387403488 + ], + "opposite_logprob_tokens": [ + -2.9340367317199707 + ], + "logprob_reward": -0.05903678387403488, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05903678387403488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1532839983701706, + 0.25904250144958496, + 0.08643504232168198, + 0.10165496915578842, + 0.02835707738995552, + 0.04360409826040268, + 0.046623799949884415, + 0.08298425376415253, + 0.05192459002137184, + 0.1460895836353302 + ], + "credit_map": [ + 0.015015540644526482, + 0.021575702354311943, + 0.018892308697104454, + 0.016598938032984734, + 0.01615257002413273, + 0.022039515897631645, + 0.01958364062011242, + 0.023425795137882233, + 0.03191574290394783, + 0.03945649042725563, + 0.03426448628306389, + 0.045603446662425995, + 0.025318780913949013, + 0.030807102099061012, + 0.02232896164059639, + 0.029347509145736694, + 0.009286867454648018, + 0.00872605200856924, + 0.01031942293047905, + 0.007722905371338129, + 0.01774434559047222, + 0.013575978577136993, + 0.006227476056665182, + 0.012832005508244038, + 0.013575070537626743, + 0.013246257789433002, + 0.008445723913609982, + 0.008503684774041176, + 0.017320705577731133, + 0.013035713694989681, + 0.012389380484819412, + 0.015138440765440464, + 0.003557227086275816, + 0.005196685437113047, + 0.0027303092647343874, + 0.0036260359920561314, + 0.003402613801881671, + 0.00333631900139153, + 0.003325308905914426, + 0.0031825811602175236, + 0.004172021523118019, + 0.00476226769387722, + 0.004080924671143293, + 0.003693685866892338, + 0.0035535653587430716, + 0.012696683406829834, + 0.004080872051417828, + 0.0065640839748084545, + 0.0046560270711779594, + 0.006828145124018192, + 0.006777223199605942, + 0.004630096256732941, + 0.006131519563496113, + 0.006821564864367247, + 0.005856091156601906, + 0.004923134110867977, + 0.011075587011873722, + 0.010657328180968761, + 0.012848910875618458, + 0.013040370307862759, + 0.006558505818247795, + 0.011214283294975758, + 0.008554748259484768, + 0.009034532122313976, + 0.006239354610443115, + 0.004583856090903282, + 0.006044657435268164, + 0.008061526343226433, + 0.009434957057237625, + 0.0068555306643247604, + 0.006480155512690544, + 0.0042245578952133656, + 0.01990942284464836, + 0.022432008758187294, + 0.018360685557127, + 0.020834224298596382, + 0.030771983787417412, + 0.013537393882870674, + 0.009837946854531765, + 0.010405938141047955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.05903678387403488 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052597238682210445 + ], + "opposite_logprob_tokens": [ + -5.2552595138549805 + ], + "logprob_reward": -0.0052597238682210445, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052597238682210445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09524228423833847, + 0.3024597465991974, + 0.17513811588287354, + 0.07746182382106781, + 0.04273094981908798, + 0.049936775118112564, + 0.04890596494078636, + 0.09555982798337936, + 0.0556405670940876, + 0.05692389979958534 + ], + "credit_map": [ + 0.010426219552755356, + 0.01476185955107212, + 0.011870657093822956, + 0.0120683079585433, + 0.00933797750622034, + 0.012982981279492378, + 0.011232838034629822, + 0.012561446987092495, + 0.040013860911130905, + 0.057491645216941833, + 0.03378948196768761, + 0.048717960715293884, + 0.031322021037340164, + 0.03337351232767105, + 0.027402102947235107, + 0.030349180102348328, + 0.021914616227149963, + 0.023665884509682655, + 0.0264851413667202, + 0.02063610404729843, + 0.02254905179142952, + 0.01894231326878071, + 0.017431898042559624, + 0.02351314015686512, + 0.009319910779595375, + 0.007624810561537743, + 0.006751670967787504, + 0.007339527364820242, + 0.013683213852345943, + 0.010913810692727566, + 0.009051084518432617, + 0.012777810916304588, + 0.0039030006155371666, + 0.0059554968029260635, + 0.004380802158266306, + 0.0053346529603004456, + 0.004392345901578665, + 0.006357167847454548, + 0.0074068717658519745, + 0.005000615026801825, + 0.006142703350633383, + 0.005943693220615387, + 0.006087743677198887, + 0.004664828069508076, + 0.003408916527405381, + 0.01159692369401455, + 0.004789910279214382, + 0.007302061654627323, + 0.00572114996612072, + 0.00512218801304698, + 0.009720159694552422, + 0.005592973902821541, + 0.004268614575266838, + 0.006504985969513655, + 0.005041190888732672, + 0.006934706587344408, + 0.013203756883740425, + 0.011926724575459957, + 0.015731768682599068, + 0.012011565268039703, + 0.008309856988489628, + 0.01533981878310442, + 0.009448754601180553, + 0.00958759244531393, + 0.007991465739905834, + 0.007389903534203768, + 0.008879058994352818, + 0.007038049399852753, + 0.004372102674096823, + 0.004693158436566591, + 0.003919865004718304, + 0.011356968432664871, + 0.007820059545338154, + 0.008845234289765358, + 0.006128051783889532, + 0.007591592147946358, + 0.008463578298687935, + 0.005187734495848417, + 0.006521784700453281, + 0.00636587617918849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.0052597238682210445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015470415353775 + ], + "opposite_logprob_tokens": [ + -1.7015470266342163 + ], + "logprob_reward": -0.2015470415353775, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015470415353775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2208331972360611, + 0.16839607059955597, + 0.16345207393169403, + 0.10172589123249054, + 0.03737107291817665, + 0.05784691870212555, + 0.04902990534901619, + 0.06763362884521484, + 0.03960734233260155, + 0.09410389512777328 + ], + "credit_map": [ + 0.03481365740299225, + 0.023372961208224297, + 0.03009483404457569, + 0.016711635515093803, + 0.016733311116695404, + 0.01688399165868759, + 0.023911774158477783, + 0.05831104889512062, + 0.022021904587745667, + 0.021991370245814323, + 0.01928565837442875, + 0.029654743149876595, + 0.015278441831469536, + 0.02640683390200138, + 0.016678934916853905, + 0.01707819290459156, + 0.014021238312125206, + 0.011872719042003155, + 0.014585578814148903, + 0.01276230439543724, + 0.04426758736371994, + 0.030448798090219498, + 0.013119298033416271, + 0.02237456478178501, + 0.012518975883722305, + 0.009483963251113892, + 0.010506208054721355, + 0.010220512747764587, + 0.015681909397244453, + 0.01412107702344656, + 0.012721282429993153, + 0.016471978276968002, + 0.005316184367984533, + 0.004867435898631811, + 0.003145017195492983, + 0.0044781360775232315, + 0.0037366440519690514, + 0.006363994441926479, + 0.005206654313951731, + 0.004257009830325842, + 0.003982925321906805, + 0.0064744180999696255, + 0.004531588871032, + 0.0037182022351771593, + 0.004935570992529392, + 0.01881023682653904, + 0.004371934570372105, + 0.011022043414413929, + 0.006773847620934248, + 0.00721090380102396, + 0.005939314607530832, + 0.004922505933791399, + 0.004774992819875479, + 0.009192571975290775, + 0.005095170345157385, + 0.005120598245412111, + 0.011455883271992207, + 0.005649153143167496, + 0.014577384106814861, + 0.008082891814410686, + 0.00813228264451027, + 0.010320687666535378, + 0.004659352824091911, + 0.004755999892950058, + 0.0042966087348759174, + 0.0041315993294119835, + 0.009522596374154091, + 0.004262842703610659, + 0.0032548983581364155, + 0.004696677438914776, + 0.0033934577368199825, + 0.006048665847629309, + 0.014759255573153496, + 0.015217356383800507, + 0.010723862797021866, + 0.01134081557393074, + 0.01669403351843357, + 0.010280243121087551, + 0.008046567440032959, + 0.007041762582957745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.2015470415353775 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534293346107006 + ], + "opposite_logprob_tokens": [ + -4.387534141540527 + ], + "logprob_reward": -0.012534293346107006, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534293346107006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11870726943016052, + 0.17001944780349731, + 0.0597066804766655, + 0.10501294583082199, + 0.06700626760721207, + 0.24100658297538757, + 0.07691596448421478, + 0.027000384405255318, + 0.039084091782569885, + 0.09554038941860199 + ], + "credit_map": [ + 0.015294116921722889, + 0.010364139452576637, + 0.01419643685221672, + 0.012301348149776459, + 0.01748833991587162, + 0.017761871218681335, + 0.012398135848343372, + 0.01890288107097149, + 0.016502583399415016, + 0.015389803797006607, + 0.034387171268463135, + 0.015801167115569115, + 0.022064195945858955, + 0.027262769639492035, + 0.01280858926475048, + 0.02580316923558712, + 0.005466942675411701, + 0.004207179881632328, + 0.0060104550793766975, + 0.004198121372610331, + 0.01598701812326908, + 0.009921119548380375, + 0.005391756538301706, + 0.008524090051651001, + 0.008838877081871033, + 0.007788177113980055, + 0.006732708774507046, + 0.00955758523195982, + 0.01858402229845524, + 0.015667365863919258, + 0.012758967466652393, + 0.025085236877202988, + 0.006723530124872923, + 0.011880268342792988, + 0.008066628128290176, + 0.008087125606834888, + 0.0072158752009272575, + 0.0083130719140172, + 0.010114029049873352, + 0.006605734582990408, + 0.018787739798426628, + 0.020383430644869804, + 0.01752273365855217, + 0.01773579604923725, + 0.02259995974600315, + 0.08007506281137466, + 0.02346343919634819, + 0.04043840616941452, + 0.009392378851771355, + 0.010647103190422058, + 0.013561563566327095, + 0.008187207393348217, + 0.006644986569881439, + 0.011738982982933521, + 0.008090540766716003, + 0.00865320023149252, + 0.00379987177439034, + 0.003573935478925705, + 0.004065860528498888, + 0.003280997509136796, + 0.0024187720846384764, + 0.004791822750121355, + 0.002317017875611782, + 0.002752105938270688, + 0.006033198907971382, + 0.0038710525259375572, + 0.006642651278525591, + 0.005275241564959288, + 0.003909250255674124, + 0.003283206606283784, + 0.003903636010363698, + 0.006165854632854462, + 0.01728677935898304, + 0.014668969437479973, + 0.011832713149487972, + 0.011647908017039299, + 0.013813190162181854, + 0.011217971332371235, + 0.008104514330625534, + 0.006968345958739519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.012534293346107006 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733584217727184 + ], + "opposite_logprob_tokens": [ + -5.006733417510986 + ], + "logprob_reward": -0.006733584217727184, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733584217727184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16907095909118652, + 0.1128288209438324, + 0.056742213666439056, + 0.07757660001516342, + 0.05780292674899101, + 0.10838329792022705, + 0.06767921149730682, + 0.07956498861312866, + 0.11474458873271942, + 0.15560637414455414 + ], + "credit_map": [ + 0.018356598913669586, + 0.024959255009889603, + 0.01611708290874958, + 0.0181820597499609, + 0.02451332099735737, + 0.021700572222471237, + 0.0247800350189209, + 0.020462024956941605, + 0.014949643053114414, + 0.009670259431004524, + 0.01867528073489666, + 0.013069577515125275, + 0.010033675469458103, + 0.016391567885875702, + 0.010239695198833942, + 0.019799115136265755, + 0.00624488852918148, + 0.009049344807863235, + 0.0077756731770932674, + 0.008256923407316208, + 0.008947337046265602, + 0.006693427450954914, + 0.0037776371464133263, + 0.005996981635689735, + 0.009157479740679264, + 0.008513659238815308, + 0.00735454773530364, + 0.008288506418466568, + 0.010097222402691841, + 0.00916697271168232, + 0.011355608701705933, + 0.013642600737512112, + 0.008058564737439156, + 0.006602188106626272, + 0.0046706171706318855, + 0.005374233704060316, + 0.004724191967397928, + 0.013544154353439808, + 0.007236410863697529, + 0.00759256212040782, + 0.00926334597170353, + 0.011552350595593452, + 0.012063540518283844, + 0.0105854207649827, + 0.011499177664518356, + 0.02688610926270485, + 0.00949734728783369, + 0.017036009579896927, + 0.008117694407701492, + 0.006747082807123661, + 0.012474547140300274, + 0.009204723872244358, + 0.008274178951978683, + 0.008211557753384113, + 0.006570208352059126, + 0.008079221472144127, + 0.006534746382385492, + 0.007342285476624966, + 0.018467146903276443, + 0.015512106008827686, + 0.00786484032869339, + 0.006823615171015263, + 0.007039072923362255, + 0.009981177747249603, + 0.01701752282679081, + 0.017934300005435944, + 0.01913962885737419, + 0.011321463622152805, + 0.010571078397333622, + 0.010730327107012272, + 0.009005571715533733, + 0.019024694338440895, + 0.018488770350813866, + 0.023686286062002182, + 0.02414356917142868, + 0.021272864192724228, + 0.025509033352136612, + 0.014750365167856216, + 0.011864766478538513, + 0.015890726819634438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.006733584217727184 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, dynamic aerial chase shot, captured from a slightly elevated rear perspective, captures a divine monkey clad in magnificent golden armor. He rides a swirling cloud, its edges radiating an ethereal golden glow, hurtling at breathtaking speed across the peaks of a vast, endless mountain range, steep as an ink painting. The divine monkey's golden armor is intricately carved with dragon patterns, each plate reflecting a dazzling light in the intense dawn sunlight. Two slender phoenix feathers from the purple-gold crown on his head sway violently backwards in the high-speed air. Through his helmet, his focused, resolute gaze is visible, his golden pupils gleaming with a divine radiance. The mist beneath his feet is not ethereal but a solid chariot, cleaving the air with immense momentum, leaving a clear, turbulent trail in its wake. The camera swoops and rises with him, passing over deep canyons shrouded in morning mist and steep ridges covered with dark green pines. The distant skyline is dyed a brilliant orange-red by the rising sun. The golden sunlight penetrates the clouds, casting magnificent beams of light, rendering the entire scene majestic and mythical, creating an epic atmosphere full of power and freedom.", + "seed": 3852, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_10/index_8145_seed_3852_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04804784087464213, + "vlm_logprob": -0.04804784087464213, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.04804784087464213, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03229799121618271 + ], + "opposite_logprob_tokens": [ + -3.6572980880737305 + ], + "logprob_reward": -0.03229799121618271, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03229799121618271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10475359857082367, + 0.055347222834825516, + 0.03303530067205429, + 0.040310315787792206, + 0.11767180263996124, + 0.05295521020889282, + 0.18305477499961853, + 0.04625111073255539, + 0.0927940234541893, + 0.2738266587257385 + ], + "credit_map": [ + 0.011849692091345787, + 0.010112684220075607, + 0.012878729030489922, + 0.015062292106449604, + 0.011557028628885746, + 0.016063615679740906, + 0.012541861273348331, + 0.014687690883874893, + 0.008825397118926048, + 0.006731628440320492, + 0.007119747344404459, + 0.007086733356118202, + 0.005641941912472248, + 0.006076361518353224, + 0.007608217652887106, + 0.006257194560021162, + 0.0026508739683777094, + 0.0032775953877717257, + 0.0030994717963039875, + 0.0054892427287995815, + 0.0032901130616664886, + 0.004076096229255199, + 0.006375525146722794, + 0.004776378162205219, + 0.0038655574899166822, + 0.004391359630972147, + 0.004644862841814756, + 0.007368880789726973, + 0.0039099445566535, + 0.0052601033821702, + 0.0065078274346888065, + 0.004361780360341072, + 0.014298299327492714, + 0.014694941230118275, + 0.011990556493401527, + 0.02230040729045868, + 0.018203284591436386, + 0.01097890455275774, + 0.01690608076751232, + 0.008299329318106174, + 0.0066681792959570885, + 0.004851858131587505, + 0.011922040022909641, + 0.006743643898516893, + 0.006335108075290918, + 0.006505915895104408, + 0.004851899575442076, + 0.005076561123132706, + 0.027349364012479782, + 0.023715415969491005, + 0.01602480188012123, + 0.016442380845546722, + 0.04061482101678848, + 0.01607099175453186, + 0.022787563502788544, + 0.02004944160580635, + 0.006369903218001127, + 0.008910370990633965, + 0.004581636283546686, + 0.004643919412046671, + 0.005368024110794067, + 0.0031145343091338873, + 0.006435706280171871, + 0.006827017292380333, + 0.01691720262169838, + 0.00965079478919506, + 0.010574797168374062, + 0.011622411198914051, + 0.007509892340749502, + 0.008189612068235874, + 0.013943195343017578, + 0.014386125840246677, + 0.04418395459651947, + 0.026998063549399376, + 0.03033233992755413, + 0.0362655445933342, + 0.036949899047613144, + 0.04036707431077957, + 0.03092232160270214, + 0.02780747599899769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.03229799121618271 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258894059807062 + ], + "opposite_logprob_tokens": [ + -5.255259037017822 + ], + "logprob_reward": -0.005258894059807062, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258894059807062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09468567371368408, + 0.07687657326459885, + 0.03598083183169365, + 0.04531773552298546, + 0.047901563346385956, + 0.12499918043613434, + 0.10488574951887131, + 0.076546311378479, + 0.09389466047286987, + 0.2989117205142975 + ], + "credit_map": [ + 0.012442600913345814, + 0.01271202601492405, + 0.011612417176365852, + 0.01299923937767744, + 0.006870085373520851, + 0.015027977526187897, + 0.011648206040263176, + 0.011373121291399002, + 0.009723887778818607, + 0.009743747301399708, + 0.01010203454643488, + 0.010190092027187347, + 0.008747915737330914, + 0.009203667752444744, + 0.010132892057299614, + 0.009032336995005608, + 0.0034716909285634756, + 0.0023115475196391344, + 0.004370059818029404, + 0.003368962090462446, + 0.0028722791466861963, + 0.004883668385446072, + 0.007420813199132681, + 0.007281809579581022, + 0.005798929836601019, + 0.004871624521911144, + 0.010135713033378124, + 0.006724976003170013, + 0.00450873700901866, + 0.004443477373570204, + 0.004800683353096247, + 0.004033596720546484, + 0.005136073101311922, + 0.006038876250386238, + 0.010000145994126797, + 0.00736861489713192, + 0.004959353245794773, + 0.004102853126823902, + 0.0065438514575362206, + 0.0037517971359193325, + 0.014997545629739761, + 0.0149975111708045, + 0.021718330681324005, + 0.017260393127799034, + 0.014048339799046516, + 0.008564259856939316, + 0.02215161733329296, + 0.011261185631155968, + 0.012257101945579052, + 0.014801524579524994, + 0.015282589942216873, + 0.018037928268313408, + 0.014388504438102245, + 0.01052022259682417, + 0.011148102581501007, + 0.008449777029454708, + 0.011148104444146156, + 0.011385190300643444, + 0.008561719208955765, + 0.007048237603157759, + 0.010425751097500324, + 0.007058967836201191, + 0.00911850854754448, + 0.011799834668636322, + 0.01421027909964323, + 0.010691584087908268, + 0.012769737280905247, + 0.01189626194536686, + 0.00939459353685379, + 0.011456445790827274, + 0.012531881220638752, + 0.010943882167339325, + 0.04965776950120926, + 0.02630578726530075, + 0.030848808586597443, + 0.04368726536631584, + 0.03941472992300987, + 0.04238263517618179, + 0.034345246851444244, + 0.03226948529481888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.005258894059807062 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986346781253815 + ], + "opposite_logprob_tokens": [ + -1.804863452911377 + ], + "logprob_reward": -0.17986346781253815, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986346781253815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05516654998064041, + 0.046535417437553406, + 0.02568991854786873, + 0.03553299233317375, + 0.08419427275657654, + 0.11814367026090622, + 0.09635679423809052, + 0.11867911368608475, + 0.2796856462955475, + 0.1400156021118164 + ], + "credit_map": [ + 0.006092889700084925, + 0.010069706477224827, + 0.007190893869847059, + 0.005365799646824598, + 0.004951232112944126, + 0.008527814410626888, + 0.0074120392091572285, + 0.00555617269128561, + 0.008447619155049324, + 0.003949073608964682, + 0.006514914333820343, + 0.005657833069562912, + 0.006231728941202164, + 0.005403239279985428, + 0.005651324056088924, + 0.004679687786847353, + 0.0025445243809372187, + 0.002438388764858246, + 0.0027659006882458925, + 0.002789189340546727, + 0.0025591047015041113, + 0.003706221701577306, + 0.004036117345094681, + 0.004850469995290041, + 0.0033279398921877146, + 0.004474620800465345, + 0.005058001726865768, + 0.005371407605707645, + 0.0035016117617487907, + 0.0044939350336790085, + 0.004369452130049467, + 0.004936025943607092, + 0.009383311495184898, + 0.010974488221108913, + 0.009374288842082024, + 0.017252974212169647, + 0.008897622115910053, + 0.004325667396187782, + 0.015934202820062637, + 0.008051717653870583, + 0.015081750229001045, + 0.013623500242829323, + 0.026818107813596725, + 0.012525047175586224, + 0.016364825889468193, + 0.014427175745368004, + 0.012490613386034966, + 0.0068126549012959, + 0.011576124466955662, + 0.017314018681645393, + 0.01388693880289793, + 0.013250110670924187, + 0.010111788287758827, + 0.010905465111136436, + 0.010236390866339207, + 0.009075960144400597, + 0.017900899052619934, + 0.016990235075354576, + 0.01564469374716282, + 0.011524088680744171, + 0.01557862013578415, + 0.009013754315674305, + 0.015095651149749756, + 0.01693117804825306, + 0.041815049946308136, + 0.03222711384296417, + 0.03391736373305321, + 0.037018053233623505, + 0.030509818345308304, + 0.033545609563589096, + 0.03812367096543312, + 0.03252897784113884, + 0.024144120514392853, + 0.011095439083874226, + 0.01723603717982769, + 0.01799163594841957, + 0.013739903457462788, + 0.02061556652188301, + 0.02035536617040634, + 0.014837529510259628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.17986346781253815 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008632130920886993 + ], + "opposite_logprob_tokens": [ + -4.758632183074951 + ], + "logprob_reward": -0.008632130920886993, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008632130920886993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19404831528663635, + 0.10798453539609909, + 0.1159629076719284, + 0.036906544119119644, + 0.0761822760105133, + 0.05484282597899437, + 0.06228106841444969, + 0.0631113350391388, + 0.17309027910232544, + 0.1155899167060852 + ], + "credit_map": [ + 0.017379123717546463, + 0.010184844955801964, + 0.01794903352856636, + 0.02515212632715702, + 0.017937423661351204, + 0.03922072425484657, + 0.029983047395944595, + 0.03624200448393822, + 0.014317425899207592, + 0.0099377715960145, + 0.011540132574737072, + 0.013628407381474972, + 0.008087517693638802, + 0.015956204384565353, + 0.01493826787918806, + 0.019578807055950165, + 0.009743628092110157, + 0.0065132430754601955, + 0.014233897440135479, + 0.009514269419014454, + 0.0089547298848629, + 0.020067617297172546, + 0.023960070684552193, + 0.02297544851899147, + 0.004601441323757172, + 0.003391716629266739, + 0.0036381236277520657, + 0.005239563994109631, + 0.0035814440343528986, + 0.007087546866387129, + 0.004848578479140997, + 0.004518128465861082, + 0.007484009023755789, + 0.007541715633124113, + 0.01813489757478237, + 0.009928404353559017, + 0.010360931977629662, + 0.006420127581804991, + 0.009949189610779285, + 0.006363004446029663, + 0.0075401850044727325, + 0.007282450329512358, + 0.010372070595622063, + 0.007205805741250515, + 0.008002541027963161, + 0.006292065139859915, + 0.004570527467876673, + 0.0035771820694208145, + 0.00912031065672636, + 0.0077375550754368305, + 0.009806346148252487, + 0.006053735502064228, + 0.006176924332976341, + 0.005904613994061947, + 0.008104091510176659, + 0.00937748420983553, + 0.013891875743865967, + 0.007222064770758152, + 0.011369329877197742, + 0.006678631063550711, + 0.006401100195944309, + 0.004718453157693148, + 0.006935977842658758, + 0.0058938972651958466, + 0.024955350905656815, + 0.02248435840010643, + 0.02887652814388275, + 0.022773146629333496, + 0.017053933814167976, + 0.01842775195837021, + 0.019106337800621986, + 0.019412877038121223, + 0.015723159536719322, + 0.010253232903778553, + 0.014843134209513664, + 0.016565872356295586, + 0.01613067276775837, + 0.016013534739613533, + 0.013297215104103088, + 0.012763096019625664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.008632130920886993 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186720363795757 + ], + "opposite_logprob_tokens": [ + -4.264186859130859 + ], + "logprob_reward": -0.014186720363795757, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186720363795757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13111966848373413, + 0.12267165631055832, + 0.04453931376338005, + 0.05959020555019379, + 0.06138542667031288, + 0.061343204230070114, + 0.08447854220867157, + 0.15450085699558258, + 0.1861247420310974, + 0.09424640983343124 + ], + "credit_map": [ + 0.021517667919397354, + 0.013213081285357475, + 0.016888337209820747, + 0.010075719095766544, + 0.010023269802331924, + 0.01980227418243885, + 0.018972568213939667, + 0.02062675729393959, + 0.015792423859238625, + 0.013151437975466251, + 0.01608545519411564, + 0.0138845881447196, + 0.018030107021331787, + 0.01657484471797943, + 0.015533903613686562, + 0.013618884608149529, + 0.004561517387628555, + 0.005131596699357033, + 0.006533553823828697, + 0.005637108348309994, + 0.005288559943437576, + 0.007293088361620903, + 0.004766369704157114, + 0.005327519029378891, + 0.007124070543795824, + 0.006641071755439043, + 0.009339123032987118, + 0.006766481325030327, + 0.006090566515922546, + 0.008967982605099678, + 0.007625756785273552, + 0.007035154849290848, + 0.00532452343031764, + 0.006939462851732969, + 0.009485672228038311, + 0.009015578776597977, + 0.005378100089728832, + 0.0052099996246397495, + 0.013999340124428272, + 0.006032748147845268, + 0.007847518660128117, + 0.005487143062055111, + 0.011374433524906635, + 0.007914463989436626, + 0.006622995715588331, + 0.006120767444372177, + 0.009586396627128124, + 0.00638948567211628, + 0.011258120648562908, + 0.012490909546613693, + 0.010098930448293686, + 0.010638348758220673, + 0.008430469781160355, + 0.012138444930315018, + 0.010979704558849335, + 0.008443616330623627, + 0.031101541593670845, + 0.020198499783873558, + 0.03189864754676819, + 0.0184174757450819, + 0.01644309051334858, + 0.01018538884818554, + 0.014831846579909325, + 0.011424369178712368, + 0.022360285744071007, + 0.015468710102140903, + 0.02229646034538746, + 0.023204654455184937, + 0.01617884822189808, + 0.021569665521383286, + 0.03163176774978638, + 0.033414360135793686, + 0.01939900778234005, + 0.011775972321629524, + 0.008535671979188919, + 0.009512759745121002, + 0.009762479923665524, + 0.011222923174500465, + 0.009991765953600407, + 0.014045828022062778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.014186720363795757 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, dynamic aerial chase shot, captured from a slightly elevated rear perspective, captures a divine monkey clad in magnificent golden armor. He rides a swirling cloud, its edges radiating an ethereal golden glow, hurtling at breathtaking speed across the peaks of a vast, endless mountain range, steep as an ink painting. The divine monkey's golden armor is intricately carved with dragon patterns, each plate reflecting a dazzling light in the intense dawn sunlight. Two slender phoenix feathers from the purple-gold crown on his head sway violently backwards in the high-speed air. Through his helmet, his focused, resolute gaze is visible, his golden pupils gleaming with a divine radiance. The mist beneath his feet is not ethereal but a solid chariot, cleaving the air with immense momentum, leaving a clear, turbulent trail in its wake. The camera swoops and rises with him, passing over deep canyons shrouded in morning mist and steep ridges covered with dark green pines. The distant skyline is dyed a brilliant orange-red by the rising sun. The golden sunlight penetrates the clouds, casting magnificent beams of light, rendering the entire scene majestic and mythical, creating an epic atmosphere full of power and freedom.", + "seed": 3853, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_10/index_8145_seed_3853_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.041962855588644744, + "vlm_logprob": -0.041962855588644744, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.041962855588644744, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02208799496293068 + ], + "opposite_logprob_tokens": [ + -3.897088050842285 + ], + "logprob_reward": -0.02208799496293068, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02208799496293068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14240117371082306, + 0.10835593193769455, + 0.15853509306907654, + 0.09199633449316025, + 0.05628160759806633, + 0.052148595452308655, + 0.13925133645534515, + 0.03840300068259239, + 0.07113178074359894, + 0.14149516820907593 + ], + "credit_map": [ + 0.01819682866334915, + 0.016962049528956413, + 0.02258034236729145, + 0.013292036950588226, + 0.017926784232258797, + 0.017359748482704163, + 0.015777764841914177, + 0.020305611193180084, + 0.014526910148561, + 0.014199336059391499, + 0.008153998292982578, + 0.00923248939216137, + 0.01781044341623783, + 0.017176570370793343, + 0.017134398221969604, + 0.010121789760887623, + 0.02135992981493473, + 0.015205114148557186, + 0.022792240604758263, + 0.015038112178444862, + 0.03245437517762184, + 0.010324613191187382, + 0.02344297245144844, + 0.017917731776833534, + 0.008621257729828358, + 0.005342741496860981, + 0.02081499993801117, + 0.0071456036530435085, + 0.005518192425370216, + 0.01170436106622219, + 0.007549760397523642, + 0.025299418717622757, + 0.004768560640513897, + 0.004093287978321314, + 0.010220760479569435, + 0.004147789906710386, + 0.005824950523674488, + 0.008944310247898102, + 0.005599529016762972, + 0.012682418338954449, + 0.0058530704118311405, + 0.003970910329371691, + 0.009281652048230171, + 0.004771667066961527, + 0.005904597230255604, + 0.011261094361543655, + 0.004780150949954987, + 0.0063254572451114655, + 0.01287926733493805, + 0.00932260137051344, + 0.03095223382115364, + 0.010789262130856514, + 0.009604811668395996, + 0.014794974587857723, + 0.01010465994477272, + 0.04080352187156677, + 0.004138660617172718, + 0.0027099556755274534, + 0.005153580568730831, + 0.004601237829774618, + 0.002929775044322014, + 0.0069281128235161304, + 0.0046059261076152325, + 0.007335755508393049, + 0.005866135936230421, + 0.007636052556335926, + 0.010638713836669922, + 0.007874304428696632, + 0.011193123646080494, + 0.012737046927213669, + 0.007941873744130135, + 0.007244525011628866, + 0.016691943630576134, + 0.016280759125947952, + 0.018717799335718155, + 0.017209632322192192, + 0.022166145965456963, + 0.01351406704634428, + 0.01846788078546524, + 0.018446946516633034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.02208799496293068 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525581045076251 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.00525581045076251, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525581045076251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1119263768196106, + 0.185602605342865, + 0.10789816081523895, + 0.08328435570001602, + 0.13253000378608704, + 0.066762775182724, + 0.1478862762451172, + 0.09272082895040512, + 0.023206477984786034, + 0.04818214476108551 + ], + "credit_map": [ + 0.014205358922481537, + 0.014102219603955746, + 0.018648721277713776, + 0.009509683586657047, + 0.015213276259601116, + 0.013345595449209213, + 0.010730353184044361, + 0.01617116667330265, + 0.03365100175142288, + 0.02553749457001686, + 0.027606338262557983, + 0.019021326676011086, + 0.01614445261657238, + 0.023087946698069572, + 0.013641338795423508, + 0.026912689208984375, + 0.01355732511729002, + 0.010475526563823223, + 0.013407919555902481, + 0.011188692413270473, + 0.022788794711232185, + 0.008010796271264553, + 0.015782581642270088, + 0.01268651895225048, + 0.008984641171991825, + 0.007950023747980595, + 0.0131678506731987, + 0.008748207241296768, + 0.009306744672358036, + 0.008803424425423145, + 0.013490090146660805, + 0.01283336989581585, + 0.013680923730134964, + 0.010693775489926338, + 0.022276122123003006, + 0.012111269868910313, + 0.014528006315231323, + 0.022583911195397377, + 0.013603251427412033, + 0.02305273897945881, + 0.009133984334766865, + 0.004611885640770197, + 0.011547909118235111, + 0.0051833949983119965, + 0.006884902715682983, + 0.015731099992990494, + 0.005768798291683197, + 0.007900804281234741, + 0.012338769622147083, + 0.012889942154288292, + 0.02903807908296585, + 0.011027762666344643, + 0.017427684739232063, + 0.015426119789481163, + 0.013018968515098095, + 0.036718953400850296, + 0.008513370528817177, + 0.010083942674100399, + 0.011410475708544254, + 0.015756899490952492, + 0.007854616269469261, + 0.012790421955287457, + 0.01326884888112545, + 0.013042256236076355, + 0.002019178355112672, + 0.0029922816902399063, + 0.00250809988938272, + 0.003266002284362912, + 0.0036238031461834908, + 0.003229772672057152, + 0.002616713522002101, + 0.0029506252612918615, + 0.006841772701591253, + 0.008801710791885853, + 0.006534731015563011, + 0.004652407951653004, + 0.007297571282833815, + 0.005003317259252071, + 0.004780288320034742, + 0.004270343109965324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.00525581045076251 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033631563186646 + ], + "opposite_logprob_tokens": [ + -1.9103362560272217 + ], + "logprob_reward": -0.16033631563186646, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033631563186646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10576443374156952, + 0.23421220481395721, + 0.07660989463329315, + 0.09124168008565903, + 0.08322062343358994, + 0.06785199791193008, + 0.04144503176212311, + 0.13856470584869385, + 0.03306369483470917, + 0.12802563607692719 + ], + "credit_map": [ + 0.019736818969249725, + 0.010021209716796875, + 0.018731873482465744, + 0.008846422657370567, + 0.007492189295589924, + 0.011279456317424774, + 0.009172647260129452, + 0.020483817905187607, + 0.056403305381536484, + 0.015117566101253033, + 0.02078859508037567, + 0.012149899266660213, + 0.015378636308014393, + 0.0791991725564003, + 0.01051531732082367, + 0.024659723043441772, + 0.008631681092083454, + 0.006926857866346836, + 0.012378999963402748, + 0.007738503161817789, + 0.010188021697103977, + 0.008285542950034142, + 0.008364269509911537, + 0.01409602165222168, + 0.01048625260591507, + 0.008369646035134792, + 0.010171494446694851, + 0.009658969938755035, + 0.011729392223060131, + 0.014569131657481194, + 0.013221466913819313, + 0.013035327196121216, + 0.007291192188858986, + 0.005450822412967682, + 0.015816563740372658, + 0.004408519249409437, + 0.00710398331284523, + 0.02581438422203064, + 0.004511810373514891, + 0.012823347002267838, + 0.008682570420205593, + 0.006489644292742014, + 0.010645722039043903, + 0.006693349685519934, + 0.005734992679208517, + 0.013194309547543526, + 0.0067964885383844376, + 0.009614921174943447, + 0.006875381339341402, + 0.00378938065841794, + 0.006415522191673517, + 0.004077409394085407, + 0.006518098991364241, + 0.006095017306506634, + 0.003647545352578163, + 0.00402667885646224, + 0.0128291305154562, + 0.015787512063980103, + 0.015433450229465961, + 0.018909698352217674, + 0.016132105141878128, + 0.01863311603665352, + 0.027003536000847816, + 0.01383616216480732, + 0.0032869630958884954, + 0.004337501712143421, + 0.003726996947079897, + 0.0037935227155685425, + 0.006770201027393341, + 0.003730406053364277, + 0.0038733058609068394, + 0.0035447983536869287, + 0.016988739371299744, + 0.025177275761961937, + 0.022751417011022568, + 0.012739892117679119, + 0.018086446449160576, + 0.010650635696947575, + 0.011707110330462456, + 0.009924136102199554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.16033631563186646 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066253297030926 + ], + "opposite_logprob_tokens": [ + -4.511066436767578 + ], + "logprob_reward": -0.011066253297030926, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066253297030926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18561102449893951, + 0.11753692477941513, + 0.07736009359359741, + 0.029100783169269562, + 0.11854188144207001, + 0.1999312788248062, + 0.04284188151359558, + 0.1726677417755127, + 0.02735356241464615, + 0.029054906219244003 + ], + "credit_map": [ + 0.015003886073827744, + 0.02287624217569828, + 0.02726043201982975, + 0.017908208072185516, + 0.031135857105255127, + 0.02462327666580677, + 0.024715663865208626, + 0.02208743244409561, + 0.019750414416193962, + 0.014021922834217548, + 0.013240461237728596, + 0.01107644010335207, + 0.014225772581994534, + 0.019398905336856842, + 0.013420377857983112, + 0.012402627617120743, + 0.009995149448513985, + 0.006835815962404013, + 0.012458785437047482, + 0.006906475871801376, + 0.014924640767276287, + 0.005852700211107731, + 0.010932950302958488, + 0.00945356860756874, + 0.0025451569817960262, + 0.002678824821487069, + 0.004111956804990768, + 0.0026439609937369823, + 0.002861212706193328, + 0.00546242343261838, + 0.0033810404129326344, + 0.00541620422154665, + 0.014124078676104546, + 0.008078593760728836, + 0.020722614601254463, + 0.007990331389009953, + 0.011627157218754292, + 0.024321429431438446, + 0.008202268742024899, + 0.023475397378206253, + 0.018444212153553963, + 0.013619322329759598, + 0.03544778376817703, + 0.017498893663287163, + 0.032215896993875504, + 0.03922688215970993, + 0.020407527685165405, + 0.023070743307471275, + 0.006508247461169958, + 0.004057515878230333, + 0.006150108762085438, + 0.004302613902837038, + 0.005206181202083826, + 0.004746587481349707, + 0.005675643682479858, + 0.0061949798837304115, + 0.017802558839321136, + 0.013399960473179817, + 0.026185542345046997, + 0.02595767378807068, + 0.010593806393444538, + 0.027953986078500748, + 0.015729209408164024, + 0.03504497930407524, + 0.003141402266919613, + 0.0031870678067207336, + 0.0049516139551997185, + 0.0019386272178962827, + 0.0038069530855864286, + 0.0057988218031823635, + 0.0022215123753994703, + 0.0023075605276972055, + 0.002828051568940282, + 0.004421180579811335, + 0.00364866154268384, + 0.0037645844276994467, + 0.004274892155081034, + 0.003395214444026351, + 0.003641656134277582, + 0.0030806621070951223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.011066253297030926 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a monkey wearing intricate golden armor with dragon patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monkey riding a swirling cloud that radiates an ethereal golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two slender purple-gold phoenix feathers visible swaying violently on the monkey's head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a vast mountain range with steep ridges covered in dark green pines and deep canyons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a sweeping aerial shot that swoops and rises while passing over the terrain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067903600633144 + ], + "opposite_logprob_tokens": [ + -4.511067867279053 + ], + "logprob_reward": -0.011067903600633144, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067903600633144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18825501203536987, + 0.14615695178508759, + 0.05865716561675072, + 0.15180809795856476, + 0.1170426458120346, + 0.10789945721626282, + 0.09178446978330612, + 0.05007914453744888, + 0.02518252283334732, + 0.06313449144363403 + ], + "credit_map": [ + 0.01622886024415493, + 0.0240765567868948, + 0.022019924595952034, + 0.0169330183416605, + 0.034903384745121, + 0.02485213242471218, + 0.026042813435196877, + 0.02319832146167755, + 0.022573266178369522, + 0.014950756914913654, + 0.011521296575665474, + 0.010263956151902676, + 0.014756997115910053, + 0.040487032383680344, + 0.015033518895506859, + 0.01657014526426792, + 0.008475269190967083, + 0.006688552908599377, + 0.008163715712726116, + 0.006920541636645794, + 0.006708838976919651, + 0.007416702341288328, + 0.005564600229263306, + 0.008718946948647499, + 0.017162632197141647, + 0.013525314629077911, + 0.021749388426542282, + 0.013799055479466915, + 0.021038569509983063, + 0.020865637809038162, + 0.022389646619558334, + 0.021277863532304764, + 0.011624865233898163, + 0.006339447572827339, + 0.02713785693049431, + 0.006173056084662676, + 0.008961334824562073, + 0.02012806572020054, + 0.005674103740602732, + 0.03100392036139965, + 0.013036434538662434, + 0.008012259379029274, + 0.018941368907690048, + 0.007785065099596977, + 0.015000365674495697, + 0.02597888745367527, + 0.008484245277941227, + 0.010660831816494465, + 0.011528367176651955, + 0.009069769643247128, + 0.016875797882676125, + 0.007114802021533251, + 0.0129899512976408, + 0.015752090141177177, + 0.010115785524249077, + 0.008337916806340218, + 0.0077413334511220455, + 0.004470492713153362, + 0.006714265793561935, + 0.006238466594368219, + 0.004136078525334597, + 0.00844744686037302, + 0.0057213143445551395, + 0.006609752774238586, + 0.0026850581634789705, + 0.0036359503865242004, + 0.0031193874310702085, + 0.0025688842870295048, + 0.003624188480898738, + 0.003232415299862623, + 0.003481512423604727, + 0.002835127990692854, + 0.00811056885868311, + 0.008976005017757416, + 0.005452738609164953, + 0.007963660173118114, + 0.01048604492098093, + 0.005576280876994133, + 0.009066143073141575, + 0.00750305550172925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.011067903600633144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_11/index_tensor([2693, 2693])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_11/index_tensor([2693, 2693])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8d8323a2cde1af151d6f0e1b8fdcfc15bff30344 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_11/index_tensor([2693, 2693])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide-angle shot, rendered in a gritty, comic-book style, slowly advances across a desolate, sun-drenched desert highway, where every element is defined by bold black ink outlines and stark, high-contrast shadows. Under the intense, unforgiving midday sun, a visible heat haze shimmers, distorting the air and rising from the cracked asphalt. At the center of this oppressive scene, a flock of Coragyps atratus feeds en masse. They are a writhing mass of black feathers, huddled over invisible carrion. Their featherless, swarthy heads are deeply wrinkled and leathery, and they viciously tear and pull with their sharp, hooked beaks, bobbing and weaving. Their powerful bodies cast deep, sharp black shadows on the scorching pavement, creating stark, graphic patterns. The vultures aggressively jockey for position, their movements primal and focused, creating a brutal, raw atmosphere of natural decay across the empty, sun-bleached road.", + "seed": 32, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_11/index_2693_seed_32_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13880870789289473, + "vlm_logprob": -0.13880870789289473, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.13880870789289473, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09149332344532013 + ], + "opposite_logprob_tokens": [ + -2.4664933681488037 + ], + "logprob_reward": -0.09149332344532013, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09149332344532013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17146019637584686, + 0.056642692536115646, + 0.15169404447078705, + 0.024785280227661133, + 0.1672125905752182, + 0.08815480768680573, + 0.15974503755569458, + 0.039492227137088776, + 0.10616862028837204, + 0.034644484519958496 + ], + "credit_map": [ + 0.02010882832109928, + 0.02719867043197155, + 0.02894735150039196, + 0.027131684124469757, + 0.01577175036072731, + 0.01815582811832428, + 0.020127706229686737, + 0.01401838194578886, + 0.009002210572361946, + 0.0076484596356749535, + 0.009036916308104992, + 0.008142535574734211, + 0.007049301639199257, + 0.004660137929022312, + 0.006365005858242512, + 0.004738124553114176, + 0.006749968975782394, + 0.00477613927796483, + 0.04074285179376602, + 0.017171934247016907, + 0.007946169935166836, + 0.004214515909552574, + 0.04079792648553848, + 0.029294533655047417, + 0.002519869478419423, + 0.0034424185287207365, + 0.002803332172334194, + 0.0039923833683133125, + 0.003410754259675741, + 0.0027841520495712757, + 0.0032935042399913073, + 0.0025388661306351423, + 0.00771151389926672, + 0.011799752712249756, + 0.007887383922934532, + 0.05812744051218033, + 0.009472462348639965, + 0.007442684844136238, + 0.053505074232816696, + 0.011266276240348816, + 0.01187355536967516, + 0.024881534278392792, + 0.0041748215444386005, + 0.00597029784694314, + 0.01607256755232811, + 0.007897959090769291, + 0.011390883475542068, + 0.00589319271966815, + 0.019054895266890526, + 0.00821766722947359, + 0.007454180624336004, + 0.012389358133077621, + 0.008786276914179325, + 0.03295043483376503, + 0.04147566854953766, + 0.029416562989354134, + 0.004272928461432457, + 0.0035480319056659937, + 0.00785925891250372, + 0.007234444376081228, + 0.004127573221921921, + 0.0037443451583385468, + 0.004102795384824276, + 0.004602849017828703, + 0.017946986481547356, + 0.02198733016848564, + 0.010176331736147404, + 0.00810662005096674, + 0.018916888162493706, + 0.012539236806333065, + 0.007030109409242868, + 0.009465114213526249, + 0.00483035808429122, + 0.005355477798730135, + 0.005998929962515831, + 0.0036957107950001955, + 0.0043419720605015755, + 0.003989193122833967, + 0.003305067541077733, + 0.003127774689346552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.09149332344532013 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048662614077329636 + ], + "opposite_logprob_tokens": [ + -3.0486626625061035 + ], + "logprob_reward": -0.048662614077329636, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048662614077329636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07854561507701874, + 0.057191696017980576, + 0.1135379895567894, + 0.022283557802438736, + 0.2349637746810913, + 0.09656082093715668, + 0.16777004301548004, + 0.12733091413974762, + 0.067250557243824, + 0.03456505760550499 + ], + "credit_map": [ + 0.006768283899873495, + 0.009349915198981762, + 0.010591947473585606, + 0.015280948020517826, + 0.005465523339807987, + 0.007760973647236824, + 0.0070046316832304, + 0.0163233894854784, + 0.008651314303278923, + 0.007767365779727697, + 0.010359440930187702, + 0.008435623720288277, + 0.005750623997300863, + 0.005270158872008324, + 0.0055272579193115234, + 0.005429905839264393, + 0.005634799133986235, + 0.003471993375569582, + 0.04739522933959961, + 0.013024440966546535, + 0.005152053665369749, + 0.0032072430476546288, + 0.028160464018583298, + 0.007491755299270153, + 0.002776915207505226, + 0.0030262023210525513, + 0.0024890126660466194, + 0.004450619220733643, + 0.0023631586227566004, + 0.002417858224362135, + 0.002636749530211091, + 0.00212303944863379, + 0.015247867442667484, + 0.018354488536715508, + 0.017951538786292076, + 0.06763017922639847, + 0.014226764440536499, + 0.012089436873793602, + 0.06789029389619827, + 0.021573197096586227, + 0.006839563138782978, + 0.007007142994552851, + 0.004133460111916065, + 0.025559628382325172, + 0.009502594359219074, + 0.006478574592620134, + 0.0297053474932909, + 0.007334505673497915, + 0.02047877572476864, + 0.013150476850569248, + 0.00969111267477274, + 0.02750973589718342, + 0.007987214252352715, + 0.025400418788194656, + 0.02696780487895012, + 0.036584511399269104, + 0.009980056434869766, + 0.006912959273904562, + 0.014924976043403149, + 0.015533242374658585, + 0.009097598493099213, + 0.008096467703580856, + 0.024886656552553177, + 0.0378989540040493, + 0.010423475876450539, + 0.014176291413605213, + 0.0064687589183449745, + 0.006558221764862537, + 0.010757898911833763, + 0.0077730948105454445, + 0.005810961127281189, + 0.005281856283545494, + 0.005648913327604532, + 0.006502573378384113, + 0.003032160922884941, + 0.0038302652537822723, + 0.005733456462621689, + 0.0036994426045566797, + 0.0034402175806462765, + 0.0026780269108712673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.048662614077329636 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497201159596443 + ], + "opposite_logprob_tokens": [ + -2.929971933364868 + ], + "logprob_reward": -0.05497201159596443, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497201159596443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08162630349397659, + 0.1347789466381073, + 0.16390009224414825, + 0.09692247211933136, + 0.15562617778778076, + 0.07312579452991486, + 0.1420324444770813, + 0.0694170817732811, + 0.035032738000154495, + 0.047537945210933685 + ], + "credit_map": [ + 0.009548156522214413, + 0.01109262090176344, + 0.013037974946200848, + 0.011291141621768475, + 0.006313797552138567, + 0.009754815138876438, + 0.009061811491847038, + 0.011525987647473812, + 0.016119034960865974, + 0.013466611504554749, + 0.024996886029839516, + 0.019742732867598534, + 0.01571066677570343, + 0.014960027299821377, + 0.01503298431634903, + 0.014750003814697266, + 0.0072618103586137295, + 0.006126587279140949, + 0.04683737829327583, + 0.035622261464595795, + 0.006370130926370621, + 0.004189212806522846, + 0.02835705876350403, + 0.029135651886463165, + 0.008576761931180954, + 0.006589281838387251, + 0.009187967516481876, + 0.02206840179860592, + 0.007787118665874004, + 0.007017181254923344, + 0.008167309686541557, + 0.027528448030352592, + 0.011333892121911049, + 0.018765047192573547, + 0.0127158397808671, + 0.037590812891721725, + 0.012289825826883316, + 0.011072354391217232, + 0.03636852279305458, + 0.015489884652197361, + 0.009039787575602531, + 0.028022579848766327, + 0.0023284379858523607, + 0.010622289963066578, + 0.006049074698239565, + 0.006158730015158653, + 0.006512179039418697, + 0.004392716567963362, + 0.020960617810487747, + 0.008646413683891296, + 0.006383108906447887, + 0.01622658781707287, + 0.007144508417695761, + 0.03876860812306404, + 0.022907428443431854, + 0.020995162427425385, + 0.005271999165415764, + 0.0035531725734472275, + 0.008540784008800983, + 0.00834882352501154, + 0.005269431974738836, + 0.004395336844027042, + 0.01351398229598999, + 0.020523549988865852, + 0.006400425452739, + 0.005889663938432932, + 0.0031942008063197136, + 0.0030697695910930634, + 0.006011555436998606, + 0.004494570195674896, + 0.0028022308833897114, + 0.003170320298522711, + 0.011591905727982521, + 0.008813987486064434, + 0.005327803082764149, + 0.004389642737805843, + 0.005547177046537399, + 0.004070436581969261, + 0.00509287603199482, + 0.002704115118831396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.05497201159596443 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288995563983917 + ], + "opposite_logprob_tokens": [ + -1.0538995265960693 + ], + "logprob_reward": -0.4288995563983917, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288995563983917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06437312811613083, + 0.04689335823059082, + 0.18813210725784302, + 0.03651609271764755, + 0.07700595259666443, + 0.1728072464466095, + 0.09591690450906754, + 0.03210930898785591, + 0.23182044923305511, + 0.0544254370033741 + ], + "credit_map": [ + 0.009054647758603096, + 0.011423992924392223, + 0.00826717633754015, + 0.008427457883954048, + 0.006113063544034958, + 0.0063364580273628235, + 0.007091250270605087, + 0.007659084629267454, + 0.007006199564784765, + 0.0068818056024611, + 0.004356148187071085, + 0.006323181092739105, + 0.006008600816130638, + 0.0068069687113165855, + 0.005661603529006243, + 0.0038488516584038734, + 0.008338327519595623, + 0.006298520602285862, + 0.04913632944226265, + 0.03425084426999092, + 0.008922155946493149, + 0.006730058696120977, + 0.049762580543756485, + 0.024693284183740616, + 0.00358089548535645, + 0.004067051224410534, + 0.0032919212244451046, + 0.010230101644992828, + 0.0032803539652377367, + 0.0023354862350970507, + 0.003626411547884345, + 0.006103869061917067, + 0.00429716007784009, + 0.005923978518694639, + 0.004383508116006851, + 0.024289553984999657, + 0.0041405814699828625, + 0.003808878129348159, + 0.024330096319317818, + 0.005832199472934008, + 0.01383510883897543, + 0.050540741533041, + 0.005517673213034868, + 0.05603143200278282, + 0.009172271937131882, + 0.007305232807993889, + 0.024659840390086174, + 0.005744941998273134, + 0.009251493029296398, + 0.01404123567044735, + 0.005112174432724714, + 0.009773205034434795, + 0.004828951321542263, + 0.014839829877018929, + 0.015430789440870285, + 0.022639229893684387, + 0.0028904934879392385, + 0.001952323131263256, + 0.003691465128213167, + 0.003687467658892274, + 0.0028393324464559555, + 0.002477819100022316, + 0.005958279594779015, + 0.008612128905951977, + 0.03999188169836998, + 0.04855826497077942, + 0.020602330565452576, + 0.022200513631105423, + 0.038297876715660095, + 0.02579333819448948, + 0.017589885741472244, + 0.018786365166306496, + 0.008826668374240398, + 0.00847262516617775, + 0.009961439296603203, + 0.005457533989101648, + 0.007026663050055504, + 0.005872426554560661, + 0.0045289588160812855, + 0.004279123619198799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.4288995563983917 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700160339474678 + ], + "opposite_logprob_tokens": [ + -2.6950161457061768 + ], + "logprob_reward": -0.0700160339474678, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700160339474678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06610534340143204, + 0.020296167582273483, + 0.18340516090393066, + 0.07131225615739822, + 0.12336532026529312, + 0.32610663771629333, + 0.09264793992042542, + 0.025639163330197334, + 0.032487768679857254, + 0.05863421410322189 + ], + "credit_map": [ + 0.005487210117280483, + 0.008234908804297447, + 0.011132937856018543, + 0.01011582650244236, + 0.007246060762554407, + 0.007687238976359367, + 0.00679897191002965, + 0.009402184747159481, + 0.0020672117825597525, + 0.002275231759995222, + 0.003747951937839389, + 0.0029129572212696075, + 0.0019776830449700356, + 0.002820280846208334, + 0.002458516275510192, + 0.002036335878074169, + 0.008557231165468693, + 0.00880469661206007, + 0.014331183396279812, + 0.05982041731476784, + 0.005311962217092514, + 0.009208458475768566, + 0.010990148410201073, + 0.0663810670375824, + 0.004793703090399504, + 0.005397546570748091, + 0.005143188871443272, + 0.018414204940199852, + 0.0044467868283391, + 0.0063314917497336864, + 0.004836108535528183, + 0.02194923162460327, + 0.011708050034940243, + 0.014446999877691269, + 0.006114603020250797, + 0.040303174406290054, + 0.010358982719480991, + 0.008083080872893333, + 0.025701725855469704, + 0.006648707669228315, + 0.05657738819718361, + 0.1422627568244934, + 0.008586381562054157, + 0.05178876221179962, + 0.02244543470442295, + 0.021400228142738342, + 0.009402396157383919, + 0.013643285259604454, + 0.009687365964055061, + 0.012354288250207901, + 0.0036401201505213976, + 0.004870263393968344, + 0.005115481559187174, + 0.022413544356822968, + 0.015343725681304932, + 0.019223155453801155, + 0.003090630518272519, + 0.002620924264192581, + 0.0035419664345681667, + 0.004239019937813282, + 0.0034119051415473223, + 0.002794711384922266, + 0.003863212652504444, + 0.00207679346203804, + 0.0057692877016961575, + 0.006716266740113497, + 0.0028695138171315193, + 0.003067400772124529, + 0.005697314161807299, + 0.0033649131655693054, + 0.0024370511528104544, + 0.0025660197716206312, + 0.012526796199381351, + 0.011146808043122292, + 0.009411873295903206, + 0.006600336171686649, + 0.005312419030815363, + 0.005948122125118971, + 0.004209455568343401, + 0.0034784036688506603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.0700160339474678 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot, rendered in a gritty, comic-book style, slowly advances across a desolate, sun-drenched desert highway, where every element is defined by bold black ink outlines and stark, high-contrast shadows. Under the intense, unforgiving midday sun, a visible heat haze shimmers, distorting the air and rising from the cracked asphalt. At the center of this oppressive scene, a flock of Coragyps atratus feeds en masse. They are a writhing mass of black feathers, huddled over invisible carrion. Their featherless, swarthy heads are deeply wrinkled and leathery, and they viciously tear and pull with their sharp, hooked beaks, bobbing and weaving. Their powerful bodies cast deep, sharp black shadows on the scorching pavement, creating stark, graphic patterns. The vultures aggressively jockey for position, their movements primal and focused, creating a brutal, raw atmosphere of natural decay across the empty, sun-bleached road.", + "seed": 33, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_11/index_2693_seed_33_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06644489746540785, + "vlm_logprob": -0.06644489746540785, + "vlm_yes_no_margin": 3.0999999761581423, + "vlm_token_credit": -0.06644489746540785, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017835358157753944 + ], + "opposite_logprob_tokens": [ + -4.2678351402282715 + ], + "logprob_reward": -0.017835358157753944, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017835358157753944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1514570713043213, + 0.15959559381008148, + 0.1021747812628746, + 0.09368154406547546, + 0.088178351521492, + 0.028877774253487587, + 0.020078329369425774, + 0.02111300453543663, + 0.27659356594085693, + 0.05824996903538704 + ], + "credit_map": [ + 0.008266608230769634, + 0.00865662656724453, + 0.02636078931391239, + 0.011285915039479733, + 0.009813053533434868, + 0.007328209467232227, + 0.056848980486392975, + 0.02289688028395176, + 0.005896777380257845, + 0.007898391224443913, + 0.013113776221871376, + 0.053823746740818024, + 0.006142203230410814, + 0.004969384055584669, + 0.061822377145290375, + 0.00592893036082387, + 0.009729468263685703, + 0.014054710976779461, + 0.01846722699701786, + 0.010654033161699772, + 0.00889920350164175, + 0.013018129393458366, + 0.020831363275647163, + 0.0065206498838961124, + 0.00721703702583909, + 0.023999065160751343, + 0.026280587539076805, + 0.005900734569877386, + 0.004268247168511152, + 0.00788767822086811, + 0.005348966922610998, + 0.012779227457940578, + 0.01486560981720686, + 0.0140247056260705, + 0.009510583244264126, + 0.01003014761954546, + 0.006981047801673412, + 0.011461107060313225, + 0.011659429408609867, + 0.009645719081163406, + 0.0023913870099931955, + 0.0032832880970090628, + 0.0028478866443037987, + 0.0030271972063928843, + 0.003241980681195855, + 0.005778010468930006, + 0.0035018264316022396, + 0.00480619678273797, + 0.0029253927059471607, + 0.0018474315293133259, + 0.004105341155081987, + 0.0017367691034451127, + 0.0016818118747323751, + 0.002140598138794303, + 0.0023335558362305164, + 0.003307430073618889, + 0.003930188715457916, + 0.0027560992166399956, + 0.002831167308613658, + 0.0018307865830138326, + 0.0014850706793367863, + 0.0026521107647567987, + 0.0025711774360388517, + 0.0030564044136554003, + 0.055127259343862534, + 0.007795346435159445, + 0.11106175184249878, + 0.013172763399779797, + 0.013309857808053493, + 0.00829846691340208, + 0.026612242683768272, + 0.041215892881155014, + 0.007990353740751743, + 0.005910155363380909, + 0.009988189674913883, + 0.005545809864997864, + 0.008452778682112694, + 0.006338923238217831, + 0.006830895785242319, + 0.00719286035746336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.017835358157753944 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543593533337116 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543593533337116, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543593533337116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1991409957408905, + 0.10299281775951385, + 0.06548947840929031, + 0.22397468984127045, + 0.05722350627183914, + 0.028202328830957413, + 0.009305892512202263, + 0.010014178231358528, + 0.26140111684799194, + 0.0422549843788147 + ], + "credit_map": [ + 0.014253538101911545, + 0.006782200187444687, + 0.02025873027741909, + 0.017516881227493286, + 0.008024024777114391, + 0.01459530834108591, + 0.06633280962705612, + 0.051377490162849426, + 0.005037867929786444, + 0.007119978312402964, + 0.010483408346772194, + 0.028731293976306915, + 0.004950498230755329, + 0.00512907886877656, + 0.03633388504385948, + 0.005206804722547531, + 0.007048046216368675, + 0.005527547560632229, + 0.01359385997056961, + 0.006311090663075447, + 0.005077857058495283, + 0.004938075318932533, + 0.017506958916783333, + 0.005486042704433203, + 0.012207512743771076, + 0.06954006105661392, + 0.06432466208934784, + 0.013699602335691452, + 0.01017774548381567, + 0.018918683752417564, + 0.009675794281065464, + 0.02543061040341854, + 0.0071443840861320496, + 0.007990339770913124, + 0.008251271210610867, + 0.006708418484777212, + 0.0045906719751656055, + 0.006370088085532188, + 0.01100823562592268, + 0.005160090979188681, + 0.0015526963397860527, + 0.0023406913969665766, + 0.0014720511389896274, + 0.0024432858917862177, + 0.002042824402451515, + 0.007303173188120127, + 0.0032179094851017, + 0.0078296959400177, + 0.000990786124020815, + 0.0011381160002201796, + 0.0016123753739520907, + 0.0013149696169421077, + 0.0007666867459192872, + 0.001193398260511458, + 0.0011949815088883042, + 0.0010945782996714115, + 0.0011716869194060564, + 0.0012232103617861867, + 0.0017416093032807112, + 0.0012975321151316166, + 0.0006666032131761312, + 0.0014887015568092465, + 0.001076694461517036, + 0.001348139951005578, + 0.033266302198171616, + 0.006929656025022268, + 0.11028214544057846, + 0.01150762103497982, + 0.008801467716693878, + 0.0058953347615897655, + 0.04755287989974022, + 0.03716567903757095, + 0.005022470839321613, + 0.004782866220921278, + 0.006094351876527071, + 0.004917507525533438, + 0.004701514262706041, + 0.0045446776784956455, + 0.0055642458610236645, + 0.00662734592333436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.012543593533337116 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003770768642426 + ], + "opposite_logprob_tokens": [ + -2.695037603378296 + ], + "logprob_reward": -0.07003770768642426, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003770768642426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4304398000240326, + 0.10245548188686371, + 0.057167019695043564, + 0.056022293865680695, + 0.06266993284225464, + 0.03961504250764847, + 0.009318912401795387, + 0.010692792944610119, + 0.18191179633140564, + 0.049706920981407166 + ], + "credit_map": [ + 0.023457825183868408, + 0.015167148783802986, + 0.058492086827754974, + 0.03261347860097885, + 0.01664019376039505, + 0.02744481712579727, + 0.16936668753623962, + 0.08725756406784058, + 0.003543806727975607, + 0.0039466870948672295, + 0.008338838815689087, + 0.035972557961940765, + 0.0026706152129918337, + 0.0030608316883444786, + 0.04064008593559265, + 0.004282062407582998, + 0.007232070900499821, + 0.008021192625164986, + 0.008211052976548672, + 0.0050352285616099834, + 0.00623368052765727, + 0.007712882477790117, + 0.010105789639055729, + 0.004615123849362135, + 0.007003354374319315, + 0.016636857762932777, + 0.011311182752251625, + 0.0049668787978589535, + 0.0035885535180568695, + 0.004077717196196318, + 0.00498221255838871, + 0.003455536440014839, + 0.00779507914558053, + 0.007248383481055498, + 0.012406996451318264, + 0.007400532253086567, + 0.003953009843826294, + 0.006625478621572256, + 0.011565307155251503, + 0.00567514868453145, + 0.0022445321083068848, + 0.0024365587159991264, + 0.00194421480409801, + 0.0034405854530632496, + 0.00195637671276927, + 0.012105398811399937, + 0.0033272337168455124, + 0.012160141952335835, + 0.0009906744817271829, + 0.0013922805665060878, + 0.0016617041546851397, + 0.001154451398178935, + 0.0007589203887619078, + 0.001282648416236043, + 0.0012593775754794478, + 0.0008188557112589478, + 0.0014834137400612235, + 0.0013522827066481113, + 0.0015797773376107216, + 0.0008649700903333724, + 0.0009144509676843882, + 0.0019726392347365618, + 0.0011275098659098148, + 0.0013977487105876207, + 0.02618005871772766, + 0.006262896116822958, + 0.07439307868480682, + 0.008414401672780514, + 0.006173109635710716, + 0.004698388744145632, + 0.02939261868596077, + 0.026397231966257095, + 0.006158546078950167, + 0.005063266027718782, + 0.008443454280495644, + 0.005173900630325079, + 0.004966932348906994, + 0.0052026123739778996, + 0.0067184120416641235, + 0.007979795336723328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.07003770768642426 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279301464557648 + ], + "opposite_logprob_tokens": [ + -2.0177929401397705 + ], + "logprob_reward": -0.14279301464557648, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14279301464557648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11805085837841034, + 0.24853229522705078, + 0.08489334583282471, + 0.18370461463928223, + 0.05041835457086563, + 0.08945472538471222, + 0.03722378611564636, + 0.021467193961143494, + 0.12908904254436493, + 0.0371658094227314 + ], + "credit_map": [ + 0.011099586263298988, + 0.007058961782604456, + 0.04363790154457092, + 0.00735219893977046, + 0.005587334278970957, + 0.008589225821197033, + 0.021433966234326363, + 0.013291670940816402, + 0.009479503147304058, + 0.013110850937664509, + 0.019453616812825203, + 0.08681200444698334, + 0.007036198396235704, + 0.009968647733330727, + 0.09297757595777512, + 0.009693887084722519, + 0.007982674054801464, + 0.014656261540949345, + 0.010569410398602486, + 0.01128841470927, + 0.007732035592198372, + 0.007855302654206753, + 0.018021386116743088, + 0.006787853315472603, + 0.008521816693246365, + 0.059020258486270905, + 0.05256171524524689, + 0.01362333819270134, + 0.008486464619636536, + 0.015502373687922955, + 0.006562737748026848, + 0.019425898790359497, + 0.005379462614655495, + 0.007082812953740358, + 0.00821468885987997, + 0.0061482093296945095, + 0.004585367627441883, + 0.004948457702994347, + 0.009850265458226204, + 0.004209083504974842, + 0.006096121855080128, + 0.007560489699244499, + 0.007136011030524969, + 0.008476408198475838, + 0.008112621493637562, + 0.02129715122282505, + 0.010494356974959373, + 0.020281556993722916, + 0.004944243468344212, + 0.003996624611318111, + 0.007753301411867142, + 0.003348077181726694, + 0.0032913589384406805, + 0.004018954001367092, + 0.0038861020002514124, + 0.005985121708363295, + 0.0033898765686899424, + 0.002903139218688011, + 0.003442285815253854, + 0.0017618467099964619, + 0.0018498608842492104, + 0.0033416959922760725, + 0.0022950973361730576, + 0.002483389340341091, + 0.013690146617591381, + 0.0072744181379675865, + 0.04942256957292557, + 0.006614107638597488, + 0.003327235346660018, + 0.003581185359507799, + 0.027528464794158936, + 0.01765090599656105, + 0.004285317845642567, + 0.004690700676292181, + 0.0051732491701841354, + 0.003166717244312167, + 0.004947269801050425, + 0.0039015626534819603, + 0.004977208096534014, + 0.006023779045790434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.14279301464557648 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901481330394745 + ], + "opposite_logprob_tokens": [ + -2.464014768600464 + ], + "logprob_reward": -0.08901481330394745, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901481330394745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17198382318019867, + 0.19397704303264618, + 0.02517370693385601, + 0.17696401476860046, + 0.0876665860414505, + 0.05584936961531639, + 0.014093237929046154, + 0.02688840590417385, + 0.2240157425403595, + 0.02338811568915844 + ], + "credit_map": [ + 0.0107265068218112, + 0.009351222775876522, + 0.0587848424911499, + 0.009228923358023167, + 0.00913727656006813, + 0.0116021279245615, + 0.05334717407822609, + 0.00980574544519186, + 0.005299564450979233, + 0.008594110608100891, + 0.012730288319289684, + 0.07302089780569077, + 0.005474565085023642, + 0.006117863114923239, + 0.07483876496553421, + 0.007901000790297985, + 0.0025811856612563133, + 0.004662283696234226, + 0.00270567717961967, + 0.002877757651731372, + 0.002629687311127782, + 0.0035383901558816433, + 0.00434646662324667, + 0.0018322588875889778, + 0.010318807326257229, + 0.05874527990818024, + 0.040601905435323715, + 0.013316064141690731, + 0.009799919091165066, + 0.01550297811627388, + 0.01438251230865717, + 0.014296558685600758, + 0.010287944227457047, + 0.013623476959764957, + 0.016301840543746948, + 0.012173553928732872, + 0.008164876140654087, + 0.007415891159325838, + 0.013886711560189724, + 0.0058122961781919, + 0.0027068296913057566, + 0.003846315201371908, + 0.002334947930648923, + 0.004744352772831917, + 0.0020915428176522255, + 0.017949288710951805, + 0.00473142322152853, + 0.0174446702003479, + 0.0017491562757641077, + 0.00181910814717412, + 0.002386739943176508, + 0.0019289430929347873, + 0.0014431687304750085, + 0.0012101837201043963, + 0.001946991658769548, + 0.001608947291970253, + 0.0035372753627598286, + 0.002389263827353716, + 0.003618896473199129, + 0.0018335432978346944, + 0.0015852219657972455, + 0.004411212634295225, + 0.003718694904819131, + 0.005794295109808445, + 0.04467878118157387, + 0.006747047416865826, + 0.0890255868434906, + 0.010576545260846615, + 0.011496501043438911, + 0.007164608221501112, + 0.022896386682987213, + 0.03143027424812317, + 0.0024675584863871336, + 0.002471236977726221, + 0.0033325268886983395, + 0.0028791932854801416, + 0.0022281331475824118, + 0.003691867459565401, + 0.0028792214579880238, + 0.003438378684222698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.08901481330394745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot, rendered in a gritty, comic-book style, slowly advances across a desolate, sun-drenched desert highway, where every element is defined by bold black ink outlines and stark, high-contrast shadows. Under the intense, unforgiving midday sun, a visible heat haze shimmers, distorting the air and rising from the cracked asphalt. At the center of this oppressive scene, a flock of Coragyps atratus feeds en masse. They are a writhing mass of black feathers, huddled over invisible carrion. Their featherless, swarthy heads are deeply wrinkled and leathery, and they viciously tear and pull with their sharp, hooked beaks, bobbing and weaving. Their powerful bodies cast deep, sharp black shadows on the scorching pavement, creating stark, graphic patterns. The vultures aggressively jockey for position, their movements primal and focused, creating a brutal, raw atmosphere of natural decay across the empty, sun-bleached road.", + "seed": 34, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_11/index_2693_seed_34_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.061316514201462266, + "vlm_logprob": -0.061316514201462266, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.061316514201462266, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009581885300576687 + ], + "opposite_logprob_tokens": [ + -5.134582042694092 + ], + "logprob_reward": -0.009581885300576687, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009581885300576687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28020307421684265, + 0.16303876042366028, + 0.029793713241815567, + 0.05965717136859894, + 0.04716508463025093, + 0.03979770839214325, + 0.060160376131534576, + 0.08511006832122803, + 0.13325360417366028, + 0.10182036459445953 + ], + "credit_map": [ + 0.05209507793188095, + 0.044377706944942474, + 0.032433491200208664, + 0.04582839459180832, + 0.0239995326846838, + 0.02494898810982704, + 0.03470400720834732, + 0.02181587740778923, + 0.0417628176510334, + 0.022069672122597694, + 0.015695862472057343, + 0.018055709078907967, + 0.017818957567214966, + 0.021583586931228638, + 0.010924612171947956, + 0.015127560123801231, + 0.005101396702229977, + 0.002457659924402833, + 0.002512576524168253, + 0.004656119737774134, + 0.0029406833928078413, + 0.006567843724042177, + 0.0021075536496937275, + 0.003449880052357912, + 0.006978905294090509, + 0.007390943821519613, + 0.008462885394692421, + 0.010105145163834095, + 0.007494758348912001, + 0.006215832196176052, + 0.00589970126748085, + 0.007109001744538546, + 0.0046089724637568, + 0.004317335784435272, + 0.005500706844031811, + 0.012164812535047531, + 0.004151810891926289, + 0.002356089884415269, + 0.009216171689331532, + 0.004849182907491922, + 0.003342348150908947, + 0.002503864699974656, + 0.002661301987245679, + 0.013450466096401215, + 0.0024311637971550226, + 0.002490841317921877, + 0.009499465115368366, + 0.0034182555973529816, + 0.008282438851892948, + 0.007555018644779921, + 0.004391690716147423, + 0.007141199428588152, + 0.006392550654709339, + 0.007365152705460787, + 0.01371069811284542, + 0.005321627017110586, + 0.009136742912232876, + 0.009048786014318466, + 0.008134515956044197, + 0.020107552409172058, + 0.00664833327755332, + 0.00727760000154376, + 0.011778723448514938, + 0.012977814301848412, + 0.01689455658197403, + 0.014300709590315819, + 0.010976394638419151, + 0.03526534512639046, + 0.010509645566344261, + 0.011537629179656506, + 0.0241820327937603, + 0.009587295353412628, + 0.017101353034377098, + 0.010603453032672405, + 0.010484800674021244, + 0.015782156959176064, + 0.009890681132674217, + 0.013251915574073792, + 0.013019994832575321, + 0.011686012148857117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.009581885300576687 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070948913693428 + ], + "opposite_logprob_tokens": [ + -4.14107084274292 + ], + "logprob_reward": -0.016070948913693428, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070948913693428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11162404716014862, + 0.21657313406467438, + 0.017573827877640724, + 0.059561505913734436, + 0.07798145711421967, + 0.03468458727002144, + 0.027423100546002388, + 0.04802751541137695, + 0.2542700171470642, + 0.15228082239627838 + ], + "credit_map": [ + 0.019668616354465485, + 0.020424915477633476, + 0.010476497933268547, + 0.016002777963876724, + 0.01155747752636671, + 0.010816920548677444, + 0.014991735108196735, + 0.007685099262744188, + 0.05090770870447159, + 0.02919636107981205, + 0.024212148040533066, + 0.023860137909650803, + 0.023810576647520065, + 0.029457638040184975, + 0.015041468665003777, + 0.020087098702788353, + 0.002723808167502284, + 0.0026385453529655933, + 0.0025124638341367245, + 0.0018412237986922264, + 0.0023015157785266638, + 0.0021060726139694452, + 0.001856963848695159, + 0.001593233784660697, + 0.007052568718791008, + 0.007581024896353483, + 0.00716012017801404, + 0.011056949384510517, + 0.006040203385055065, + 0.0060666524805128574, + 0.007057341281324625, + 0.007546640001237392, + 0.00894218496978283, + 0.008311099372804165, + 0.007225973065942526, + 0.015618007630109787, + 0.00818986538797617, + 0.004539159592241049, + 0.01662985421717167, + 0.0085253044962883, + 0.0020399584900587797, + 0.0023232868406921625, + 0.0025489970576018095, + 0.010278237983584404, + 0.002480149269104004, + 0.0028064274229109287, + 0.009111432358622551, + 0.0030960955191403627, + 0.004283761139959097, + 0.003836223855614662, + 0.0026657800190150738, + 0.0034168842248618603, + 0.0037593746092170477, + 0.0027418341487646103, + 0.004291198682039976, + 0.002428041072562337, + 0.006404928397387266, + 0.007438561413437128, + 0.004430954344570637, + 0.0051668849773705006, + 0.00823984295129776, + 0.005525978282094002, + 0.006608316209167242, + 0.004212042782455683, + 0.027348719537258148, + 0.02862263284623623, + 0.019093455746769905, + 0.06961755454540253, + 0.017333125695586205, + 0.023864256218075752, + 0.04618655517697334, + 0.022203700616955757, + 0.020971523597836494, + 0.014951755292713642, + 0.0163285993039608, + 0.027477843686938286, + 0.01381661370396614, + 0.017864173278212547, + 0.018788959830999374, + 0.022081339731812477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.016070948913693428 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905833214521408 + ], + "opposite_logprob_tokens": [ + -2.4640583992004395 + ], + "logprob_reward": -0.08905833214521408, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905833214521408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15819033980369568, + 0.13999994099140167, + 0.031838309019804, + 0.04356897622346878, + 0.09816551953554153, + 0.09260621666908264, + 0.028789881616830826, + 0.05517829582095146, + 0.17820724844932556, + 0.17345531284809113 + ], + "credit_map": [ + 0.027879970148205757, + 0.02656053751707077, + 0.015929361805319786, + 0.0263985488563776, + 0.013855901546776295, + 0.014964630827307701, + 0.02155175991356373, + 0.011049631983041763, + 0.020151222124695778, + 0.01765136606991291, + 0.019920622929930687, + 0.018823759630322456, + 0.015524950809776783, + 0.019063692539930344, + 0.011075975373387337, + 0.017788348719477654, + 0.003732362063601613, + 0.003361423732712865, + 0.0028300199192017317, + 0.003145101945847273, + 0.004651259630918503, + 0.007513184100389481, + 0.0037805107422173023, + 0.0028244450222700834, + 0.006297380663454533, + 0.0036638074088841677, + 0.007215449586510658, + 0.0075128087773919106, + 0.004764590412378311, + 0.00422935513779521, + 0.00607686210423708, + 0.003808721899986267, + 0.010909054428339005, + 0.010214601643383503, + 0.010354485362768173, + 0.018220795318484306, + 0.010479137301445007, + 0.005940005648881197, + 0.02010488137602806, + 0.011942551471292973, + 0.005691316444426775, + 0.006203466095030308, + 0.003873682115226984, + 0.030977992340922356, + 0.004194341134279966, + 0.004216835834085941, + 0.03198598697781563, + 0.0054626003839075565, + 0.003223407780751586, + 0.003615080378949642, + 0.002501947805285454, + 0.003776280442252755, + 0.0037652808241546154, + 0.003231152892112732, + 0.005617058370262384, + 0.0030596707947552204, + 0.00619516521692276, + 0.007208550814539194, + 0.0048078736290335655, + 0.00684655737131834, + 0.010820516385138035, + 0.0059066456742584705, + 0.007904022932052612, + 0.005488961469382048, + 0.02123098634183407, + 0.018782654777169228, + 0.013437382876873016, + 0.04694351181387901, + 0.012614980340003967, + 0.01576557755470276, + 0.033888280391693115, + 0.01554386131465435, + 0.02436761185526848, + 0.01640130765736103, + 0.018129555508494377, + 0.029642755165696144, + 0.01678118109703064, + 0.02066800370812416, + 0.021825386211276054, + 0.025639502331614494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.08905833214521408 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896344363689423 + ], + "opposite_logprob_tokens": [ + -2.5789635181427 + ], + "logprob_reward": -0.07896344363689423, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896344363689423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10139348357915878, + 0.16439400613307953, + 0.08008143305778503, + 0.034733373671770096, + 0.1557396948337555, + 0.10098812729120255, + 0.05466854199767113, + 0.17178523540496826, + 0.06694718450307846, + 0.06926897168159485 + ], + "credit_map": [ + 0.027436593547463417, + 0.012147747911512852, + 0.014593376778066158, + 0.011373069137334824, + 0.007489252369850874, + 0.009265931323170662, + 0.009378032758831978, + 0.00970946904271841, + 0.046121302992105484, + 0.016793329268693924, + 0.01822468265891075, + 0.014731922186911106, + 0.01638227328658104, + 0.02046784944832325, + 0.011776461265981197, + 0.01989615522325039, + 0.008889887481927872, + 0.013093089684844017, + 0.0069018336944282055, + 0.012651130557060242, + 0.012881411239504814, + 0.009287617169320583, + 0.01035008579492569, + 0.006026366259902716, + 0.003356631612405181, + 0.0036930462811142206, + 0.004029886331409216, + 0.008167632855474949, + 0.00400480767711997, + 0.0038640592247247696, + 0.0043036616407334805, + 0.0033136429265141487, + 0.010075700469315052, + 0.008521112613379955, + 0.012496935203671455, + 0.05416867136955261, + 0.007891952991485596, + 0.005564103368669748, + 0.04582711309194565, + 0.011194095015525818, + 0.006327727809548378, + 0.0054551842622458935, + 0.005616555921733379, + 0.03589576482772827, + 0.004817134700715542, + 0.004379805643111467, + 0.03243701159954071, + 0.006058929022401571, + 0.006048270501196384, + 0.007997775450348854, + 0.005091555882245302, + 0.0070888628251850605, + 0.010244131088256836, + 0.006229793187230825, + 0.0076535968109965324, + 0.004314549267292023, + 0.018654676154255867, + 0.02032632566988468, + 0.017201296985149384, + 0.03281347081065178, + 0.018686523661017418, + 0.01945505477488041, + 0.022730937227606773, + 0.02191692404448986, + 0.006869926583021879, + 0.006982934195548296, + 0.005810331087559462, + 0.014724846929311752, + 0.005856998264789581, + 0.0059700519777834415, + 0.01197114959359169, + 0.00876094214618206, + 0.009652768261730671, + 0.006533695384860039, + 0.009540151804685593, + 0.009344527497887611, + 0.007365991361439228, + 0.0077978335320949554, + 0.009789055213332176, + 0.009244943037629128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.07896344363689423 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290796101093292 + ], + "opposite_logprob_tokens": [ + -2.237907886505127 + ], + "logprob_reward": -0.11290796101093292, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290796101093292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13070961833000183, + 0.1961759775876999, + 0.044153857976198196, + 0.08878854662179947, + 0.05138513818383217, + 0.11234888434410095, + 0.09037917107343674, + 0.14727096259593964, + 0.058115359395742416, + 0.08067256212234497 + ], + "credit_map": [ + 0.026321904733777046, + 0.01640106737613678, + 0.014160754159092903, + 0.015522257424890995, + 0.014921655878424644, + 0.015958556905388832, + 0.015512119978666306, + 0.0119113028049469, + 0.052779003977775574, + 0.0224620271474123, + 0.020029030740261078, + 0.019321756437420845, + 0.019839081913232803, + 0.023454749956727028, + 0.014227034524083138, + 0.024063263088464737, + 0.006510264705866575, + 0.0048347958363592625, + 0.003991237375885248, + 0.00945291481912136, + 0.004557725973427296, + 0.004862375557422638, + 0.0054461355321109295, + 0.004498402588069439, + 0.016741247847676277, + 0.009978970512747765, + 0.012726756744086742, + 0.011045639403164387, + 0.011717509478330612, + 0.009038109332323074, + 0.008743989281356335, + 0.008796310052275658, + 0.006232806947082281, + 0.005656477529555559, + 0.008398469537496567, + 0.008753602392971516, + 0.004809280391782522, + 0.00415730569511652, + 0.005218764767050743, + 0.008158428594470024, + 0.009674563072621822, + 0.010301210917532444, + 0.0060009765438735485, + 0.030708616599440575, + 0.0077886986546218395, + 0.008614090271294117, + 0.0321178175508976, + 0.0071428935043513775, + 0.012645483016967773, + 0.01346386969089508, + 0.00731975631788373, + 0.011104948818683624, + 0.014476793818175793, + 0.008622253313660622, + 0.01360202394425869, + 0.00914402399212122, + 0.015428782440721989, + 0.017870420590043068, + 0.013879098929464817, + 0.03357607498764992, + 0.013510656543076038, + 0.01183293852955103, + 0.020192595198750496, + 0.02098037861287594, + 0.008309652097523212, + 0.005448354408144951, + 0.005835372023284435, + 0.010798136703670025, + 0.00549673568457365, + 0.0053327446803450584, + 0.010431683622300625, + 0.006462675053626299, + 0.009665032848715782, + 0.008502507582306862, + 0.008136231452226639, + 0.013092189095914364, + 0.009174201637506485, + 0.009618349373340607, + 0.010779932141304016, + 0.011704107746481895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.11290796101093292 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot, rendered in a gritty, comic-book style, slowly advances across a desolate, sun-drenched desert highway, where every element is defined by bold black ink outlines and stark, high-contrast shadows. Under the intense, unforgiving midday sun, a visible heat haze shimmers, distorting the air and rising from the cracked asphalt. At the center of this oppressive scene, a flock of Coragyps atratus feeds en masse. They are a writhing mass of black feathers, huddled over invisible carrion. Their featherless, swarthy heads are deeply wrinkled and leathery, and they viciously tear and pull with their sharp, hooked beaks, bobbing and weaving. Their powerful bodies cast deep, sharp black shadows on the scorching pavement, creating stark, graphic patterns. The vultures aggressively jockey for position, their movements primal and focused, creating a brutal, raw atmosphere of natural decay across the empty, sun-bleached road.", + "seed": 35, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_11/index_2693_seed_35_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06962308324873448, + "vlm_logprob": -0.06962308324873448, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.06962308324873448, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010057131759822369 + ], + "opposite_logprob_tokens": [ + -5.135056972503662 + ], + "logprob_reward": -0.010057131759822369, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.010057131759822369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23461364209651947, + 0.10382954776287079, + 0.1381620168685913, + 0.08007103204727173, + 0.029766583815217018, + 0.09648871421813965, + 0.027194712311029434, + 0.043165747076272964, + 0.047923743724823, + 0.1987842321395874 + ], + "credit_map": [ + 0.053653035312891006, + 0.016649547964334488, + 0.036740127950906754, + 0.0223940908908844, + 0.036346182227134705, + 0.02785152569413185, + 0.023633725941181183, + 0.017345430329442024, + 0.02045980468392372, + 0.007046460174024105, + 0.006948260124772787, + 0.007751214783638716, + 0.008745180442929268, + 0.03710116818547249, + 0.006506228353828192, + 0.009271234273910522, + 0.015949809923768044, + 0.009855969808995724, + 0.0228809155523777, + 0.02010601945221424, + 0.017033683136105537, + 0.017617342993617058, + 0.015120030380785465, + 0.01959824375808239, + 0.014871427789330482, + 0.008940138854086399, + 0.006935620214790106, + 0.020448604598641396, + 0.00669579952955246, + 0.00545716006308794, + 0.010269098915159702, + 0.006453179754316807, + 0.0033095735125243664, + 0.00465863524004817, + 0.004352064803242683, + 0.0038446038961410522, + 0.003503787098452449, + 0.0034449808299541473, + 0.003669353201985359, + 0.002983587561175227, + 0.009997145272791386, + 0.009557424113154411, + 0.010485749691724777, + 0.010649827308952808, + 0.012128032743930817, + 0.017393678426742554, + 0.019382774829864502, + 0.0068940832279622555, + 0.004638538230210543, + 0.0020573472138494253, + 0.004056461621075869, + 0.0050501576624810696, + 0.0023690895177423954, + 0.003394131548702717, + 0.0027746614068746567, + 0.002854328602552414, + 0.005005188751965761, + 0.004843804985284805, + 0.006064840592443943, + 0.005357922054827213, + 0.006450043059885502, + 0.007635659538209438, + 0.003803311148658395, + 0.004004981834441423, + 0.006418965291231871, + 0.00504435645416379, + 0.006251962389796972, + 0.0043407040648162365, + 0.005039320793002844, + 0.005331113003194332, + 0.0036656216252595186, + 0.011831706389784813, + 0.019062083214521408, + 0.026125295087695122, + 0.02404436282813549, + 0.038045309484004974, + 0.02984808385372162, + 0.027977462857961655, + 0.015985755249857903, + 0.017695879563689232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.010057131759822369 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198473654687405 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198473654687405, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198473654687405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1462959200143814, + 0.1257026344537735, + 0.07002158463001251, + 0.17916566133499146, + 0.09019609540700912, + 0.029267767444252968, + 0.03378060832619667, + 0.021887052804231644, + 0.08880474418401718, + 0.2148779332637787 + ], + "credit_map": [ + 0.04079720377922058, + 0.013157257810235023, + 0.028578298166394234, + 0.014409800060093403, + 0.01194383017718792, + 0.016135524958372116, + 0.009979245252907276, + 0.011294756084680557, + 0.015397549606859684, + 0.007743882481008768, + 0.018511932343244553, + 0.034658629447221756, + 0.012775403447449207, + 0.0120188994333148, + 0.012434370815753937, + 0.012161963619291782, + 0.007369372062385082, + 0.006494040135294199, + 0.010378391481935978, + 0.01153014600276947, + 0.009828420355916023, + 0.00787836778908968, + 0.0077310651540756226, + 0.008811773732304573, + 0.02986103482544422, + 0.016718078404664993, + 0.020359106361865997, + 0.04078196361660957, + 0.022474100813269615, + 0.01606578752398491, + 0.019915657117962837, + 0.012989932671189308, + 0.009734800085425377, + 0.01052885316312313, + 0.015678349882364273, + 0.014870707876980305, + 0.011672680266201496, + 0.008441274985671043, + 0.012492034584283829, + 0.006777388043701649, + 0.0030991295352578163, + 0.0031030236277729273, + 0.0037134322337806225, + 0.0033848413731902838, + 0.00399794802069664, + 0.004499534610658884, + 0.004652670118957758, + 0.002817185828462243, + 0.003697473555803299, + 0.0032328260131180286, + 0.006691461894661188, + 0.004889138508588076, + 0.0026677516289055347, + 0.00533767556771636, + 0.00293094664812088, + 0.004333328455686569, + 0.0032266045454889536, + 0.0018263583770021796, + 0.0028795546386390924, + 0.001963538583368063, + 0.0025532403960824013, + 0.006419104989618063, + 0.0015016054967418313, + 0.0015170433325693011, + 0.008908024989068508, + 0.006918692961335182, + 0.006193290464580059, + 0.006702074781060219, + 0.019781844690442085, + 0.017360005527734756, + 0.006631936877965927, + 0.016308866441249847, + 0.0194470826536417, + 0.028827590867877007, + 0.02452549897134304, + 0.04041972756385803, + 0.03341830521821976, + 0.030164403840899467, + 0.018671318888664246, + 0.019404007121920586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.014198473654687405 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496445298194885 + ], + "opposite_logprob_tokens": [ + -2.929964542388916 + ], + "logprob_reward": -0.05496445298194885, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496445298194885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09280227869749069, + 0.345443993806839, + 0.13186173141002655, + 0.06663627177476883, + 0.053792256861925125, + 0.025547269731760025, + 0.08040481060743332, + 0.02585921436548233, + 0.038423456251621246, + 0.1392287313938141 + ], + "credit_map": [ + 0.012906121090054512, + 0.012149476446211338, + 0.014279191382229328, + 0.01954309083521366, + 0.007715815212577581, + 0.010725955478847027, + 0.0074113900773227215, + 0.008071236312389374, + 0.056570664048194885, + 0.026354560628533363, + 0.02387155219912529, + 0.03239172324538231, + 0.026506241410970688, + 0.15285374224185944, + 0.013438746333122253, + 0.013456746004521847, + 0.01540431845933199, + 0.009966530837118626, + 0.024873150512576103, + 0.016522690653800964, + 0.015636207535862923, + 0.025751587003469467, + 0.010732895694673061, + 0.012974345125257969, + 0.009963291697204113, + 0.011318049393594265, + 0.008989355526864529, + 0.007180082146078348, + 0.007065649144351482, + 0.010863769799470901, + 0.0049753496423363686, + 0.006280722562223673, + 0.0065358467400074005, + 0.010594451799988747, + 0.00796407088637352, + 0.006280013360083103, + 0.007610556203871965, + 0.004557387437671423, + 0.005127377808094025, + 0.005122554954141378, + 0.0033792576286941767, + 0.0016944995149970055, + 0.0032624166924506426, + 0.002774412976577878, + 0.0022550655994564295, + 0.004411197733134031, + 0.004843209404498339, + 0.0029272071551531553, + 0.014608437195420265, + 0.00591743690893054, + 0.005267612636089325, + 0.012582930736243725, + 0.006520110182464123, + 0.01742514595389366, + 0.009918199852108955, + 0.008164933882653713, + 0.0033842918928712606, + 0.0031930841505527496, + 0.0049617113545536995, + 0.0029570607002824545, + 0.00507369264960289, + 0.002201144117861986, + 0.0020119475666433573, + 0.0020762821659445763, + 0.0042900447733700275, + 0.00213025719858706, + 0.0029800436459481716, + 0.002921740058809519, + 0.0040105488151311874, + 0.01063083577901125, + 0.0029482056852430105, + 0.008511780761182308, + 0.014033745974302292, + 0.019429078325629234, + 0.015724925324320793, + 0.02173643372952938, + 0.021085023880004883, + 0.02325994148850441, + 0.011116252280771732, + 0.01284333411604166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.05496445298194885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798730194568634 + ], + "opposite_logprob_tokens": [ + -1.804872989654541 + ], + "logprob_reward": -0.1798730194568634, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798730194568634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2311764657497406, + 0.2430596649646759, + 0.06837457418441772, + 0.06395578384399414, + 0.06342360377311707, + 0.050011344254016876, + 0.032514479011297226, + 0.05678155645728111, + 0.05453367903828621, + 0.13616889715194702 + ], + "credit_map": [ + 0.03748558461666107, + 0.0286849532276392, + 0.028312386944890022, + 0.029863694682717323, + 0.022141678258776665, + 0.036736201494932175, + 0.018106089904904366, + 0.029845867305994034, + 0.043553344905376434, + 0.018126122653484344, + 0.022362804040312767, + 0.009388047270476818, + 0.020809616893529892, + 0.1094716489315033, + 0.008370555005967617, + 0.010977519676089287, + 0.007627792190760374, + 0.005844862665981054, + 0.008607993833720684, + 0.010583409108221531, + 0.010045533068478107, + 0.010556627996265888, + 0.00596731249243021, + 0.009141044691205025, + 0.005505284294486046, + 0.005810001865029335, + 0.005475768353790045, + 0.00990284327417612, + 0.009951630607247353, + 0.006967928260564804, + 0.011033717542886734, + 0.009308611042797565, + 0.0063622514717280865, + 0.005197550635784864, + 0.006602223962545395, + 0.00791594386100769, + 0.018072398379445076, + 0.00686436565592885, + 0.007697092369198799, + 0.004711780231446028, + 0.004840334877371788, + 0.005724045913666487, + 0.005619085393846035, + 0.0062357173301279545, + 0.007753418758511543, + 0.00672248937189579, + 0.008915917947888374, + 0.004200335126370192, + 0.0033709316048771143, + 0.002205485478043556, + 0.002615638542920351, + 0.005274482071399689, + 0.0022951907012611628, + 0.00489608570933342, + 0.008016007021069527, + 0.003840656718239188, + 0.007093858905136585, + 0.005341889802366495, + 0.005208723247051239, + 0.005480142310261726, + 0.008671863935887814, + 0.016419298946857452, + 0.0031329968478530645, + 0.005432785488665104, + 0.006048236973583698, + 0.004719670861959457, + 0.0049674613401293755, + 0.004897162318229675, + 0.004996791016310453, + 0.01681646890938282, + 0.004205887671560049, + 0.00788199994713068, + 0.012495252303779125, + 0.018992913886904716, + 0.014683840796351433, + 0.024604806676506996, + 0.02149350941181183, + 0.01970580965280533, + 0.011739839799702168, + 0.012452918104827404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.1798730194568634 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty comic-book style with bold black ink outlines and high-contrast shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a desolate, sun-drenched desert highway with cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a visible heat haze shimmering and distorting the air above the road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a flock of Coragyps atratus (vultures) visible feeding aggressively in the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward across the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902233839035034 + ], + "opposite_logprob_tokens": [ + -2.464022397994995 + ], + "logprob_reward": -0.08902233839035034, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902233839035034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1960321068763733, + 0.37003961205482483, + 0.040319930762052536, + 0.06644057482481003, + 0.07914327830076218, + 0.027229633182287216, + 0.025397662073373795, + 0.11064539849758148, + 0.02940421737730503, + 0.05534766986966133 + ], + "credit_map": [ + 0.029169877991080284, + 0.013446143828332424, + 0.035641979426145554, + 0.043113820254802704, + 0.014231802895665169, + 0.03148362785577774, + 0.015401226468384266, + 0.013543612323701382, + 0.053575437515974045, + 0.02729390747845173, + 0.05010130628943443, + 0.019799329340457916, + 0.031214334070682526, + 0.16347724199295044, + 0.011502951383590698, + 0.013075084425508976, + 0.00557318888604641, + 0.0031417831778526306, + 0.006636836566030979, + 0.004754012916237116, + 0.0035583062563091516, + 0.006718948483467102, + 0.0037474294658750296, + 0.006189420353621244, + 0.00763262715190649, + 0.008176764473319054, + 0.0063109323382377625, + 0.007746179588139057, + 0.010297255590558052, + 0.006047028582543135, + 0.01024401094764471, + 0.009985772892832756, + 0.00819415133446455, + 0.006817819084972143, + 0.008603638038039207, + 0.011447995901107788, + 0.014400048181414604, + 0.007954015396535397, + 0.012996736913919449, + 0.008728869259357452, + 0.0035668767523020506, + 0.004159777890890837, + 0.0025117991026490927, + 0.003027532948181033, + 0.0035654488019645214, + 0.004581364803016186, + 0.003078026697039604, + 0.0027388029266148806, + 0.0036566928029060364, + 0.002482707379385829, + 0.004039212595671415, + 0.003953302279114723, + 0.0030663369689136744, + 0.003039856441318989, + 0.0024866927415132523, + 0.002672858303412795, + 0.01265390869230032, + 0.010872945189476013, + 0.01305744331330061, + 0.010610371828079224, + 0.019594067707657814, + 0.027811234816908836, + 0.00653228722512722, + 0.009513131342828274, + 0.003366137621924281, + 0.0023568421602249146, + 0.003202321007847786, + 0.002233646111562848, + 0.0023899597581475973, + 0.008398661389946938, + 0.0022557333577424288, + 0.005200912710279226, + 0.004400097299367189, + 0.005980892572551966, + 0.007496132981032133, + 0.014313186518847942, + 0.00843519065529108, + 0.004547223448753357, + 0.005115848500281572, + 0.005059090442955494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.08902233839035034 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_12/index_tensor([6867, 6867])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_12/index_tensor([6867, 6867])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0e02eff2a3a0a3a0c83555a82f76eb3202991ab7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_12/index_tensor([6867, 6867])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up, framed with poetic precision in the signature Makoto Shinkai style, focuses on the stoic face of a majestic bronze guardian lion statue. Its weathered surface, a rich tapestry of dark, bronze, and pale green patina stripes, shimmers with the sudden, torrential downpour. Each individual raindrop is rendered with surreal, crystalline clarity, capturing the ethereal, warm glow of a nearby, out-of-frame lantern as they streak along the cold metal. A delicate cherry blossom petal, drenched in a translucent pink, soaked by the rain, clings tenaciously to the statue's cheek, creating a fragile contrast against the enduring metal. The glistening water meticulously highlights the intricate and deeply carved details of its awe-inspiring mane, droplets pooling and dripping from each sculpted curl. In the background, the world completely dissolves into a beautiful, dreamy bokeh of distant city lights, their vibrant gold, blue, and crimson orbs creating a soft, atmospheric blur that imbues the entire scene with a tranquil and melancholic beauty.", + "seed": 7131, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_12/index_6867_seed_7131_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07377983890473842, + "vlm_logprob": -0.07377983890473842, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.07377983890473842, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.29288989305496216 + ], + "opposite_logprob_tokens": [ + -1.4178898334503174 + ], + "logprob_reward": -0.29288989305496216, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.29288989305496216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15254327654838562, + 0.1479949951171875, + 0.09863242506980896, + 0.10120269656181335, + 0.08475197106599808, + 0.025266772136092186, + 0.05627836659550667, + 0.13251984119415283, + 0.09571240842342377, + 0.10509718209505081 + ], + "credit_map": [ + 0.020868731662631035, + 0.02117554284632206, + 0.020972885191440582, + 0.021335167810320854, + 0.017332864925265312, + 0.021082155406475067, + 0.017766457051038742, + 0.012009487487375736, + 0.016108952462673187, + 0.012956428341567516, + 0.03749769181013107, + 0.012806173413991928, + 0.01574089378118515, + 0.0117863230407238, + 0.032164644449949265, + 0.008933912962675095, + 0.0098151545971632, + 0.007952511310577393, + 0.022943636402487755, + 0.010343953035771847, + 0.00817523617297411, + 0.013603153638541698, + 0.019359545782208443, + 0.006439248565584421, + 0.010002505034208298, + 0.007111451588571072, + 0.014560901559889317, + 0.01973005384206772, + 0.010225655511021614, + 0.016716457903385162, + 0.01155062299221754, + 0.011305064894258976, + 0.009564056992530823, + 0.012314810417592525, + 0.012092355638742447, + 0.014559376053512096, + 0.00803794153034687, + 0.010560930706560612, + 0.009786675684154034, + 0.007835831493139267, + 0.0019948005210608244, + 0.0030106266494840384, + 0.003393439343199134, + 0.003918759059160948, + 0.004194543696939945, + 0.0034889867529273033, + 0.003365449607372284, + 0.001900169998407364, + 0.0050874375738203526, + 0.006049188319593668, + 0.005678731948137283, + 0.010118740610778332, + 0.0068793827667832375, + 0.01083875261247158, + 0.005109783262014389, + 0.006516352761536837, + 0.011634758673608303, + 0.01955159567296505, + 0.014294091612100601, + 0.017477471381425858, + 0.020060202106833458, + 0.018204335123300552, + 0.018561039119958878, + 0.012736358679831028, + 0.008222379721701145, + 0.01356435101479292, + 0.008656919933855534, + 0.017473356798291206, + 0.01224500872194767, + 0.01374537032097578, + 0.01167968101799488, + 0.010125351138412952, + 0.00760277034714818, + 0.018900061026215553, + 0.010787892155349255, + 0.016848508268594742, + 0.013107148930430412, + 0.010946286842226982, + 0.013577085919678211, + 0.01332743838429451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.29288989305496216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014238902367651463 + ], + "opposite_logprob_tokens": [ + -4.2642388343811035 + ], + "logprob_reward": -0.014238902367651463, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014238902367651463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1472756266593933, + 0.26192957162857056, + 0.030305787920951843, + 0.12326508015394211, + 0.05333912372589111, + 0.023333100602030754, + 0.09095774590969086, + 0.1878565102815628, + 0.03599613159894943, + 0.04574128985404968 + ], + "credit_map": [ + 0.021109970286488533, + 0.019660908728837967, + 0.016772888600826263, + 0.017797356471419334, + 0.01859317719936371, + 0.02550368383526802, + 0.016786381602287292, + 0.011051279492676258, + 0.02662164717912674, + 0.02682761661708355, + 0.0599503293633461, + 0.024307847023010254, + 0.03213617578148842, + 0.02396531216800213, + 0.05177386477589607, + 0.016346797347068787, + 0.002762684365734458, + 0.0025298588443547487, + 0.006488041020929813, + 0.005595730151981115, + 0.002918361686170101, + 0.0022343506570905447, + 0.005581252742558718, + 0.002195509849116206, + 0.013416528701782227, + 0.007798902690410614, + 0.01795078068971634, + 0.011406149715185165, + 0.011852533556520939, + 0.02652381919324398, + 0.01644967310130596, + 0.017866693437099457, + 0.004038305953145027, + 0.006726581137627363, + 0.0055200085043907166, + 0.006672574207186699, + 0.005994726438075304, + 0.008930264040827751, + 0.005450672935694456, + 0.010005995631217957, + 0.0020563150756061077, + 0.002999919466674328, + 0.0027995281852781773, + 0.0032419331837445498, + 0.003102228045463562, + 0.003963189199566841, + 0.0025229977909475565, + 0.0026469924487173557, + 0.00926396157592535, + 0.009419881738722324, + 0.010245258919894695, + 0.015534423291683197, + 0.012075730599462986, + 0.01683487743139267, + 0.007786949165165424, + 0.009796665981411934, + 0.01568126119673252, + 0.022372927516698837, + 0.02064845710992813, + 0.026476183906197548, + 0.026906171813607216, + 0.03363065421581268, + 0.022697122767567635, + 0.019443737342953682, + 0.004665515385568142, + 0.004477004054933786, + 0.003570730332285166, + 0.006880529224872589, + 0.004659358412027359, + 0.004468108993023634, + 0.003832470625638962, + 0.0034424164332449436, + 0.0036770550068467855, + 0.007080016192048788, + 0.005050069186836481, + 0.006502766627818346, + 0.007417873479425907, + 0.004112978465855122, + 0.00620014825835824, + 0.005700387060642242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.014238902367651463 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233067087829113 + ], + "opposite_logprob_tokens": [ + -3.7733066082000732 + ], + "logprob_reward": -0.0233067087829113, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233067087829113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.169498011469841, + 0.24646174907684326, + 0.1243654191493988, + 0.03650059178471565, + 0.031200282275676727, + 0.010380088351666927, + 0.06240440532565117, + 0.13630077242851257, + 0.12764526903629303, + 0.05524345487356186 + ], + "credit_map": [ + 0.01733742654323578, + 0.017201319336891174, + 0.027364859357476234, + 0.040651846677064896, + 0.019303858280181885, + 0.014581015333533287, + 0.021676139906048775, + 0.01138155348598957, + 0.02167757973074913, + 0.02457987330853939, + 0.051322899758815765, + 0.042526550590991974, + 0.029731936752796173, + 0.01810348965227604, + 0.044996533542871475, + 0.013522887602448463, + 0.010707386769354343, + 0.016779808327555656, + 0.026275932788848877, + 0.01789824105799198, + 0.01314740814268589, + 0.012282433919608593, + 0.019979391247034073, + 0.007294821552932262, + 0.0035120202228426933, + 0.0038597472012043, + 0.0039637526497244835, + 0.002206442877650261, + 0.004458374343812466, + 0.008530504070222378, + 0.004307992290705442, + 0.005661758594214916, + 0.0029666598420590162, + 0.003387161297723651, + 0.0039780521765351295, + 0.005537380464375019, + 0.004275941289961338, + 0.004445668775588274, + 0.0038974841590970755, + 0.0027119361329823732, + 0.0010320325382053852, + 0.0014606114709749818, + 0.001769459806382656, + 0.0011017281794920564, + 0.0013703186996281147, + 0.001436732243746519, + 0.0010921714128926396, + 0.0011170341167598963, + 0.005073252134025097, + 0.006935811601579189, + 0.007171991281211376, + 0.007361798547208309, + 0.008543620817363262, + 0.014666962437331676, + 0.006679047830402851, + 0.005971923004835844, + 0.01096359547227621, + 0.011032147333025932, + 0.02368275076150894, + 0.013543561100959778, + 0.016867144033312798, + 0.03765198215842247, + 0.010584548115730286, + 0.011975038796663284, + 0.00783886481076479, + 0.015613779425621033, + 0.009399482980370522, + 0.029815148562192917, + 0.016862275078892708, + 0.021265607327222824, + 0.01579045131802559, + 0.011059661395847797, + 0.0031744681764394045, + 0.010398684069514275, + 0.005617688409984112, + 0.008650850504636765, + 0.007548753637820482, + 0.005057165864855051, + 0.007906341925263405, + 0.006889500189572573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0233067087829113 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863910373300314 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.00863910373300314, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863910373300314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16552208364009857, + 0.10937629640102386, + 0.03818753734230995, + 0.06458885967731476, + 0.030692650005221367, + 0.04569052904844284, + 0.19305704534053802, + 0.23028363287448883, + 0.07022325694561005, + 0.052378080785274506 + ], + "credit_map": [ + 0.022506611421704292, + 0.020441345870494843, + 0.02814333140850067, + 0.02996794879436493, + 0.019104663282632828, + 0.013862503692507744, + 0.020410330966114998, + 0.011085345409810543, + 0.013642408885061741, + 0.01463943999260664, + 0.014622348360717297, + 0.013476107269525528, + 0.01661967858672142, + 0.012956189922988415, + 0.015118452720344067, + 0.008301659487187862, + 0.0039035717491060495, + 0.0042082988657057285, + 0.006016481667757034, + 0.005413119215518236, + 0.003966508898884058, + 0.0063125863671302795, + 0.005597271025180817, + 0.002769698388874531, + 0.005694541614502668, + 0.007922002114355564, + 0.006549446377903223, + 0.004066632129251957, + 0.008592198602855206, + 0.015621118247509003, + 0.007740105036646128, + 0.008402815088629723, + 0.003018513787537813, + 0.004443969111889601, + 0.004386669956147671, + 0.002277290215715766, + 0.0035669340286403894, + 0.004763221368193626, + 0.0032890106085687876, + 0.0049470411613583565, + 0.0032733185216784477, + 0.0073419720865786076, + 0.003793747164309025, + 0.008589896373450756, + 0.007188407704234123, + 0.006053284276276827, + 0.006063586100935936, + 0.0033863128628581762, + 0.017959635704755783, + 0.018497629091143608, + 0.023388363420963287, + 0.028198186308145523, + 0.025122327730059624, + 0.040801040828228, + 0.015894010663032532, + 0.023195860907435417, + 0.01167308073490858, + 0.01821998693048954, + 0.027206502854824066, + 0.020655246451497078, + 0.026517927646636963, + 0.07894472032785416, + 0.025481559336185455, + 0.02158460021018982, + 0.007568133529275656, + 0.00697542866691947, + 0.013148675672709942, + 0.014739691279828548, + 0.008856517262756824, + 0.005690986756235361, + 0.007350422441959381, + 0.005893393885344267, + 0.004429765976965427, + 0.007865550927817822, + 0.004906552843749523, + 0.006644533481448889, + 0.008604341186583042, + 0.006097747478634119, + 0.006793787237256765, + 0.007035799324512482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.00863910373300314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982458658516407 + ], + "opposite_logprob_tokens": [ + -3.5298244953155518 + ], + "logprob_reward": -0.02982458658516407, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982458658516407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1599803864955902, + 0.1590842306613922, + 0.060638535767793655, + 0.05516258254647255, + 0.034705016762018204, + 0.12869365513324738, + 0.2237292230129242, + 0.03209372237324715, + 0.0845712348818779, + 0.061341430991888046 + ], + "credit_map": [ + 0.023901404812932014, + 0.020227398723363876, + 0.026962563395500183, + 0.02013472281396389, + 0.01778011955320835, + 0.021308546885848045, + 0.018039219081401825, + 0.01162640005350113, + 0.016536347568035126, + 0.013445285148918629, + 0.042272794991731644, + 0.016610100865364075, + 0.0121614970266819, + 0.01051961351186037, + 0.03804260119795799, + 0.009495980106294155, + 0.005906571634113789, + 0.009030989371240139, + 0.009360631927847862, + 0.008433373644948006, + 0.005757132079452276, + 0.009485714137554169, + 0.008067932911217213, + 0.004596186336129904, + 0.006867334246635437, + 0.004825812764465809, + 0.008931069634854794, + 0.007470386568456888, + 0.00556615786626935, + 0.008135895244777203, + 0.0065711382776498795, + 0.006794783752411604, + 0.0035267837811261415, + 0.003030139720067382, + 0.0067711444571614265, + 0.0034109842963516712, + 0.0037012493703514338, + 0.003381081623956561, + 0.003904331009835005, + 0.006979299243539572, + 0.009075241163372993, + 0.01219121739268303, + 0.01733323186635971, + 0.023440107703208923, + 0.025070231407880783, + 0.015519790351390839, + 0.015980727970600128, + 0.010083114728331566, + 0.012595907784998417, + 0.023884454742074013, + 0.024177048355340958, + 0.030103182420134544, + 0.031297504901885986, + 0.050818733870983124, + 0.028239669278264046, + 0.022612711414694786, + 0.0032120198011398315, + 0.003959334921091795, + 0.006698985584080219, + 0.003627256490290165, + 0.003599132178351283, + 0.002732189605012536, + 0.003810746595263481, + 0.004454057198017836, + 0.0057076881639659405, + 0.011741473339498043, + 0.006474644877016544, + 0.01747909188270569, + 0.010792405344545841, + 0.015429534949362278, + 0.009563799947500229, + 0.007382599171251059, + 0.004562470130622387, + 0.009964151307940483, + 0.005448964424431324, + 0.00679011270403862, + 0.008115723729133606, + 0.007078896276652813, + 0.012249679304659367, + 0.0071314298547804356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.02982458658516407 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up, framed with poetic precision in the signature Makoto Shinkai style, focuses on the stoic face of a majestic bronze guardian lion statue. Its weathered surface, a rich tapestry of dark, bronze, and pale green patina stripes, shimmers with the sudden, torrential downpour. Each individual raindrop is rendered with surreal, crystalline clarity, capturing the ethereal, warm glow of a nearby, out-of-frame lantern as they streak along the cold metal. A delicate cherry blossom petal, drenched in a translucent pink, soaked by the rain, clings tenaciously to the statue's cheek, creating a fragile contrast against the enduring metal. The glistening water meticulously highlights the intricate and deeply carved details of its awe-inspiring mane, droplets pooling and dripping from each sculpted curl. In the background, the world completely dissolves into a beautiful, dreamy bokeh of distant city lights, their vibrant gold, blue, and crimson orbs creating a soft, atmospheric blur that imbues the entire scene with a tranquil and melancholic beauty.", + "seed": 7132, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_12/index_6867_seed_7132_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04646711922250688, + "vlm_logprob": -0.04646711922250688, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.04646711922250688, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01947297528386116 + ], + "opposite_logprob_tokens": [ + -4.269473075866699 + ], + "logprob_reward": -0.01947297528386116, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01947297528386116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05617928504943848, + 0.11427658796310425, + 0.0807880088686943, + 0.22230632603168488, + 0.07298067212104797, + 0.03262243792414665, + 0.12976354360580444, + 0.1584007441997528, + 0.10193407535552979, + 0.030748341232538223 + ], + "credit_map": [ + 0.007323719095438719, + 0.004533660598099232, + 0.005576398689299822, + 0.005687317345291376, + 0.00439433241263032, + 0.00628963066264987, + 0.00910766702145338, + 0.013266563415527344, + 0.011585639789700508, + 0.010874846018850803, + 0.022728191688656807, + 0.011474357917904854, + 0.009871465153992176, + 0.014674009755253792, + 0.017332257702946663, + 0.015735816210508347, + 0.009982271119952202, + 0.011331509798765182, + 0.010869583114981651, + 0.008670923300087452, + 0.010884983465075493, + 0.010170651599764824, + 0.007926170714199543, + 0.010951917618513107, + 0.021571574732661247, + 0.035335034132003784, + 0.03911949321627617, + 0.03228545933961868, + 0.020844070240855217, + 0.013821995817124844, + 0.02666289173066616, + 0.032665785402059555, + 0.009186751209199429, + 0.003507802030071616, + 0.023767108097672462, + 0.005518496036529541, + 0.003891497617587447, + 0.004341805819422007, + 0.01716541312634945, + 0.005601796787232161, + 0.0033494906965643167, + 0.005402827635407448, + 0.006405978463590145, + 0.0037250632885843515, + 0.003859812393784523, + 0.002438194118440151, + 0.0035625214222818613, + 0.0038785499054938555, + 0.017550982534885406, + 0.008847660385072231, + 0.024095337837934494, + 0.01825271174311638, + 0.018120666965842247, + 0.010635089129209518, + 0.020661087706685066, + 0.011600010097026825, + 0.01855863817036152, + 0.006884884089231491, + 0.056490905582904816, + 0.01275235041975975, + 0.010361921973526478, + 0.010675023309886456, + 0.015935096889734268, + 0.026741933077573776, + 0.01325499452650547, + 0.00868220068514347, + 0.02323048748075962, + 0.017654793336987495, + 0.007364132907241583, + 0.007272632792592049, + 0.009057307615876198, + 0.015417532995343208, + 0.005060675088316202, + 0.0019990724977105856, + 0.006503010168671608, + 0.004493270069360733, + 0.003610983258113265, + 0.002478956477716565, + 0.0037453724071383476, + 0.0028570001013576984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.01947297528386116 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255783811211586 + ], + "opposite_logprob_tokens": [ + -4.3875579833984375 + ], + "logprob_reward": -0.01255783811211586, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255783811211586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05985921248793602, + 0.016941621899604797, + 0.017180414870381355, + 0.03900749236345291, + 0.7478501200675964, + 0.010662876069545746, + 0.05070963874459267, + 0.023485006764531136, + 0.01439574547111988, + 0.019907834008336067 + ], + "credit_map": [ + 0.0068294634111225605, + 0.006591966841369867, + 0.007325300946831703, + 0.007149390410631895, + 0.005801815073937178, + 0.006607734598219395, + 0.008316718973219395, + 0.011236830614507198, + 0.0017683202167972922, + 0.001160099753178656, + 0.0036883139982819557, + 0.001787486020475626, + 0.0015236751642078161, + 0.0016188165172934532, + 0.002924684202298522, + 0.0024702269583940506, + 0.0021813195198774338, + 0.0015962886391207576, + 0.002615244360640645, + 0.0025063345674425364, + 0.0014197343261912465, + 0.0025276453234255314, + 0.0019852041732519865, + 0.002348644658923149, + 0.0031029963865876198, + 0.0028506782837212086, + 0.01082666590809822, + 0.0037993015721440315, + 0.004054106771945953, + 0.003963159862905741, + 0.006306871306151152, + 0.004103717394173145, + 0.07222811877727509, + 0.02501029148697853, + 0.27544111013412476, + 0.05338561162352562, + 0.04441852122545242, + 0.045697372406721115, + 0.19667936861515045, + 0.03498981520533562, + 0.0013456655433401465, + 0.001263043493963778, + 0.0018717654747888446, + 0.0008032379555515945, + 0.0018496512202546, + 0.0011771164136007428, + 0.0012160997139289975, + 0.0011362972436472774, + 0.005849230568856001, + 0.00462229922413826, + 0.007385329809039831, + 0.004924733191728592, + 0.007615968119353056, + 0.004876913037151098, + 0.008229590952396393, + 0.007205581292510033, + 0.001984917325899005, + 0.001541816396638751, + 0.0029284872580319643, + 0.0019682932179421186, + 0.003024803241714835, + 0.003987970761954784, + 0.0027963235042989254, + 0.005252396687865257, + 0.0018516938434913754, + 0.0011482371482998133, + 0.0025513432919979095, + 0.0017652634996920824, + 0.0020250098314136267, + 0.0012965596979483962, + 0.0015164739452302456, + 0.002241165144369006, + 0.004807138349860907, + 0.002658221172168851, + 0.0033800569362938404, + 0.0022574409376829863, + 0.0014291707193478942, + 0.0012730375165119767, + 0.001567761180922389, + 0.002535009989514947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.01255783811211586 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002992928028107 + ], + "opposite_logprob_tokens": [ + -2.6950299739837646 + ], + "logprob_reward": -0.07002992928028107, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002992928028107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09360431879758835, + 0.04284507408738136, + 0.03032015636563301, + 0.07740459591150284, + 0.605668306350708, + 0.013397208414971828, + 0.048022542148828506, + 0.06237315759062767, + 0.012319589033722878, + 0.014045069925487041 + ], + "credit_map": [ + 0.009128834120929241, + 0.010159442201256752, + 0.01928042620420456, + 0.016877006739377975, + 0.006481108255684376, + 0.009474987164139748, + 0.008718699216842651, + 0.013483826071023941, + 0.004206347279250622, + 0.004064865875989199, + 0.006382015533745289, + 0.0036937021650373936, + 0.003631695406511426, + 0.004063118249177933, + 0.009023392572999, + 0.007779936771839857, + 0.0025576737243682146, + 0.00428477069362998, + 0.004354390315711498, + 0.004657307639718056, + 0.0031840838491916656, + 0.004246947821229696, + 0.0027221699710935354, + 0.004312809091061354, + 0.00967284757643938, + 0.004753299988806248, + 0.023997385054826736, + 0.007688490673899651, + 0.008479621261358261, + 0.005941660143435001, + 0.010412522591650486, + 0.006458769552409649, + 0.05676410347223282, + 0.02178327552974224, + 0.21680128574371338, + 0.04195939749479294, + 0.036767397075891495, + 0.03881824016571045, + 0.1595250815153122, + 0.03324951231479645, + 0.001603331882506609, + 0.002411983208730817, + 0.0017525136936455965, + 0.0015937767457216978, + 0.001980322180315852, + 0.0012578016612678766, + 0.0014846831327304244, + 0.0013127959100529552, + 0.008534547872841358, + 0.0034850402735173702, + 0.010863826610147953, + 0.004409140907227993, + 0.006983616854995489, + 0.00359159242361784, + 0.00600311066955328, + 0.004151673521846533, + 0.007726198527961969, + 0.003956308588385582, + 0.02151484042406082, + 0.004054198041558266, + 0.003366200951859355, + 0.00465968856588006, + 0.0056744650937616825, + 0.011421255767345428, + 0.001413831952959299, + 0.001032570144161582, + 0.0026947231963276863, + 0.0018615658627822995, + 0.0014863727847114205, + 0.001081143389455974, + 0.001061013201251626, + 0.0016883669886738062, + 0.0017489707097411156, + 0.0019204366253688931, + 0.0022601985838264227, + 0.0019684589933604, + 0.0017380312783643603, + 0.0013788905926048756, + 0.0015478041023015976, + 0.0014822783414274454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.07002992928028107 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031935188453644514 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.0031935188453644514, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031935188453644514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12001390010118484, + 0.06573620438575745, + 0.09613882005214691, + 0.04558844864368439, + 0.07802391052246094, + 0.046080417931079865, + 0.14024797081947327, + 0.1963292360305786, + 0.026700226590037346, + 0.18514086306095123 + ], + "credit_map": [ + 0.012689664959907532, + 0.012496612966060638, + 0.02122589386999607, + 0.019016660749912262, + 0.009395388886332512, + 0.01071125641465187, + 0.014887655153870583, + 0.019590767100453377, + 0.007192058954387903, + 0.0063684433698654175, + 0.012175288051366806, + 0.005436412990093231, + 0.009152363054454327, + 0.005539061035960913, + 0.009170965291559696, + 0.010701616294682026, + 0.011118361726403236, + 0.012430155649781227, + 0.013747398741543293, + 0.01600736752152443, + 0.012079210951924324, + 0.010992784053087234, + 0.008716895245015621, + 0.011046646162867546, + 0.005184778477996588, + 0.004076274577528238, + 0.010325874201953411, + 0.005473599303513765, + 0.004787793383002281, + 0.0041756886057555676, + 0.006790871731936932, + 0.004773570690304041, + 0.01037884782999754, + 0.006297145504504442, + 0.017743205651640892, + 0.006233033258467913, + 0.0055152652785182, + 0.00684924004599452, + 0.014540635049343109, + 0.010466543026268482, + 0.005283615086227655, + 0.005642842035740614, + 0.010024183429777622, + 0.004272524733096361, + 0.006922711152583361, + 0.0036438300739973783, + 0.006238238885998726, + 0.00405247462913394, + 0.02940230257809162, + 0.010621262714266777, + 0.03672376647591591, + 0.01570620946586132, + 0.00934246089309454, + 0.007821690291166306, + 0.019468024373054504, + 0.011162251234054565, + 0.024202758446335793, + 0.013026349246501923, + 0.0663771778345108, + 0.013326924294233322, + 0.010664607398211956, + 0.01583009399473667, + 0.017279479652643204, + 0.03562183678150177, + 0.003664890071377158, + 0.0021888858173042536, + 0.005592527333647013, + 0.0032742831390351057, + 0.0038149775937199593, + 0.002202951116487384, + 0.0025837766006588936, + 0.0033779339864850044, + 0.03776174783706665, + 0.025617992505431175, + 0.04268227890133858, + 0.026561520993709564, + 0.011714579537510872, + 0.0083253663033247, + 0.014455313794314861, + 0.01802205853164196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.0031935188453644514 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708133459091187 + ], + "opposite_logprob_tokens": [ + -2.1270813941955566 + ], + "logprob_reward": -0.12708133459091187, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12708133459091187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20753945410251617, + 0.0686224102973938, + 0.08505668491125107, + 0.09279372543096542, + 0.16136790812015533, + 0.0366513766348362, + 0.17948803305625916, + 0.06658925861120224, + 0.036404792219400406, + 0.0654863715171814 + ], + "credit_map": [ + 0.018558893352746964, + 0.02578582428395748, + 0.05787511542439461, + 0.03820156306028366, + 0.011140204034745693, + 0.013797929510474205, + 0.019080914556980133, + 0.023098979145288467, + 0.006385487038642168, + 0.004315192345529795, + 0.010736731812357903, + 0.005463796202093363, + 0.005769619718194008, + 0.009121088311076164, + 0.011668169870972633, + 0.015162315219640732, + 0.012571807950735092, + 0.010431493632495403, + 0.01512286625802517, + 0.008523098193109035, + 0.009730069898068905, + 0.012572061270475388, + 0.007313795853406191, + 0.008791483007371426, + 0.008824992924928665, + 0.013583182357251644, + 0.015838034451007843, + 0.01358591578900814, + 0.011599133722484112, + 0.007346301805227995, + 0.011586247012019157, + 0.010429909452795982, + 0.016770927235484123, + 0.009836355224251747, + 0.048464126884937286, + 0.013439270667731762, + 0.01306730043143034, + 0.011800909414887428, + 0.031716279685497284, + 0.01627271994948387, + 0.004385712556540966, + 0.005189580377191305, + 0.004994647577404976, + 0.0038124308921396732, + 0.005628856364637613, + 0.004014451056718826, + 0.004555480554699898, + 0.004070214461535215, + 0.036417506635189056, + 0.012847860343754292, + 0.039892952889204025, + 0.014944816008210182, + 0.0171929020434618, + 0.017406199127435684, + 0.02843167819082737, + 0.012354107573628426, + 0.008945031091570854, + 0.005543006118386984, + 0.014397780410945415, + 0.004138284362852573, + 0.009534148499369621, + 0.006554818712174892, + 0.007028869818896055, + 0.010447314009070396, + 0.003860970726236701, + 0.0029803819488734007, + 0.006966355722397566, + 0.005231587216258049, + 0.005346513353288174, + 0.0041768718510866165, + 0.0032868122216314077, + 0.004555295221507549, + 0.01538904756307602, + 0.0064837634563446045, + 0.013965985737740993, + 0.008702602237462997, + 0.005226695444434881, + 0.004044354427605867, + 0.005872828885912895, + 0.005801084451377392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.12708133459091187 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up, framed with poetic precision in the signature Makoto Shinkai style, focuses on the stoic face of a majestic bronze guardian lion statue. Its weathered surface, a rich tapestry of dark, bronze, and pale green patina stripes, shimmers with the sudden, torrential downpour. Each individual raindrop is rendered with surreal, crystalline clarity, capturing the ethereal, warm glow of a nearby, out-of-frame lantern as they streak along the cold metal. A delicate cherry blossom petal, drenched in a translucent pink, soaked by the rain, clings tenaciously to the statue's cheek, creating a fragile contrast against the enduring metal. The glistening water meticulously highlights the intricate and deeply carved details of its awe-inspiring mane, droplets pooling and dripping from each sculpted curl. In the background, the world completely dissolves into a beautiful, dreamy bokeh of distant city lights, their vibrant gold, blue, and crimson orbs creating a soft, atmospheric blur that imbues the entire scene with a tranquil and melancholic beauty.", + "seed": 7133, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_12/index_6867_seed_7133_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04027491612359881, + "vlm_logprob": -0.04027491612359881, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.04027491612359881, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1310061812400818 + ], + "opposite_logprob_tokens": [ + -2.1310062408447266 + ], + "logprob_reward": -0.1310061812400818, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1310061812400818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20441941916942596, + 0.1048450767993927, + 0.035070016980171204, + 0.1313972920179367, + 0.08012568205595016, + 0.02677374705672264, + 0.04174350947141647, + 0.17107553780078888, + 0.07825836539268494, + 0.12629130482673645 + ], + "credit_map": [ + 0.028409497812390327, + 0.017692264169454575, + 0.024440255016088486, + 0.017294373363256454, + 0.02311738207936287, + 0.03183041140437126, + 0.02727876976132393, + 0.03435647115111351, + 0.01804480515420437, + 0.011171995662152767, + 0.013358542695641518, + 0.0064879595302045345, + 0.00997111201286316, + 0.009783224202692509, + 0.025283116847276688, + 0.010744325816631317, + 0.005798247642815113, + 0.003109488869085908, + 0.007011425215750933, + 0.0027137473225593567, + 0.0034519450273364782, + 0.0026374259032309055, + 0.006543240509927273, + 0.0038045025430619717, + 0.012741602957248688, + 0.01581442914903164, + 0.01345075387507677, + 0.016406629234552383, + 0.014264634810388088, + 0.025122223421931267, + 0.01977367140352726, + 0.013823364861309528, + 0.011322284117341042, + 0.006691684015095234, + 0.009638353250920773, + 0.007705737370997667, + 0.01059777569025755, + 0.013426315039396286, + 0.010747267864644527, + 0.009996275417506695, + 0.0032146736048161983, + 0.0025007501244544983, + 0.004162643104791641, + 0.002167150378227234, + 0.002743449294939637, + 0.0043420041911304, + 0.0032469758298248053, + 0.004396102856844664, + 0.007379636634141207, + 0.003272517118602991, + 0.010568279772996902, + 0.0034109808038920164, + 0.0037575941532850266, + 0.0038436369504779577, + 0.0050298371352255344, + 0.004481031093746424, + 0.015606783330440521, + 0.022354083135724068, + 0.02630024589598179, + 0.014669153839349747, + 0.022082161158323288, + 0.02040715143084526, + 0.03236045315861702, + 0.01729552634060383, + 0.008674717508256435, + 0.012367220595479012, + 0.006559957284480333, + 0.0065160817466676235, + 0.01225377805531025, + 0.009939002804458141, + 0.014686274342238903, + 0.007261337712407112, + 0.013125840574502945, + 0.016386473551392555, + 0.01998245343565941, + 0.011450869962573051, + 0.020173652097582817, + 0.013586233370006084, + 0.020280731841921806, + 0.011305064894258976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.1310061812400818 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012555483728647232 + ], + "opposite_logprob_tokens": [ + -4.3875555992126465 + ], + "logprob_reward": -0.012555483728647232, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012555483728647232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1637910008430481, + 0.052666909992694855, + 0.07144284248352051, + 0.11827979236841202, + 0.08434934914112091, + 0.04481211304664612, + 0.02379235066473484, + 0.08056983351707458, + 0.1239917054772377, + 0.2363041341304779 + ], + "credit_map": [ + 0.0201485026627779, + 0.01563422568142414, + 0.013129940256476402, + 0.015586314722895622, + 0.01986675150692463, + 0.023524096235632896, + 0.023365488275885582, + 0.03253566846251488, + 0.007273869588971138, + 0.007039844989776611, + 0.006407373119145632, + 0.0032831900753080845, + 0.004969647619873285, + 0.005605957005172968, + 0.011666016653180122, + 0.006421002559363842, + 0.00959201529622078, + 0.005951593164354563, + 0.009247459471225739, + 0.009322445839643478, + 0.008556830696761608, + 0.009060313925147057, + 0.012652507051825523, + 0.007059665396809578, + 0.014475551433861256, + 0.008498696610331535, + 0.03122468665242195, + 0.0069048781879246235, + 0.011884338222444057, + 0.017504841089248657, + 0.01725117675960064, + 0.010535608045756817, + 0.00798878911882639, + 0.007555147632956505, + 0.008025704883038998, + 0.005487001501023769, + 0.010044175200164318, + 0.021663974970579147, + 0.012884256429970264, + 0.010700291022658348, + 0.004124993458390236, + 0.004213768523186445, + 0.005755379796028137, + 0.003588305553421378, + 0.006909536197781563, + 0.006624196190387011, + 0.006082614883780479, + 0.007513314951211214, + 0.0040046642534434795, + 0.002335066208615899, + 0.004179806914180517, + 0.002250055084004998, + 0.0029522175900638103, + 0.0026988994795829058, + 0.0028736572712659836, + 0.0024979840964078903, + 0.008816821500658989, + 0.009515664540231228, + 0.010669568553566933, + 0.007838912308216095, + 0.012480951845645905, + 0.010671048425137997, + 0.011600605212152004, + 0.008976255543529987, + 0.018491650000214577, + 0.015611167997121811, + 0.018098236992955208, + 0.009660241194069386, + 0.019212963059544563, + 0.013830902986228466, + 0.01709490828216076, + 0.011991622857749462, + 0.02457968145608902, + 0.02751348167657852, + 0.048251599073410034, + 0.019546350464224815, + 0.04344362020492554, + 0.022855371236801147, + 0.029994886368513107, + 0.020119115710258484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.012555483728647232 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192220479249954 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192220479249954, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192220479249954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07102774828672409, + 0.04016430303454399, + 0.061655886471271515, + 0.18228691816329956, + 0.3188900053501129, + 0.030973874032497406, + 0.03226746618747711, + 0.04520402476191521, + 0.12166158854961395, + 0.09586820751428604 + ], + "credit_map": [ + 0.009825430810451508, + 0.007129869889467955, + 0.010869134217500687, + 0.005802216473966837, + 0.007836234755814075, + 0.01118666585534811, + 0.009284461848437786, + 0.00909373164176941, + 0.004998791962862015, + 0.004663950763642788, + 0.004507392644882202, + 0.002925366861745715, + 0.004193645901978016, + 0.005639621522277594, + 0.00785259623080492, + 0.00538293644785881, + 0.008717266842722893, + 0.005207652226090431, + 0.009178796783089638, + 0.006985460873693228, + 0.005764818750321865, + 0.006783952470868826, + 0.012288462370634079, + 0.006729473359882832, + 0.01988978125154972, + 0.01660894975066185, + 0.04513910785317421, + 0.012365958653390408, + 0.016403447836637497, + 0.03152978792786598, + 0.02416566200554371, + 0.016184208914637566, + 0.03206821158528328, + 0.03429238125681877, + 0.04244661703705788, + 0.03286303952336311, + 0.03983832150697708, + 0.060965802520513535, + 0.044584721326828, + 0.03183092176914215, + 0.0030256768222898245, + 0.003132613841444254, + 0.005703683476895094, + 0.002035763580352068, + 0.004139437805861235, + 0.0038842554204165936, + 0.0038677004631608725, + 0.005184743087738752, + 0.0043712337501347065, + 0.003845247905701399, + 0.006005199160426855, + 0.0032408367842435837, + 0.003185979789122939, + 0.0039724563248455524, + 0.004119109828025103, + 0.003527404274791479, + 0.006384722422808409, + 0.004445452708750963, + 0.008858392015099525, + 0.0036921899300068617, + 0.006808621343225241, + 0.0055093588307499886, + 0.00513174943625927, + 0.004373539239168167, + 0.009963509626686573, + 0.013100886717438698, + 0.014887623488903046, + 0.008417615666985512, + 0.019542129710316658, + 0.013650220818817616, + 0.0340004600584507, + 0.008099139668047428, + 0.007114203181117773, + 0.011106958612799644, + 0.011598305776715279, + 0.007574466522783041, + 0.020848503336310387, + 0.009101170115172863, + 0.020814862102270126, + 0.0077097490429878235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.018192220479249954 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015135272406041622 + ], + "opposite_logprob_tokens": [ + -6.501513481140137 + ], + "logprob_reward": -0.0015135272406041622, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015135272406041622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16099563241004944, + 0.0671122819185257, + 0.09126023203134537, + 0.14663295447826385, + 0.1084907278418541, + 0.020509710535407066, + 0.043160367757081985, + 0.1313284933567047, + 0.08249211311340332, + 0.14801748096942902 + ], + "credit_map": [ + 0.01760431006550789, + 0.01756860502064228, + 0.026250533759593964, + 0.012692400254309177, + 0.01305859349668026, + 0.017658356577157974, + 0.030093466863036156, + 0.02606937289237976, + 0.012241199612617493, + 0.007759250700473785, + 0.009685675613582134, + 0.004931533709168434, + 0.0054444605484604836, + 0.006237488240003586, + 0.013859587721526623, + 0.006953094620257616, + 0.011979389935731888, + 0.007500446867197752, + 0.013308160938322544, + 0.012483210302889347, + 0.008513020351529121, + 0.010269220918416977, + 0.018289944157004356, + 0.008916850201785564, + 0.011029734276235104, + 0.012823452241718769, + 0.022944144904613495, + 0.011658615432679653, + 0.017997289076447487, + 0.029784241691231728, + 0.024133436381816864, + 0.016262060031294823, + 0.011011885479092598, + 0.008259005844593048, + 0.014714580029249191, + 0.005923911929130554, + 0.011802266351878643, + 0.02470385655760765, + 0.01715835928916931, + 0.01491687074303627, + 0.002619158010929823, + 0.0018018570262938738, + 0.004213432315737009, + 0.0020198263227939606, + 0.0020729596726596355, + 0.002431150758638978, + 0.0023874077014625072, + 0.002963920822367072, + 0.006118081510066986, + 0.003797651268541813, + 0.007320799864828587, + 0.004668237175792456, + 0.006821233779191971, + 0.005044907797127962, + 0.005844998639076948, + 0.0035444621462374926, + 0.012804561294615269, + 0.019674574956297874, + 0.02211361937224865, + 0.013051318936049938, + 0.014066014438867569, + 0.013450796715915203, + 0.022113477811217308, + 0.014054132625460625, + 0.011987128295004368, + 0.007811160292476416, + 0.014704236760735512, + 0.005715167615562677, + 0.010940974578261375, + 0.008082432672381401, + 0.016530046239495277, + 0.006720971316099167, + 0.008916234597563744, + 0.010128749534487724, + 0.012391465716063976, + 0.007010150235146284, + 0.03617308661341667, + 0.010201329365372658, + 0.05254238471388817, + 0.01065409742295742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.0015135272406041622 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038107167929410934 + ], + "opposite_logprob_tokens": [ + -3.28810715675354 + ], + "logprob_reward": -0.038107167929410934, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038107167929410934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12010367959737778, + 0.027035389095544815, + 0.08608023077249527, + 0.07346437871456146, + 0.04870106279850006, + 0.026528870686888695, + 0.025995716452598572, + 0.17841842770576477, + 0.25829124450683594, + 0.15538102388381958 + ], + "credit_map": [ + 0.011188017204403877, + 0.014229467138648033, + 0.015252253040671349, + 0.010707361623644829, + 0.01322690024971962, + 0.018026892095804214, + 0.019634781405329704, + 0.017837997525930405, + 0.0022371900267899036, + 0.0025585347320884466, + 0.0023924584966152906, + 0.003270757617428899, + 0.002544987015426159, + 0.006290140096098185, + 0.00438189972192049, + 0.003359417663887143, + 0.0076253218576312065, + 0.008422981947660446, + 0.011349466629326344, + 0.016515258699655533, + 0.009862689301371574, + 0.011205175891518593, + 0.011113503947854042, + 0.009985820390284061, + 0.007418413180857897, + 0.007074651774019003, + 0.00892617553472519, + 0.006836759392172098, + 0.009572015143930912, + 0.013146134093403816, + 0.01304546743631363, + 0.00744475144892931, + 0.006096275523304939, + 0.004923020955175161, + 0.007166136056184769, + 0.003006431506946683, + 0.005333721172064543, + 0.009867941029369831, + 0.006869293283671141, + 0.005438236519694328, + 0.00368450116366148, + 0.0024219576735049486, + 0.0038745214696973562, + 0.0019987565465271473, + 0.003541036508977413, + 0.0045119053684175014, + 0.0035873553715646267, + 0.0029088344890624285, + 0.0037100345361977816, + 0.002380358288064599, + 0.004219844006001949, + 0.0022197854705154896, + 0.003007086692377925, + 0.003751276293769479, + 0.003716777078807354, + 0.002990551060065627, + 0.02653403952717781, + 0.022233344614505768, + 0.040670037269592285, + 0.012572129257023335, + 0.015298373997211456, + 0.01570265367627144, + 0.029496807605028152, + 0.01591101475059986, + 0.030402425676584244, + 0.022303571924567223, + 0.03968349099159241, + 0.01678568683564663, + 0.03757300227880478, + 0.02622166834771633, + 0.065264992415905, + 0.020056376233696938, + 0.013605213724076748, + 0.015338812954723835, + 0.02153262309730053, + 0.010849593207240105, + 0.03272566571831703, + 0.01380882691591978, + 0.035342104732990265, + 0.012178170494735241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.038107167929410934 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up, framed with poetic precision in the signature Makoto Shinkai style, focuses on the stoic face of a majestic bronze guardian lion statue. Its weathered surface, a rich tapestry of dark, bronze, and pale green patina stripes, shimmers with the sudden, torrential downpour. Each individual raindrop is rendered with surreal, crystalline clarity, capturing the ethereal, warm glow of a nearby, out-of-frame lantern as they streak along the cold metal. A delicate cherry blossom petal, drenched in a translucent pink, soaked by the rain, clings tenaciously to the statue's cheek, creating a fragile contrast against the enduring metal. The glistening water meticulously highlights the intricate and deeply carved details of its awe-inspiring mane, droplets pooling and dripping from each sculpted curl. In the background, the world completely dissolves into a beautiful, dreamy bokeh of distant city lights, their vibrant gold, blue, and crimson orbs creating a soft, atmospheric blur that imbues the entire scene with a tranquil and melancholic beauty.", + "seed": 7134, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_12/index_6867_seed_7134_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02684613815508783, + "vlm_logprob": -0.02684613815508783, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.02684613815508783, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018305763602256775 + ], + "opposite_logprob_tokens": [ + -4.268305778503418 + ], + "logprob_reward": -0.018305763602256775, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018305763602256775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09923838078975677, + 0.04158227518200874, + 0.024947376921772957, + 0.026131801307201385, + 0.06833184510469437, + 0.03753755986690521, + 0.04515891894698143, + 0.09946171194314957, + 0.493008553981781, + 0.0646016001701355 + ], + "credit_map": [ + 0.009353662841022015, + 0.013124925084412098, + 0.018092820420861244, + 0.017267882823944092, + 0.010461351834237576, + 0.007743197493255138, + 0.00844739843159914, + 0.014747136272490025, + 0.002496512373909354, + 0.004198143258690834, + 0.008244454860687256, + 0.008126191794872284, + 0.0029807996470481157, + 0.00785051565617323, + 0.003385343821719289, + 0.004300304688513279, + 0.0016745033208280802, + 0.003921305760741234, + 0.002996119437739253, + 0.0038711640518158674, + 0.0032904779072850943, + 0.0030436699744313955, + 0.00310127274133265, + 0.003048860002309084, + 0.0024977701250463724, + 0.003775956341996789, + 0.0034005811903625727, + 0.004231847822666168, + 0.003094303421676159, + 0.0037192488089203835, + 0.0026818793267011642, + 0.0027302131056785583, + 0.0032801919151097536, + 0.007156618405133486, + 0.0031778861302882433, + 0.0179239921271801, + 0.0028429131489247084, + 0.008442102931439877, + 0.0027436157688498497, + 0.022764520719647408, + 0.0027324839029461145, + 0.00786407571285963, + 0.0031900866888463497, + 0.008042399771511555, + 0.0039905039593577385, + 0.0037093462888151407, + 0.004695302341133356, + 0.003313356777653098, + 0.008012458682060242, + 0.006292480044066906, + 0.007937978021800518, + 0.006028646603226662, + 0.002631341340020299, + 0.004573833663016558, + 0.0031768816988915205, + 0.006505296099931002, + 0.009147893637418747, + 0.018853921443223953, + 0.011565016582608223, + 0.01665208302438259, + 0.009941146709024906, + 0.010797185823321342, + 0.009095078334212303, + 0.01340937428176403, + 0.0379079133272171, + 0.09796027094125748, + 0.02170289121568203, + 0.05097438395023346, + 0.03374599292874336, + 0.06168612092733383, + 0.021945947781205177, + 0.16708502173423767, + 0.0051399776712059975, + 0.010540933348238468, + 0.004121524281799793, + 0.014999186620116234, + 0.003878652350977063, + 0.006492216605693102, + 0.00657874159514904, + 0.01285036001354456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.018305763602256775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009803954511880875 + ], + "opposite_logprob_tokens": [ + -4.634803771972656 + ], + "logprob_reward": -0.009803954511880875, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009803954511880875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017330322414636612, + 0.01873212866485119, + 0.0058299521915614605, + 0.009724672883749008, + 0.12055668234825134, + 0.01200645137578249, + 0.013369761407375336, + 0.3128591775894165, + 0.26996394991874695, + 0.21962690353393555 + ], + "credit_map": [ + 0.0017704588826745749, + 0.0014651098754256964, + 0.003143210429698229, + 0.0029941245447844267, + 0.001850504893809557, + 0.0016391956014558673, + 0.002087617991492152, + 0.002380100777372718, + 0.00090529874432832, + 0.0022221344988793135, + 0.002666764659807086, + 0.0030408853199332952, + 0.0020918450318276882, + 0.003089535515755415, + 0.0021986381616443396, + 0.002517026849091053, + 0.0005036190850660205, + 0.0009806921007111669, + 0.0009411692735739052, + 0.0009415383683517575, + 0.0009695470216684043, + 0.0003958978923037648, + 0.0005980934365652502, + 0.0004993948386982083, + 0.00106932467315346, + 0.0012924993643537164, + 0.0013205688446760178, + 0.0011980625567957759, + 0.0012761522084474564, + 0.001602850272320211, + 0.0007729391218163073, + 0.0011922750854864717, + 0.004335032310336828, + 0.010999868623912334, + 0.004018704406917095, + 0.028863390907645226, + 0.002753153210505843, + 0.020639048889279366, + 0.003945834934711456, + 0.04500165954232216, + 0.0023149526678025723, + 0.0016630428144708276, + 0.0015588700771331787, + 0.0016326244222000241, + 0.0007105957483872771, + 0.001331876264885068, + 0.0012686585541814566, + 0.0015258314087986946, + 0.002115355571731925, + 0.0014645315241068602, + 0.0021313910838216543, + 0.0015085096238180995, + 0.0005595572874881327, + 0.0017085218569263816, + 0.0006083500338718295, + 0.0032735434360802174, + 0.02938787080347538, + 0.04838299751281738, + 0.06258541345596313, + 0.04334547743201256, + 0.025694826617836952, + 0.026338815689086914, + 0.02545025944709778, + 0.05167350545525551, + 0.020423179492354393, + 0.05941295623779297, + 0.014148355461657047, + 0.02829815074801445, + 0.021681470796465874, + 0.029837483540177345, + 0.013183800503611565, + 0.08297854661941528, + 0.014786701649427414, + 0.055104371160268784, + 0.012975840829312801, + 0.06476886570453644, + 0.01704895682632923, + 0.014151318930089474, + 0.015825921669602394, + 0.024964936077594757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.009803954511880875 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048656027764081955 + ], + "opposite_logprob_tokens": [ + -3.0486559867858887 + ], + "logprob_reward": -0.048656027764081955, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048656027764081955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0813010111451149, + 0.050288181751966476, + 0.06906995177268982, + 0.026571664959192276, + 0.08428482711315155, + 0.027017369866371155, + 0.06623474508523941, + 0.18503400683403015, + 0.2255222201347351, + 0.184675931930542 + ], + "credit_map": [ + 0.007701120804995298, + 0.005797996185719967, + 0.014553320594131947, + 0.013471933081746101, + 0.00531277060508728, + 0.008287099190056324, + 0.005762308835983276, + 0.020414475351572037, + 0.0030250612180680037, + 0.00765535281971097, + 0.007425935473293066, + 0.009781156666576862, + 0.00849836878478527, + 0.004704945255070925, + 0.005821047816425562, + 0.003376319771632552, + 0.004192416090518236, + 0.012478283606469631, + 0.007035670801997185, + 0.012489699758589268, + 0.011584091931581497, + 0.005085362587124109, + 0.008867007680237293, + 0.007337427698075771, + 0.002880578162148595, + 0.004345089662820101, + 0.003670171368867159, + 0.0038203883450478315, + 0.0029682135209441185, + 0.003297369694337249, + 0.002586389658972621, + 0.00300346571020782, + 0.00331988581456244, + 0.006779944524168968, + 0.002684197388589382, + 0.011710318736732006, + 0.0021090712398290634, + 0.022191934287548065, + 0.002583587309345603, + 0.03290589526295662, + 0.003119453089311719, + 0.004474036395549774, + 0.0028312294743955135, + 0.005850262939929962, + 0.001933503313921392, + 0.0034984902013093233, + 0.0018412641948089004, + 0.0034691314212977886, + 0.008554162457585335, + 0.007111847400665283, + 0.0129171684384346, + 0.007656453177332878, + 0.003933703526854515, + 0.009510298259556293, + 0.004725857172161341, + 0.0118252569809556, + 0.016696708276867867, + 0.03126518055796623, + 0.03581030294299126, + 0.02713189087808132, + 0.015268201939761639, + 0.014598539099097252, + 0.014565880410373211, + 0.029697317630052567, + 0.01692512258887291, + 0.05612628161907196, + 0.016325319185853004, + 0.027520161122083664, + 0.019855700433254242, + 0.021202899515628815, + 0.011902360245585442, + 0.055664412677288055, + 0.011142373085021973, + 0.054646074771881104, + 0.010381041094660759, + 0.0636800080537796, + 0.01074926182627678, + 0.00913920160382986, + 0.01256913784891367, + 0.012368850409984589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.048656027764081955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002489208010956645 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.002489208010956645, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002489208010956645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11268865317106247, + 0.059693608433008194, + 0.057901013642549515, + 0.05028275027871132, + 0.16280394792556763, + 0.034088581800460815, + 0.040630973875522614, + 0.2759105861186981, + 0.14991292357444763, + 0.05608696490526199 + ], + "credit_map": [ + 0.01024653110653162, + 0.011686094105243683, + 0.018012642860412598, + 0.024821389466524124, + 0.010862830094993114, + 0.009375827386975288, + 0.010628299787640572, + 0.017055021598935127, + 0.004356267396360636, + 0.005395264830440283, + 0.009068180806934834, + 0.009151785634458065, + 0.006226792000234127, + 0.012518594041466713, + 0.004950052127242088, + 0.008026670664548874, + 0.003951716236770153, + 0.009738597087562084, + 0.006571946199983358, + 0.010516307316720486, + 0.009237491525709629, + 0.0040552495047450066, + 0.007818354293704033, + 0.006011348217725754, + 0.005135639570653439, + 0.006156143266707659, + 0.006677646655589342, + 0.006918346043676138, + 0.003978755325078964, + 0.009791979566216469, + 0.004325769376009703, + 0.007298470940440893, + 0.006432566326111555, + 0.01366172544658184, + 0.0055838157422840595, + 0.03276888653635979, + 0.0028218745719641447, + 0.034263718873262405, + 0.004955338314175606, + 0.06231600418686867, + 0.0029115523211658, + 0.005250035785138607, + 0.00373847340233624, + 0.006963865365833044, + 0.002587329363450408, + 0.005940266419202089, + 0.0030712909065186977, + 0.0036257661413401365, + 0.004458763636648655, + 0.0059051536954939365, + 0.006730096414685249, + 0.0051413713954389095, + 0.003254217328503728, + 0.004415892530232668, + 0.00283767725341022, + 0.007887798361480236, + 0.026262734085321426, + 0.043315719813108444, + 0.05400285869836807, + 0.038586802780628204, + 0.02300056256353855, + 0.02329561673104763, + 0.023074273020029068, + 0.044372040778398514, + 0.011454561725258827, + 0.040549591183662415, + 0.012110189534723759, + 0.0197211392223835, + 0.014733470976352692, + 0.012855801731348038, + 0.0077132852748036385, + 0.030774878337979317, + 0.004529729951173067, + 0.011399832554161549, + 0.004233112093061209, + 0.014398712664842606, + 0.005181421991437674, + 0.004571711178869009, + 0.005018709693104029, + 0.006753731053322554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.002489208010956645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static extreme close-up of a bronze guardian lion statue's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the statue's surface show a weathered texture with dark bronze and pale green patina stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a translucent pink cherry blossom petal visible clinging to the statue's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background rendered as a soft, out-of-focus bokeh of distant city lights in gold, blue, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall visual style consistent with the signature Makoto Shinkai aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054975736886262894 + ], + "opposite_logprob_tokens": [ + -2.929975748062134 + ], + "logprob_reward": -0.054975736886262894, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054975736886262894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08250526338815689, + 0.06114073470234871, + 0.0559815987944603, + 0.01977446861565113, + 0.2338697761297226, + 0.03585365414619446, + 0.028468262404203415, + 0.2262241542339325, + 0.20856158435344696, + 0.047620467841625214 + ], + "credit_map": [ + 0.008278978057205677, + 0.01075684279203415, + 0.014581902883946896, + 0.011930602602660656, + 0.010739554651081562, + 0.009941836819052696, + 0.007055450230836868, + 0.009220101870596409, + 0.0038212186191231012, + 0.010076561942696571, + 0.00788428820669651, + 0.010520742274820805, + 0.009897272102534771, + 0.004649665672332048, + 0.008158110082149506, + 0.006132875103503466, + 0.004578771535307169, + 0.009133458137512207, + 0.007116121705621481, + 0.010041269473731518, + 0.009171190671622753, + 0.004731286782771349, + 0.007031653542071581, + 0.004177843686193228, + 0.0016139065846800804, + 0.0026856812182813883, + 0.0014894065679982305, + 0.0029492047615349293, + 0.003181983483955264, + 0.0028972404543310404, + 0.002422118093818426, + 0.002534928498789668, + 0.009758028201758862, + 0.026917550712823868, + 0.008890611119568348, + 0.07573181390762329, + 0.006467235740274191, + 0.019386257976293564, + 0.009157185442745686, + 0.07756111025810242, + 0.0020851227454841137, + 0.005902237258851528, + 0.0032895454205572605, + 0.006869190372526646, + 0.0025693345814943314, + 0.005528209265321493, + 0.0025721159763634205, + 0.007037899922579527, + 0.004618409555405378, + 0.003165561007335782, + 0.00587611785158515, + 0.003512732684612274, + 0.002242867136374116, + 0.0029166541062295437, + 0.002411188092082739, + 0.003724733367562294, + 0.020997364073991776, + 0.03430332615971565, + 0.04753747954964638, + 0.030422095209360123, + 0.01891220360994339, + 0.01899508759379387, + 0.018276140093803406, + 0.036780472844839096, + 0.015483911149203777, + 0.04528485983610153, + 0.010386568494141102, + 0.02148015797138214, + 0.016873322427272797, + 0.024522950872778893, + 0.010085830464959145, + 0.06444399803876877, + 0.004441241733729839, + 0.010395410470664501, + 0.002631327835842967, + 0.006705896463245153, + 0.003416667226701975, + 0.0048810760490596294, + 0.003490325529128313, + 0.011658523231744766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.054975736886262894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_13/index_tensor([4678, 4678])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_13/index_tensor([4678, 4678])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c33452743c659f8b442e251abedcc33111720254 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_13/index_tensor([4678, 4678])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fluid, observational tracking photograph follows the graceful unwinding of an ancient Chinese scroll as it tumbles and swirls through an infinite, matte-black void. Maintaining a consistent medium distance, the camera moves in perfect sync with the object's slow, zero-gravity rotation, capturing its forms with stark, objective clarity. The scroll itself is crafted from fine, visibly aged rice paper, its surface painted in a palette of warm ochres and pale yellows, with subtle creases and frayed edges testifying to its antiquity. Under even, shadowless, diffuse light, providing clinical illumination, the scroll steadily unfurls, revealing intricate ink illustrations rendered with masterful calligraphic brushstrokes. Scenes of dynastic battles come to life, with armies of armored soldiers clashing with spears and swords, while elsewhere, mythical beasts such as sinuous scaly dragons and majestic qilins are depicted in varying shades of deep black and ethereal gray. The movement is deliberate and unhurried, the paper gently ripples as it rotates, inviting the viewer to absorb the historical narrative depicted, creating an atmosphere of both objective and reverent presentation of timeless artifacts.", + "seed": 12857, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_13/index_4678_seed_12857_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.039811946405097845, + "vlm_logprob": -0.039811946405097845, + "vlm_yes_no_margin": 4.225000023841858, + "vlm_token_credit": -0.039811946405097845, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15617196261882782 + ], + "opposite_logprob_tokens": [ + -2.031172037124634 + ], + "logprob_reward": -0.15617196261882782, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.15617196261882782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2937600612640381, + 0.08242392539978027, + 0.08802139759063721, + 0.1218939945101738, + 0.04369937255978584, + 0.11608903110027313, + 0.07540038228034973, + 0.09108283370733261, + 0.06687856465578079, + 0.020750416442751884 + ], + "credit_map": [ + 0.02117328904569149, + 0.023580282926559448, + 0.024067385122179985, + 0.03237803280353546, + 0.034785058349370956, + 0.0530511811375618, + 0.043379105627536774, + 0.061345744878053665, + 0.00642747525125742, + 0.017354264855384827, + 0.006392044015228748, + 0.01151115633547306, + 0.0067627327516674995, + 0.011941875331103802, + 0.009035050868988037, + 0.012999328784644604, + 0.005531002301722765, + 0.010784503072500229, + 0.006777189671993256, + 0.012193690985441208, + 0.008170660585165024, + 0.019707700237631798, + 0.0084932129830122, + 0.01636343076825142, + 0.008903088979423046, + 0.007356557529419661, + 0.013172570616006851, + 0.011306681670248508, + 0.009533761069178581, + 0.028510285541415215, + 0.013167493045330048, + 0.02994355745613575, + 0.003716507228091359, + 0.005157879553735256, + 0.005054254084825516, + 0.0037555533926934004, + 0.004102449864149094, + 0.008080370724201202, + 0.004292602650821209, + 0.009539754129946232, + 0.008884658105671406, + 0.023817433044314384, + 0.010326392948627472, + 0.017043232917785645, + 0.010240626521408558, + 0.018625665456056595, + 0.010572989471256733, + 0.016578031703829765, + 0.005961596500128508, + 0.008246480487287045, + 0.005429205950349569, + 0.006933544296771288, + 0.004584428388625383, + 0.019944583997130394, + 0.0056001851335167885, + 0.018700357526540756, + 0.012750336900353432, + 0.008670761249959469, + 0.009973550215363503, + 0.010372483171522617, + 0.008350152522325516, + 0.014375610277056694, + 0.010778630152344704, + 0.01581130549311638, + 0.008486748673021793, + 0.009114244021475315, + 0.004160925280302763, + 0.005018993280827999, + 0.0052500441670417786, + 0.014332656748592854, + 0.0049754115752875805, + 0.01553953718394041, + 0.002313571283593774, + 0.002175541827455163, + 0.001991202123463154, + 0.0024676788598299026, + 0.002537461230531335, + 0.0036640302278101444, + 0.0021270301658660173, + 0.003473900491371751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.15617196261882782 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005278578493744135 + ], + "opposite_logprob_tokens": [ + -5.255278587341309 + ], + "logprob_reward": -0.005278578493744135, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005278578493744135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09796527773141861, + 0.07213947921991348, + 0.08510587364435196, + 0.15706783533096313, + 0.07838303595781326, + 0.06204598769545555, + 0.11451178789138794, + 0.12218854576349258, + 0.16856469213962555, + 0.04202745854854584 + ], + "credit_map": [ + 0.0067347255535423756, + 0.015031580813229084, + 0.011077363975346088, + 0.008743934333324432, + 0.008609124459326267, + 0.018210656940937042, + 0.012794922105967999, + 0.016762975603342056, + 0.006064165849238634, + 0.018914885818958282, + 0.005733323749154806, + 0.010562671348452568, + 0.0044758934527635574, + 0.01020461693406105, + 0.005700188223272562, + 0.010483733378350735, + 0.007931603118777275, + 0.009523874148726463, + 0.00947418063879013, + 0.008915333077311516, + 0.0075051914900541306, + 0.014306191354990005, + 0.010961469262838364, + 0.016488024964928627, + 0.011925425380468369, + 0.00992545485496521, + 0.015133104287087917, + 0.020639212802052498, + 0.008396806195378304, + 0.03827400505542755, + 0.011949287727475166, + 0.040824536234140396, + 0.012106360867619514, + 0.008488341234624386, + 0.01064634695649147, + 0.006644032429903746, + 0.007367254700511694, + 0.013029966503381729, + 0.006189438048750162, + 0.013911299407482147, + 0.005470233969390392, + 0.004939880222082138, + 0.0036711019929498434, + 0.005376092623919249, + 0.004573874641209841, + 0.017273664474487305, + 0.0033416200894862413, + 0.017399515956640244, + 0.006809363141655922, + 0.01028082799166441, + 0.010374161414802074, + 0.010868390090763569, + 0.006711123511195183, + 0.031070824712514877, + 0.00929644238203764, + 0.029100652784109116, + 0.020887156948447227, + 0.011511078104376793, + 0.01455435249954462, + 0.013083425350487232, + 0.014847302809357643, + 0.01800283044576645, + 0.015086674131453037, + 0.014215724542737007, + 0.017748354002833366, + 0.0217380840331316, + 0.016042612493038177, + 0.01830906979739666, + 0.019506404176354408, + 0.027659285813570023, + 0.019224965944886208, + 0.028335928916931152, + 0.004101577680557966, + 0.0055231135338544846, + 0.005366024095565081, + 0.004993930459022522, + 0.003986620344221592, + 0.006953340955078602, + 0.003808296052739024, + 0.007294556125998497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.005278578493744135 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419694535434246 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.01419694535434246, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419694535434246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18612772226333618, + 0.2908059060573578, + 0.04379342496395111, + 0.12498162686824799, + 0.07593750208616257, + 0.07264154404401779, + 0.03259189426898956, + 0.07615471631288528, + 0.041793253272771835, + 0.055172398686409 + ], + "credit_map": [ + 0.013093589805066586, + 0.015771226957440376, + 0.0142371179535985, + 0.018418973311781883, + 0.01278942171484232, + 0.040012724697589874, + 0.023683642968535423, + 0.048121027648448944, + 0.019994491711258888, + 0.0457913838326931, + 0.016754742711782455, + 0.025152932852506638, + 0.023784013465046883, + 0.06622110307216644, + 0.021926693618297577, + 0.0711805447936058, + 0.003788624657317996, + 0.004976149182766676, + 0.004412372130900621, + 0.005970056168735027, + 0.002479037269949913, + 0.008024103939533234, + 0.005036244634538889, + 0.009106836281716824, + 0.01256348192691803, + 0.00919418502599001, + 0.01104974839836359, + 0.01841125264763832, + 0.009042938239872456, + 0.028322456404566765, + 0.008829758502542973, + 0.027567807585000992, + 0.0065998206846416, + 0.005843216087669134, + 0.006762810051441193, + 0.0038758835289627314, + 0.004477060865610838, + 0.0215896088629961, + 0.00438335258513689, + 0.022405749186873436, + 0.006526814308017492, + 0.007848072797060013, + 0.006707998923957348, + 0.00955821294337511, + 0.004662888590246439, + 0.015660732984542847, + 0.006731157656759024, + 0.014945661649107933, + 0.0036158161237835884, + 0.0032531926408410072, + 0.003874619258567691, + 0.004256443120539188, + 0.002322328044101596, + 0.005959019996225834, + 0.0027558475267142057, + 0.006554628722369671, + 0.017545737326145172, + 0.007262347731739283, + 0.006563750561326742, + 0.007250642403960228, + 0.006218255031853914, + 0.012257104739546776, + 0.005762043409049511, + 0.013294837437570095, + 0.005710599012672901, + 0.0047331699170172215, + 0.004250808618962765, + 0.003331240499392152, + 0.0034743580035865307, + 0.007382798008620739, + 0.00371895101852715, + 0.00919132586568594, + 0.003661528928205371, + 0.004709204658865929, + 0.0038392487913370132, + 0.0040932889096438885, + 0.0054855854250490665, + 0.014394771307706833, + 0.0045943898148834705, + 0.014394381083548069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.01419694535434246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583242177963257 + ], + "opposite_logprob_tokens": [ + -3.895583152770996 + ], + "logprob_reward": -0.020583242177963257, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583242177963257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15004780888557434, + 0.09433799982070923, + 0.06061176210641861, + 0.14774344861507416, + 0.06990484893321991, + 0.07784926146268845, + 0.07247259467840195, + 0.10295567661523819, + 0.17778368294239044, + 0.04629295691847801 + ], + "credit_map": [ + 0.01018837746232748, + 0.018156714737415314, + 0.015468557365238667, + 0.01814846694469452, + 0.013018046505749226, + 0.027079900726675987, + 0.022789884358644485, + 0.025197867304086685, + 0.007299894001334906, + 0.0197494737803936, + 0.007797602564096451, + 0.011692595668137074, + 0.005191630218178034, + 0.017652424052357674, + 0.006959953811019659, + 0.01799442432820797, + 0.009520165622234344, + 0.006159033626317978, + 0.006611574906855822, + 0.00867806002497673, + 0.004324543755501509, + 0.009337647818028927, + 0.005685586016625166, + 0.010295149870216846, + 0.008630841039121151, + 0.006986887659877539, + 0.01171655673533678, + 0.012971555814146996, + 0.0065577090717852116, + 0.04505399242043495, + 0.008482059463858604, + 0.047343846410512924, + 0.00762700941413641, + 0.007372919004410505, + 0.008855699561536312, + 0.007580749690532684, + 0.00494408281520009, + 0.012787279672920704, + 0.00630072969943285, + 0.014436371624469757, + 0.007018238306045532, + 0.008200658485293388, + 0.0060539464466273785, + 0.008269612677395344, + 0.005855128634721041, + 0.018624858930706978, + 0.006245249416679144, + 0.01758156716823578, + 0.0056446148082613945, + 0.009587201289832592, + 0.0063512022607028484, + 0.00906485877931118, + 0.00652602594345808, + 0.015002351254224777, + 0.005667863879352808, + 0.014628477394580841, + 0.010577462613582611, + 0.009320669807493687, + 0.009384026750922203, + 0.013614910654723644, + 0.009653364308178425, + 0.017132587730884552, + 0.011824632994830608, + 0.02144802361726761, + 0.023996291682124138, + 0.017975620925426483, + 0.01303673256188631, + 0.016327008605003357, + 0.012166435830295086, + 0.039190854877233505, + 0.014017520472407341, + 0.04107320308685303, + 0.0034320298582315445, + 0.007553465198725462, + 0.004042096436023712, + 0.005372358486056328, + 0.006844439078122377, + 0.007679218426346779, + 0.004057170823216438, + 0.0073121776804327965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.020583242177963257 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002829003380611539 + ], + "opposite_logprob_tokens": [ + -5.877829074859619 + ], + "logprob_reward": -0.002829003380611539, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002829003380611539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2899961471557617, + 0.10268443077802658, + 0.06535772979259491, + 0.06266248226165771, + 0.06890995800495148, + 0.0770440474152565, + 0.07918726652860641, + 0.07939620316028595, + 0.13643702864646912, + 0.0383247435092926 + ], + "credit_map": [ + 0.021252157166600227, + 0.030585523694753647, + 0.03347708284854889, + 0.03555711358785629, + 0.029932916164398193, + 0.052319008857011795, + 0.03597406670451164, + 0.05089827999472618, + 0.008739268407225609, + 0.021074125543236732, + 0.010124476626515388, + 0.015296262688934803, + 0.008195333182811737, + 0.015241404063999653, + 0.010317274369299412, + 0.013696281239390373, + 0.006333602592349052, + 0.007765618152916431, + 0.0060187773779034615, + 0.010327639058232307, + 0.006313431542366743, + 0.011636180803179741, + 0.007044964004307985, + 0.009917515330016613, + 0.005650728940963745, + 0.004702382255345583, + 0.005975052714347839, + 0.007091129664331675, + 0.005897237453609705, + 0.012502850033342838, + 0.00816454365849495, + 0.012678547762334347, + 0.009792305529117584, + 0.009114157408475876, + 0.009862282313406467, + 0.005961605813354254, + 0.006214717868715525, + 0.009941328316926956, + 0.00900911632925272, + 0.009014436043798923, + 0.00869461428374052, + 0.009565338492393494, + 0.007718078326433897, + 0.010006248950958252, + 0.007737503852695227, + 0.01268171425908804, + 0.00888785533607006, + 0.011752692051231861, + 0.005132503341883421, + 0.0087105268612504, + 0.0056118229404091835, + 0.007482028566300869, + 0.0051535507664084435, + 0.021588969975709915, + 0.00543118454515934, + 0.02007666788995266, + 0.008723821491003036, + 0.008024105802178383, + 0.00818290188908577, + 0.009383385069668293, + 0.007252656389027834, + 0.012319318018853664, + 0.011514605022966862, + 0.013995397835969925, + 0.012184335850179195, + 0.015641674399375916, + 0.011183320544660091, + 0.012158611789345741, + 0.012125030159950256, + 0.028810370713472366, + 0.011958465911448002, + 0.03237520530819893, + 0.0031158607453107834, + 0.005811887327581644, + 0.003828876418992877, + 0.0040931738913059235, + 0.004800837021321058, + 0.0061956592835485935, + 0.0035512701142579317, + 0.006927177309989929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.002829003380611539 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, observational tracking photograph follows the graceful unwinding of an ancient Chinese scroll as it tumbles and swirls through an infinite, matte-black void. Maintaining a consistent medium distance, the camera moves in perfect sync with the object's slow, zero-gravity rotation, capturing its forms with stark, objective clarity. The scroll itself is crafted from fine, visibly aged rice paper, its surface painted in a palette of warm ochres and pale yellows, with subtle creases and frayed edges testifying to its antiquity. Under even, shadowless, diffuse light, providing clinical illumination, the scroll steadily unfurls, revealing intricate ink illustrations rendered with masterful calligraphic brushstrokes. Scenes of dynastic battles come to life, with armies of armored soldiers clashing with spears and swords, while elsewhere, mythical beasts such as sinuous scaly dragons and majestic qilins are depicted in varying shades of deep black and ethereal gray. The movement is deliberate and unhurried, the paper gently ripples as it rotates, inviting the viewer to absorb the historical narrative depicted, creating an atmosphere of both objective and reverent presentation of timeless artifacts.", + "seed": 12858, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_13/index_4678_seed_12858_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06076837824657559, + "vlm_logprob": -0.06076837824657559, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.06076837824657559, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24185632169246674 + ], + "opposite_logprob_tokens": [ + -1.616856336593628 + ], + "logprob_reward": -0.24185632169246674, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24185632169246674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10320141911506653, + 0.06412772834300995, + 0.27296507358551025, + 0.059691134840250015, + 0.12534844875335693, + 0.18002429604530334, + 0.05236280709505081, + 0.07184698432683945, + 0.03787083923816681, + 0.03256131336092949 + ], + "credit_map": [ + 0.01448539923876524, + 0.014593872241675854, + 0.0120787238702178, + 0.009530777111649513, + 0.012950927019119263, + 0.012654817663133144, + 0.016045762225985527, + 0.010861136019229889, + 0.00694455998018384, + 0.008496920578181744, + 0.006094739772379398, + 0.007496885489672422, + 0.006609054747968912, + 0.010373687371611595, + 0.00994250550866127, + 0.008169375360012054, + 0.027925025671720505, + 0.006557057145982981, + 0.11047045886516571, + 0.01616833172738552, + 0.008137628436088562, + 0.015115750953555107, + 0.009050237946212292, + 0.07954057306051254, + 0.005628905724734068, + 0.004582938272505999, + 0.00655240286141634, + 0.008155051618814468, + 0.008989923633635044, + 0.006657718680799007, + 0.0061589572578668594, + 0.012965239584445953, + 0.021122507750988007, + 0.01231931522488594, + 0.012626425363123417, + 0.01596762426197529, + 0.02142203599214554, + 0.015516943298280239, + 0.012603312730789185, + 0.013770288787782192, + 0.017058955505490303, + 0.030751245096325874, + 0.016536112874746323, + 0.027172569185495377, + 0.031149884685873985, + 0.021030880510807037, + 0.016165785491466522, + 0.02015886828303337, + 0.004508099053055048, + 0.0076383925043046474, + 0.005360851529985666, + 0.007888104766607285, + 0.012077768333256245, + 0.004489657934755087, + 0.003592925611883402, + 0.006807004567235708, + 0.008697493933141232, + 0.009936819784343243, + 0.009224778972566128, + 0.00828434620052576, + 0.009602457284927368, + 0.011088426224887371, + 0.006861383561044931, + 0.00815128069370985, + 0.004619610961526632, + 0.00451694568619132, + 0.004443267825990915, + 0.004033656790852547, + 0.004619342740625143, + 0.0059624179266393185, + 0.004506582394242287, + 0.005169015843421221, + 0.002124340273439884, + 0.003305286867544055, + 0.005310531239956617, + 0.0063135926611721516, + 0.003448601346462965, + 0.0036947722546756268, + 0.00467047281563282, + 0.0036937166005373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.24185632169246674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009799469262361526 + ], + "opposite_logprob_tokens": [ + -4.634799480438232 + ], + "logprob_reward": -0.009799469262361526, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009799469262361526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23844116926193237, + 0.15000870823860168, + 0.2065460979938507, + 0.11753199994564056, + 0.07026656717061996, + 0.059191301465034485, + 0.028486762195825577, + 0.048838160932064056, + 0.031203268095850945, + 0.049485985189676285 + ], + "credit_map": [ + 0.0369226336479187, + 0.03009561076760292, + 0.03144410252571106, + 0.016907773911952972, + 0.023324768990278244, + 0.031143678352236748, + 0.03287404030561447, + 0.03572855889797211, + 0.01900005340576172, + 0.018411921337246895, + 0.02081296592950821, + 0.0179339237511158, + 0.0183212049305439, + 0.017662761732935905, + 0.02242862991988659, + 0.015437240712344646, + 0.020541014149785042, + 0.011138263158500195, + 0.06257709115743637, + 0.017302406951785088, + 0.011365063488483429, + 0.012865954078733921, + 0.01581311598420143, + 0.05494319275021553, + 0.015290015377104282, + 0.010014680214226246, + 0.013264757581055164, + 0.0121562285348773, + 0.024050455540418625, + 0.01327965036034584, + 0.014270149171352386, + 0.015206067822873592, + 0.007376761641353369, + 0.009353406727313995, + 0.009886262007057667, + 0.00880849827080965, + 0.009929259307682514, + 0.007660611066967249, + 0.009315752424299717, + 0.007936015725135803, + 0.005648000165820122, + 0.00913024228066206, + 0.005105994176119566, + 0.009053456597030163, + 0.010247420519590378, + 0.007388688158243895, + 0.0054699331521987915, + 0.007147564087063074, + 0.0024986425414681435, + 0.003981846384704113, + 0.002477138303220272, + 0.005012857262045145, + 0.0033370647579431534, + 0.00536459032446146, + 0.002562277251854539, + 0.0032523481640964746, + 0.006154696922749281, + 0.006788903847336769, + 0.005806405562907457, + 0.0056175352074205875, + 0.005644891876727343, + 0.0077351476065814495, + 0.005764627829194069, + 0.005325953010469675, + 0.0037832073867321014, + 0.00397531921043992, + 0.003571891225874424, + 0.0033125486224889755, + 0.004222690127789974, + 0.004481716081500053, + 0.005029413383454084, + 0.0028264832217246294, + 0.005455924663692713, + 0.005304506979882717, + 0.006425189319998026, + 0.005801484454423189, + 0.006228580139577389, + 0.007544151972979307, + 0.008984356187283993, + 0.0037417912390083075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.009799469262361526 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634642831981182 + ], + "opposite_logprob_tokens": [ + -3.651346445083618 + ], + "logprob_reward": -0.02634642831981182, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634642831981182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10021316260099411, + 0.10369458794593811, + 0.25949937105178833, + 0.34926366806030273, + 0.061571694910526276, + 0.03259610757231712, + 0.014721048064529896, + 0.04666098207235336, + 0.014512954279780388, + 0.017266493290662766 + ], + "credit_map": [ + 0.018766988068819046, + 0.007604612968862057, + 0.012163319624960423, + 0.006979074329137802, + 0.009623708203434944, + 0.012791616842150688, + 0.012182725593447685, + 0.020101115107536316, + 0.013256494887173176, + 0.015194110572338104, + 0.016631700098514557, + 0.008965042419731617, + 0.012485651299357414, + 0.014066836796700954, + 0.013776151463389397, + 0.009318603202700615, + 0.023699449375271797, + 0.008507913909852505, + 0.09590931981801987, + 0.01691265031695366, + 0.009772558696568012, + 0.01283795852214098, + 0.020566632971167564, + 0.07129286974668503, + 0.04116450995206833, + 0.029313741251826286, + 0.04581765830516815, + 0.03692283108830452, + 0.05435384809970856, + 0.04087505117058754, + 0.032332610338926315, + 0.0684833899140358, + 0.006395383737981319, + 0.005212596617639065, + 0.0058133648708462715, + 0.006117298733443022, + 0.01821632869541645, + 0.007447995711117983, + 0.005469609517604113, + 0.006899120286107063, + 0.0032703736796975136, + 0.004864192102104425, + 0.002829240867868066, + 0.005076874978840351, + 0.006035453174263239, + 0.0035087098367512226, + 0.0030755524057894945, + 0.003935710061341524, + 0.0014733383432030678, + 0.0018474950920790434, + 0.001400094130076468, + 0.0022013147827237844, + 0.002465160097926855, + 0.00197157496586442, + 0.00157836580183357, + 0.0017837053164839745, + 0.0051630293019115925, + 0.00616089440882206, + 0.005611736327409744, + 0.005879005417227745, + 0.006091972813010216, + 0.00767734507098794, + 0.005024424754083157, + 0.005052576772868633, + 0.001498686382547021, + 0.0016393845435231924, + 0.0011829807190224528, + 0.0020438756328076124, + 0.0020147066097706556, + 0.0016732746735215187, + 0.0022836884018033743, + 0.0021763567347079515, + 0.0012297705980017781, + 0.0017322530038654804, + 0.002762338612228632, + 0.0026664920151233673, + 0.0022742508444935083, + 0.0021407429594546556, + 0.0029156848322600126, + 0.0015449606580659747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.02634642831981182 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058148942887783 + ], + "opposite_logprob_tokens": [ + -3.8955814838409424 + ], + "logprob_reward": -0.02058148942887783, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058148942887783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1730666160583496, + 0.10153365135192871, + 0.3446332514286041, + 0.09655456990003586, + 0.07039888948202133, + 0.062080081552267075, + 0.04480849578976631, + 0.07161829620599747, + 0.014131112024188042, + 0.021175064146518707 + ], + "credit_map": [ + 0.02329937182366848, + 0.023793375119566917, + 0.02180642820894718, + 0.014084802009165287, + 0.01731455884873867, + 0.023327264934778214, + 0.030228080227971077, + 0.01921272650361061, + 0.013077089563012123, + 0.011867408640682697, + 0.015979129821062088, + 0.011357690207660198, + 0.013434099964797497, + 0.013005293905735016, + 0.011619843542575836, + 0.011193102225661278, + 0.033064182847738266, + 0.0114144841209054, + 0.13408371806144714, + 0.019114971160888672, + 0.015205129981040955, + 0.018140800297260284, + 0.013738013803958893, + 0.09987194091081619, + 0.013479530811309814, + 0.007987305521965027, + 0.008587379939854145, + 0.011838913895189762, + 0.014908503741025925, + 0.011125282384455204, + 0.015471103601157665, + 0.01315655279904604, + 0.00918969139456749, + 0.007170361932367086, + 0.008081190288066864, + 0.006959353107959032, + 0.010558432899415493, + 0.00890486128628254, + 0.010250742547214031, + 0.009284256026148796, + 0.006539574824273586, + 0.008879398927092552, + 0.006184687372297049, + 0.009556017816066742, + 0.010858518071472645, + 0.007147364784032106, + 0.005459124222397804, + 0.007455397397279739, + 0.004210981540381908, + 0.006762832403182983, + 0.005059476010501385, + 0.0073936996050179005, + 0.00635888334363699, + 0.004967672284692526, + 0.0039633591659367085, + 0.006091592367738485, + 0.007785616908222437, + 0.008479391224682331, + 0.007148445583879948, + 0.008810843341052532, + 0.009616650640964508, + 0.011446107178926468, + 0.00953005626797676, + 0.008801182731986046, + 0.0017498753732070327, + 0.0018213634612038732, + 0.0014806787949055433, + 0.001972389640286565, + 0.0015922186430543661, + 0.001555529423058033, + 0.0027734555769711733, + 0.0011856015771627426, + 0.0021955533884465694, + 0.0029433206655085087, + 0.0033417658414691687, + 0.002354688011109829, + 0.002494528889656067, + 0.0029430678114295006, + 0.003169258823618293, + 0.0017328803660348058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.02058148942887783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258182529360056 + ], + "opposite_logprob_tokens": [ + -5.255258083343506 + ], + "logprob_reward": -0.005258182529360056, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258182529360056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12294691801071167, + 0.1356492042541504, + 0.2203378528356552, + 0.05926521494984627, + 0.15569528937339783, + 0.07910653203725815, + 0.06523020565509796, + 0.08228829503059387, + 0.04990687593817711, + 0.02957358956336975 + ], + "credit_map": [ + 0.016080865636467934, + 0.01724150776863098, + 0.013635993003845215, + 0.011825522407889366, + 0.012938739731907845, + 0.0157147329300642, + 0.01983395405113697, + 0.015675606206059456, + 0.016992202028632164, + 0.017016664147377014, + 0.02883385494351387, + 0.013036930002272129, + 0.012616308405995369, + 0.017506463453173637, + 0.014241012744605541, + 0.01540576945990324, + 0.02390895038843155, + 0.006843926850706339, + 0.0875491127371788, + 0.012713122181594372, + 0.011763576418161392, + 0.01238387729972601, + 0.00811784528195858, + 0.05705743655562401, + 0.005232648923993111, + 0.005501523148268461, + 0.008545379154384136, + 0.006828083656728268, + 0.008605457842350006, + 0.007996532134711742, + 0.006506900303065777, + 0.010048688389360905, + 0.014107819646596909, + 0.015229704789817333, + 0.012403154745697975, + 0.020054886117577553, + 0.040297217667102814, + 0.023153208196163177, + 0.012865656986832619, + 0.017583642154932022, + 0.007036848459392786, + 0.011805146001279354, + 0.007237832993268967, + 0.011858426034450531, + 0.014650888741016388, + 0.009242648258805275, + 0.008302499540150166, + 0.008972239680588245, + 0.005863957107067108, + 0.010021267458796501, + 0.007814476266503334, + 0.01221753191202879, + 0.011814713478088379, + 0.005012315232306719, + 0.004505292046815157, + 0.007980650290846825, + 0.010358772240579128, + 0.010858464986085892, + 0.011340634897351265, + 0.012018334120512009, + 0.010713442228734493, + 0.011057238094508648, + 0.007434968836605549, + 0.00850643590092659, + 0.00538542540743947, + 0.004883689805865288, + 0.004881354048848152, + 0.008601231500506401, + 0.006474288646131754, + 0.005710100755095482, + 0.005796914920210838, + 0.008173870854079723, + 0.0021970123052597046, + 0.0031395910773426294, + 0.0046218885108828545, + 0.006957093719393015, + 0.0033971904776990414, + 0.002180516254156828, + 0.0038945521228015423, + 0.0031857453286647797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.005258182529360056 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, observational tracking photograph follows the graceful unwinding of an ancient Chinese scroll as it tumbles and swirls through an infinite, matte-black void. Maintaining a consistent medium distance, the camera moves in perfect sync with the object's slow, zero-gravity rotation, capturing its forms with stark, objective clarity. The scroll itself is crafted from fine, visibly aged rice paper, its surface painted in a palette of warm ochres and pale yellows, with subtle creases and frayed edges testifying to its antiquity. Under even, shadowless, diffuse light, providing clinical illumination, the scroll steadily unfurls, revealing intricate ink illustrations rendered with masterful calligraphic brushstrokes. Scenes of dynastic battles come to life, with armies of armored soldiers clashing with spears and swords, while elsewhere, mythical beasts such as sinuous scaly dragons and majestic qilins are depicted in varying shades of deep black and ethereal gray. The movement is deliberate and unhurried, the paper gently ripples as it rotates, inviting the viewer to absorb the historical narrative depicted, creating an atmosphere of both objective and reverent presentation of timeless artifacts.", + "seed": 12859, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_13/index_4678_seed_12859_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02440545433200896, + "vlm_logprob": -0.02440545433200896, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.02440545433200896, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08218786120414734 + ], + "opposite_logprob_tokens": [ + -2.7071878910064697 + ], + "logprob_reward": -0.08218786120414734, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08218786120414734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14929385483264923, + 0.03187071904540062, + 0.13854685425758362, + 0.15215924382209778, + 0.1885627806186676, + 0.11051974445581436, + 0.02166065387427807, + 0.1370009034872055, + 0.051986075937747955, + 0.01839914731681347 + ], + "credit_map": [ + 0.019293084740638733, + 0.018074266612529755, + 0.020277895033359528, + 0.016985533758997917, + 0.02029755711555481, + 0.010806226171553135, + 0.02235858142375946, + 0.021200722083449364, + 0.0030969700310379267, + 0.006314827594906092, + 0.004682986997067928, + 0.003636974608525634, + 0.003555147908627987, + 0.0031404360197484493, + 0.003862629411742091, + 0.003580748802050948, + 0.011740664020180702, + 0.011685099452733994, + 0.03696999326348305, + 0.019961455836892128, + 0.02206536754965782, + 0.009970779530704021, + 0.012578492052853107, + 0.013575010932981968, + 0.03682742640376091, + 0.01697530224919319, + 0.016182664781808853, + 0.01925060525536537, + 0.014426618814468384, + 0.017453549429774284, + 0.013753884471952915, + 0.017289195209741592, + 0.02227441594004631, + 0.014715001918375492, + 0.03360256925225258, + 0.012927372008562088, + 0.028627656400203705, + 0.019128527492284775, + 0.02085411362349987, + 0.03643312677741051, + 0.01317878719419241, + 0.009362875483930111, + 0.012612810358405113, + 0.013243082910776138, + 0.020638195797801018, + 0.008650122210383415, + 0.012006847187876701, + 0.0208270326256752, + 0.0018415347440168262, + 0.004041270352900028, + 0.0021322937682271004, + 0.0031851159874349833, + 0.0035925367847085, + 0.002148900879547, + 0.0021672272123396397, + 0.002551775425672531, + 0.01621718518435955, + 0.017464715987443924, + 0.01574004627764225, + 0.022423477843403816, + 0.01701115444302559, + 0.016002239659428596, + 0.011207661591470242, + 0.02093442529439926, + 0.004319446161389351, + 0.008715554140508175, + 0.008344367146492004, + 0.004635057877749205, + 0.006383359897881746, + 0.006435450166463852, + 0.005326796323060989, + 0.007826047018170357, + 0.0016049747355282307, + 0.0022678717505186796, + 0.001973262056708336, + 0.001867502462118864, + 0.002634683158248663, + 0.0027904787566512823, + 0.0027482507284730673, + 0.0025121255312114954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.08218786120414734 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751345470547676 + ], + "opposite_logprob_tokens": [ + -5.006751537322998 + ], + "logprob_reward": -0.006751345470547676, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751345470547676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08264303207397461, + 0.0370696559548378, + 0.37760069966316223, + 0.04022112488746643, + 0.13831305503845215, + 0.12801936268806458, + 0.054030198603868484, + 0.07275207340717316, + 0.04931843653321266, + 0.02003234438598156 + ], + "credit_map": [ + 0.009724237024784088, + 0.009897478856146336, + 0.00809602253139019, + 0.011340085417032242, + 0.011608664877712727, + 0.007867484353482723, + 0.012217216193675995, + 0.011891849339008331, + 0.0037702082190662622, + 0.0059942868538200855, + 0.005361259449273348, + 0.00537066999822855, + 0.003674970706924796, + 0.00476469611749053, + 0.0044140806421637535, + 0.003719485364854336, + 0.03283078968524933, + 0.02684221975505352, + 0.11358032375574112, + 0.05161594972014427, + 0.06290093809366226, + 0.023077581077814102, + 0.02810727246105671, + 0.038645606487989426, + 0.007638267241418362, + 0.0034408466890454292, + 0.0047480156645178795, + 0.005901353899389505, + 0.004501963034272194, + 0.0055472939275205135, + 0.00400040578097105, + 0.004442978650331497, + 0.012572691775858402, + 0.013557891361415386, + 0.013585853390395641, + 0.006680828984826803, + 0.038404203951358795, + 0.01449570246040821, + 0.01232845801860094, + 0.02668742649257183, + 0.01446434948593378, + 0.012781714089214802, + 0.011789356358349323, + 0.015085199847817421, + 0.02616550400853157, + 0.013224691152572632, + 0.012356421910226345, + 0.022152123972773552, + 0.004384488798677921, + 0.008808736689388752, + 0.004525092896074057, + 0.0058945948258042336, + 0.009815197438001633, + 0.006595853250473738, + 0.006196649745106697, + 0.007809583563357592, + 0.00700386380776763, + 0.00900157168507576, + 0.00843583233654499, + 0.01172387134283781, + 0.008706868626177311, + 0.008603792637586594, + 0.006279758643358946, + 0.012996511533856392, + 0.00457044830545783, + 0.007127964869141579, + 0.008511917665600777, + 0.006478285416960716, + 0.005405214615166187, + 0.004429983906447887, + 0.005466609261929989, + 0.007328012026846409, + 0.0016752913361415267, + 0.0026076948270201683, + 0.002133340574800968, + 0.0021711240988224745, + 0.002784746466204524, + 0.002904355525970459, + 0.002882089465856552, + 0.002873701276257634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.006751345470547676 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606801524758339 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.01606801524758339, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606801524758339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13255499303340912, + 0.11072138696908951, + 0.13234657049179077, + 0.07975565642118454, + 0.38902321457862854, + 0.035470880568027496, + 0.04138825088739395, + 0.048176173120737076, + 0.013869130052626133, + 0.01669365167617798 + ], + "credit_map": [ + 0.020440159365534782, + 0.012475607916712761, + 0.017658943310379982, + 0.014376942068338394, + 0.02152356691658497, + 0.007829627022147179, + 0.017727453261613846, + 0.0205227117985487, + 0.01236714981496334, + 0.024684002622961998, + 0.0143660893663764, + 0.009883882477879524, + 0.011808345094323158, + 0.01416828390210867, + 0.012338985688984394, + 0.011104663833975792, + 0.011876435950398445, + 0.013999376446008682, + 0.025145653635263443, + 0.02250441163778305, + 0.02014707773923874, + 0.011378839612007141, + 0.01509210467338562, + 0.012202686630189419, + 0.018001018092036247, + 0.008101743645966053, + 0.0069097005762159824, + 0.011261453852057457, + 0.009018192999064922, + 0.011237598024308681, + 0.007366318721324205, + 0.007859639823436737, + 0.053589094430208206, + 0.037131909281015396, + 0.031207697466015816, + 0.02051691897213459, + 0.09133772552013397, + 0.05092253535985947, + 0.03686719015240669, + 0.06745019555091858, + 0.0032007985282689333, + 0.0061237323097884655, + 0.004567478317767382, + 0.0041640629060566425, + 0.005825710482895374, + 0.0041238307021558285, + 0.003665377153083682, + 0.0037998948246240616, + 0.003344933269545436, + 0.005530864931643009, + 0.0032908334396779537, + 0.005075887311249971, + 0.007453316356986761, + 0.004344552289694548, + 0.005085711367428303, + 0.007262160535901785, + 0.004047672729939222, + 0.007371441461145878, + 0.0055619459599256516, + 0.007595219183713198, + 0.007149297744035721, + 0.006464161444455385, + 0.003651693696156144, + 0.006334749516099691, + 0.0019523119553923607, + 0.0016647609882056713, + 0.002023168373852968, + 0.0016136878402903676, + 0.0019349181093275547, + 0.0014476360520347953, + 0.0012852762592956424, + 0.0019473722204566002, + 0.0013765443582087755, + 0.002479644026607275, + 0.0014150702627375722, + 0.0011816145852208138, + 0.0031539949122816324, + 0.0028747599571943283, + 0.0020650874357670546, + 0.0021469395142048597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.01606801524758339 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193067327141762 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014193067327141762, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193067327141762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19238121807575226, + 0.05204935744404793, + 0.07318389415740967, + 0.060715269297361374, + 0.11358962208032608, + 0.1952327936887741, + 0.0569775216281414, + 0.1454346477985382, + 0.0647701695561409, + 0.04566555842757225 + ], + "credit_map": [ + 0.031709954142570496, + 0.02230161800980568, + 0.023343337699770927, + 0.02072104997932911, + 0.029312629252672195, + 0.012976093217730522, + 0.024552563205361366, + 0.02746395394206047, + 0.005350280087441206, + 0.01042480394244194, + 0.007346716709434986, + 0.007518702652305365, + 0.005390736740082502, + 0.005212870426476002, + 0.00541403004899621, + 0.005391215439885855, + 0.0074103716760873795, + 0.007026560604572296, + 0.014523366466164589, + 0.008778158575296402, + 0.014909856021404266, + 0.008065875619649887, + 0.006680739112198353, + 0.005788959562778473, + 0.006444149650633335, + 0.007415712345391512, + 0.006083888001739979, + 0.013569358736276627, + 0.007390662562102079, + 0.006004394497722387, + 0.008160948753356934, + 0.005646150559186935, + 0.010140174068510532, + 0.012665803544223309, + 0.011693337932229042, + 0.007769786287099123, + 0.030315430834889412, + 0.012919323518872261, + 0.012056218460202217, + 0.016029536724090576, + 0.028911612927913666, + 0.021503755822777748, + 0.02550196275115013, + 0.017901919782161713, + 0.030028210952878, + 0.014836915768682957, + 0.018383648246526718, + 0.038164764642715454, + 0.004939147271215916, + 0.011457609944045544, + 0.004782898351550102, + 0.006457339972257614, + 0.013557957485318184, + 0.0047935014590620995, + 0.004032904282212257, + 0.006956158671528101, + 0.016069838777184486, + 0.018093394115567207, + 0.0166216678917408, + 0.022907033562660217, + 0.017288418486714363, + 0.018812788650393486, + 0.014501214027404785, + 0.021140282973647118, + 0.004827951546758413, + 0.012676292099058628, + 0.01033666543662548, + 0.00728228222578764, + 0.006916759070008993, + 0.005365250166505575, + 0.008093824610114098, + 0.009271137416362762, + 0.003643093630671501, + 0.006204608362168074, + 0.004152266308665276, + 0.0033551494125276804, + 0.0069041200913488865, + 0.008580442517995834, + 0.0065302555449306965, + 0.0062956176698207855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.014193067327141762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028269824106246233 + ], + "opposite_logprob_tokens": [ + -5.877827167510986 + ], + "logprob_reward": -0.0028269824106246233, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028269824106246233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24430853128433228, + 0.0997253805398941, + 0.10149902105331421, + 0.03993195667862892, + 0.11242018640041351, + 0.1305285096168518, + 0.07843006402254105, + 0.08544499427080154, + 0.061019591987133026, + 0.04669172316789627 + ], + "credit_map": [ + 0.029555944725871086, + 0.02819993533194065, + 0.026417382061481476, + 0.03250456228852272, + 0.036612678319215775, + 0.021580372005701065, + 0.036978770047426224, + 0.03245887905359268, + 0.011685643345117569, + 0.023102281615138054, + 0.013430586084723473, + 0.010018087923526764, + 0.010258927941322327, + 0.010595189407467842, + 0.011455667205154896, + 0.009179004468023777, + 0.013367182575166225, + 0.008577401749789715, + 0.023147093132138252, + 0.010801596567034721, + 0.02049414999783039, + 0.009809707291424274, + 0.006664158310741186, + 0.00863772351294756, + 0.006229851394891739, + 0.0039762891829013824, + 0.0035227176267653704, + 0.006782106589525938, + 0.005439680069684982, + 0.006655449513345957, + 0.003597626695409417, + 0.0037282344419509172, + 0.011567291803658009, + 0.00947846844792366, + 0.016036257147789, + 0.008590217679738998, + 0.01977054961025715, + 0.013094366528093815, + 0.014183617196977139, + 0.019699405878782272, + 0.020704973489046097, + 0.014826455153524876, + 0.007418341469019651, + 0.006447544787079096, + 0.030533000826835632, + 0.00822348240762949, + 0.007051944732666016, + 0.03532276675105095, + 0.007675910834223032, + 0.009223801083862782, + 0.00772093003615737, + 0.007294523064047098, + 0.011901876889169216, + 0.0091001708060503, + 0.009582794271409512, + 0.0159300547093153, + 0.009495582431554794, + 0.01171016227453947, + 0.012742111459374428, + 0.009013879112899303, + 0.012335343286395073, + 0.008867045864462852, + 0.005209601018577814, + 0.016071267426013947, + 0.004382619634270668, + 0.008165406994521618, + 0.008114999160170555, + 0.005594726651906967, + 0.011724064126610756, + 0.009328852407634258, + 0.005584498401731253, + 0.00812442321330309, + 0.003966533113270998, + 0.0060338787734508514, + 0.004074902273714542, + 0.0037343280855566263, + 0.007460893131792545, + 0.008552796207368374, + 0.0067252349108457565, + 0.006143155042082071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0028269824106246233 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, observational tracking photograph follows the graceful unwinding of an ancient Chinese scroll as it tumbles and swirls through an infinite, matte-black void. Maintaining a consistent medium distance, the camera moves in perfect sync with the object's slow, zero-gravity rotation, capturing its forms with stark, objective clarity. The scroll itself is crafted from fine, visibly aged rice paper, its surface painted in a palette of warm ochres and pale yellows, with subtle creases and frayed edges testifying to its antiquity. Under even, shadowless, diffuse light, providing clinical illumination, the scroll steadily unfurls, revealing intricate ink illustrations rendered with masterful calligraphic brushstrokes. Scenes of dynastic battles come to life, with armies of armored soldiers clashing with spears and swords, while elsewhere, mythical beasts such as sinuous scaly dragons and majestic qilins are depicted in varying shades of deep black and ethereal gray. The movement is deliberate and unhurried, the paper gently ripples as it rotates, inviting the viewer to absorb the historical narrative depicted, creating an atmosphere of both objective and reverent presentation of timeless artifacts.", + "seed": 12860, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_13/index_4678_seed_12860_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09536885223351418, + "vlm_logprob": -0.09536885223351418, + "vlm_yes_no_margin": 3.999999988079071, + "vlm_token_credit": -0.09536885223351418, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43801552057266235 + ], + "opposite_logprob_tokens": [ + -1.0630154609680176 + ], + "logprob_reward": -0.43801552057266235, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.43801552057266235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17040598392486572, + 0.10503517836332321, + 0.13030152022838593, + 0.08701534569263458, + 0.06714428216218948, + 0.26363271474838257, + 0.08100393414497375, + 0.035667747259140015, + 0.026661571115255356, + 0.03313170000910759 + ], + "credit_map": [ + 0.03876844421029091, + 0.014848655089735985, + 0.03982925042510033, + 0.014164100401103497, + 0.01473834179341793, + 0.015409234911203384, + 0.013124086894094944, + 0.019523857161402702, + 0.026163766160607338, + 0.007671538274735212, + 0.01982797123491764, + 0.009272641502320766, + 0.011774367652833462, + 0.008753931149840355, + 0.009605664759874344, + 0.011965293437242508, + 0.02308446168899536, + 0.01856289990246296, + 0.020305559039115906, + 0.011876728385686874, + 0.01597108505666256, + 0.012742680497467518, + 0.013714104890823364, + 0.014044005423784256, + 0.011914028786122799, + 0.010607358068227768, + 0.014001034200191498, + 0.011824318207800388, + 0.0117020970210433, + 0.009779942221939564, + 0.008116085082292557, + 0.009070483967661858, + 0.0157184861600399, + 0.005797241348773241, + 0.01072488073259592, + 0.0068055386655032635, + 0.007035341579467058, + 0.00583255011588335, + 0.005647675134241581, + 0.009582568891346455, + 0.025256624445319176, + 0.008676042780280113, + 0.10898197442293167, + 0.011960913427174091, + 0.010464908555150032, + 0.011571268551051617, + 0.012515494599938393, + 0.07420551776885986, + 0.009355274960398674, + 0.013212627731263638, + 0.008525529876351357, + 0.01104458887130022, + 0.010185602121055126, + 0.009744676761329174, + 0.008077424950897694, + 0.010858207009732723, + 0.00356146227568388, + 0.004920760169625282, + 0.004070417024195194, + 0.006815129891037941, + 0.004191161599010229, + 0.003348407568410039, + 0.0036058584228157997, + 0.005154546815901995, + 0.003904797602444887, + 0.0031109617557376623, + 0.0037408550269901752, + 0.002456812420859933, + 0.003289902349933982, + 0.00357668730430305, + 0.0021412945352494717, + 0.004440260119736195, + 0.00610910402610898, + 0.003712309058755636, + 0.0038881588261574507, + 0.0036562737077474594, + 0.0032356888987123966, + 0.0034867620561271906, + 0.003846270963549614, + 0.005197132471948862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.43801552057266235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059714484959840775 + ], + "opposite_logprob_tokens": [ + -5.130971431732178 + ], + "logprob_reward": -0.0059714484959840775, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059714484959840775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16870667040348053, + 0.1865941286087036, + 0.12709705531597137, + 0.11811186373233795, + 0.04504210129380226, + 0.17095431685447693, + 0.04502294585108757, + 0.03371481969952583, + 0.03369387239217758, + 0.07106214761734009 + ], + "credit_map": [ + 0.03951539471745491, + 0.009993740357458591, + 0.02728947438299656, + 0.01089118979871273, + 0.011238272301852703, + 0.017810126766562462, + 0.010068010538816452, + 0.04190045967698097, + 0.022620894014835358, + 0.021038997918367386, + 0.028780994936823845, + 0.028827611356973648, + 0.024219492450356483, + 0.01829208992421627, + 0.017790861427783966, + 0.02502317912876606, + 0.022765392437577248, + 0.016529923304915428, + 0.019513288512825966, + 0.013396542519330978, + 0.018835965543985367, + 0.010843583382666111, + 0.013035858049988747, + 0.012176502496004105, + 0.02273380197584629, + 0.013229475356638432, + 0.019677290692925453, + 0.015311596915125847, + 0.013349766843020916, + 0.013273537158966064, + 0.010290073230862617, + 0.010246322490274906, + 0.007180360145866871, + 0.005784640088677406, + 0.004557543899863958, + 0.004843933042138815, + 0.00517474627122283, + 0.0049268826842308044, + 0.007090047933161259, + 0.005483947228640318, + 0.024018626660108566, + 0.0067154415883123875, + 0.07143975794315338, + 0.009721006266772747, + 0.008972050622105598, + 0.008217566646635532, + 0.008507083170115948, + 0.03336279094219208, + 0.0049319788813591, + 0.006215574219822884, + 0.0052292305044829845, + 0.006797866430133581, + 0.0044288309291005135, + 0.005293253343552351, + 0.006180890370160341, + 0.0059453221037983894, + 0.0048700300976634026, + 0.0038727102801203728, + 0.004123606253415346, + 0.00357546703889966, + 0.004594291560351849, + 0.00407453253865242, + 0.0038133126217871904, + 0.0047908672131598, + 0.003968424629420042, + 0.0039908853359520435, + 0.004570480436086655, + 0.0053652082569897175, + 0.004655137658119202, + 0.0030253990553319454, + 0.003059694077819586, + 0.0050586434081196785, + 0.010675950907170773, + 0.007439733482897282, + 0.01023222878575325, + 0.006344058085232973, + 0.007745558861643076, + 0.008984461426734924, + 0.00949912890791893, + 0.010141023434698582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.0059714484959840775 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645959198474884 + ], + "opposite_logprob_tokens": [ + -4.758646011352539 + ], + "logprob_reward": -0.008645959198474884, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645959198474884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1050056666135788, + 0.0661356970667839, + 0.17723718285560608, + 0.4359610378742218, + 0.02371743693947792, + 0.08006138354539871, + 0.020310500636696815, + 0.0430600643157959, + 0.013102622702717781, + 0.03540840744972229 + ], + "credit_map": [ + 0.031336694955825806, + 0.0066892956383526325, + 0.020251719281077385, + 0.007194762118160725, + 0.007601342163980007, + 0.007502926513552666, + 0.00644953828305006, + 0.017979396507143974, + 0.01460087951272726, + 0.006142960395663977, + 0.012404079549014568, + 0.007241560611873865, + 0.00846251193434, + 0.00483065377920866, + 0.005364969838410616, + 0.0070880865678191185, + 0.03502015396952629, + 0.018465740606188774, + 0.028687968850135803, + 0.01627248153090477, + 0.021723594516515732, + 0.01716487482190132, + 0.016638705506920815, + 0.02326369099318981, + 0.1039193868637085, + 0.03356154263019562, + 0.08570723235607147, + 0.0557621493935585, + 0.04609261080622673, + 0.04817531630396843, + 0.02894572727382183, + 0.03379707783460617, + 0.00334603781811893, + 0.002442321041598916, + 0.003949607256799936, + 0.003261120989918709, + 0.0027720185462385416, + 0.0022199361119419336, + 0.0025917382445186377, + 0.003134658560156822, + 0.008347972296178341, + 0.004012678284198046, + 0.03164339065551758, + 0.003985131159424782, + 0.003872248576954007, + 0.0035045749973505735, + 0.004255130887031555, + 0.02044026181101799, + 0.004568355157971382, + 0.0022786015179008245, + 0.0033796203788369894, + 0.0020263460464775562, + 0.0016064582159742713, + 0.002078084275126457, + 0.0015149067621678114, + 0.0028581288643181324, + 0.007447483949363232, + 0.0063567692413926125, + 0.005087438505142927, + 0.005502792075276375, + 0.004499521572142839, + 0.0039036294911056757, + 0.0034053015988320112, + 0.006857133470475674, + 0.002151284832507372, + 0.0012211119756102562, + 0.001866802922450006, + 0.0011592828668653965, + 0.0015930818626657128, + 0.0016305497847497463, + 0.0015099721495062113, + 0.001970537705346942, + 0.0064898221753537655, + 0.0037711667828261852, + 0.0049559734761714935, + 0.004603336565196514, + 0.002837172243744135, + 0.0030739116482436657, + 0.0035710216034203768, + 0.006106007378548384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.008645959198474884 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583825185894966 + ], + "opposite_logprob_tokens": [ + -3.8955838680267334 + ], + "logprob_reward": -0.020583825185894966, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583825185894966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12044751644134521, + 0.1356426179409027, + 0.11412698775529861, + 0.13353893160820007, + 0.0702495351433754, + 0.06948229670524597, + 0.08043034374713898, + 0.09636156260967255, + 0.09121569246053696, + 0.08850443363189697 + ], + "credit_map": [ + 0.034846335649490356, + 0.00938215758651495, + 0.02654077485203743, + 0.008370401337742805, + 0.010970517061650753, + 0.009263393469154835, + 0.007663730066269636, + 0.013410216197371483, + 0.02861161343753338, + 0.008159836754202843, + 0.022213149815797806, + 0.014977099373936653, + 0.01695738174021244, + 0.011366845108568668, + 0.014697813428938389, + 0.018658898770809174, + 0.01667935959994793, + 0.013076637871563435, + 0.017697792500257492, + 0.012913400307297707, + 0.013318592682480812, + 0.011595256626605988, + 0.012239828705787659, + 0.016606129705905914, + 0.020474357530474663, + 0.01528181228786707, + 0.024035422131419182, + 0.016418900340795517, + 0.01853087730705738, + 0.012836289592087269, + 0.015560219064354897, + 0.010401078499853611, + 0.008017490617930889, + 0.00729507626965642, + 0.014357196167111397, + 0.005368269048631191, + 0.005379322916269302, + 0.0054599447175860405, + 0.009912299923598766, + 0.014459943398833275, + 0.008286015130579472, + 0.005302081350237131, + 0.021255772560834885, + 0.004895394202321768, + 0.006675164680927992, + 0.005698360037058592, + 0.006911647971719503, + 0.010457867756485939, + 0.016798343509435654, + 0.009216106496751308, + 0.013972153887152672, + 0.009005749598145485, + 0.008419663645327091, + 0.006546459626406431, + 0.005951660219579935, + 0.010520216077566147, + 0.01843445748090744, + 0.013708367012441158, + 0.012043759226799011, + 0.009825166314840317, + 0.009739420376718044, + 0.009822371415793896, + 0.00788891315460205, + 0.014899123460054398, + 0.01680837944149971, + 0.01045263186097145, + 0.01328306831419468, + 0.009467393159866333, + 0.010112804360687733, + 0.009033235721290112, + 0.007589792367070913, + 0.014468394219875336, + 0.013174479827284813, + 0.00905972346663475, + 0.011003254912793636, + 0.011380289681255817, + 0.009774507023394108, + 0.010196560062468052, + 0.008858428336679935, + 0.015057199634611607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.020583825185894966 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient Chinese scroll made of visibly aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scroll moving in a zero-gravity rotation while tumbling and swirling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background an infinite matte-black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a consistent medium distance and move in sync with the scroll's rotation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the scroll's illustrations rendered in warm ochres, pale yellows, deep black, and ethereal gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036275077145546675 + ], + "opposite_logprob_tokens": [ + -5.628627300262451 + ], + "logprob_reward": -0.0036275077145546675, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036275077145546675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14182394742965698, + 0.08138274401426315, + 0.10024388879537582, + 0.08454793691635132, + 0.07202337682247162, + 0.2868463695049286, + 0.04690748080611229, + 0.08709846436977386, + 0.028673596680164337, + 0.07045216858386993 + ], + "credit_map": [ + 0.036540836095809937, + 0.011238298378884792, + 0.021305439993739128, + 0.012937800027430058, + 0.014518042095005512, + 0.01029564905911684, + 0.012114998884499073, + 0.022872887551784515, + 0.020759141072630882, + 0.006244753021746874, + 0.016046660020947456, + 0.0061231511645019054, + 0.008351366966962814, + 0.006732691545039415, + 0.008762028068304062, + 0.00836296658962965, + 0.016543615609407425, + 0.011310088448226452, + 0.016232771798968315, + 0.010088644921779633, + 0.011562367901206017, + 0.011547288857400417, + 0.010695272125303745, + 0.012263845652341843, + 0.019748378545045853, + 0.006711394991725683, + 0.015448967926204205, + 0.008943526074290276, + 0.010212507098913193, + 0.009072099812328815, + 0.007726901210844517, + 0.006684169173240662, + 0.010129570960998535, + 0.010058222338557243, + 0.0076014031656086445, + 0.011263209395110607, + 0.008935756050050259, + 0.007693144492805004, + 0.010780862532556057, + 0.005561213940382004, + 0.030002057552337646, + 0.013239878229796886, + 0.11674054712057114, + 0.015480748377740383, + 0.01217626128345728, + 0.011989924125373363, + 0.014458375051617622, + 0.07275861501693726, + 0.0063654896803200245, + 0.0083319628611207, + 0.004197058733552694, + 0.00635073846206069, + 0.004593274090439081, + 0.005894708447158337, + 0.0050973608158528805, + 0.006076889578253031, + 0.012344403192400932, + 0.013505269773304462, + 0.010645818896591663, + 0.009871021844446659, + 0.009652228094637394, + 0.008405723609030247, + 0.00948959868401289, + 0.013184410519897938, + 0.0046441033482551575, + 0.002810159232467413, + 0.004268208052963018, + 0.002684162463992834, + 0.003217682708054781, + 0.004113602917641401, + 0.002421586075797677, + 0.0045140935108065605, + 0.015124996192753315, + 0.008690605871379375, + 0.008424620144069195, + 0.0061304508708417416, + 0.0060265278443694115, + 0.007028098218142986, + 0.007183659356087446, + 0.011843213811516762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.0036275077145546675 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_14/index_tensor([8471, 8471])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_14/index_tensor([8471, 8471])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ff14f7d3ee181415cdce65bd3e5176a8e8b35490 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_14/index_tensor([8471, 8471])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro close-up, shot at a high frame rate to create a breathtaking super-slow-motion effect, focuses on a crimson rose bud, about to bloom. In the soft, diffused, cool light of early morning, each velvety rose petal is clung to countless glistening, plump beads of morning dew. The background, a completely blurred, hazy emerald green, underscores the rose's aloofness and purity. As the flower is lifted upwards by an invisible force, the camera tilts slightly upward, capturing the subtle effects of gravity: The spherical dewdrops begin to lose their grip, stretching across the petal's smooth surface, forming tiny streaks before gracefully peeling off the edge, one by one, and sliding off. Captured in super-slow motion, each droplet, in its moment of descent, resembles a sparkling diamond, its surface shimmering, its form subtly shifting in mid-air before vanishing beyond the depth of field. The entire process is imbued with tranquility, poetry, and a fleeting, exquisite beauty.", + "seed": 11364, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_14/index_8471_seed_11364_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15116116516292094, + "vlm_logprob": -0.15116116516292094, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.15116116516292094, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00958271138370037 + ], + "opposite_logprob_tokens": [ + -5.00958251953125 + ], + "logprob_reward": -0.00958271138370037, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00958271138370037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1001117005944252, + 0.19257719814777374, + 0.05220583453774452, + 0.3884906470775604, + 0.06246159225702286, + 0.08013007044792175, + 0.04268923029303551, + 0.009805602952837944, + 0.015171850100159645, + 0.056356240063905716 + ], + "credit_map": [ + 0.011294915340840816, + 0.005294304806739092, + 0.017641840502619743, + 0.010489429347217083, + 0.003796132979914546, + 0.014577873051166534, + 0.00640462152659893, + 0.030612586066126823, + 0.014848456718027592, + 0.00658889627084136, + 0.008777206763625145, + 0.01240063738077879, + 0.0067671192809939384, + 0.06820666044950485, + 0.007670942693948746, + 0.06731726229190826, + 0.0035761010367423296, + 0.003836410352960229, + 0.004389425739645958, + 0.005203924607485533, + 0.004109697416424751, + 0.011624300852417946, + 0.007831434719264507, + 0.011634537018835545, + 0.011181138455867767, + 0.016434818506240845, + 0.01687800884246826, + 0.019350968301296234, + 0.008471699431538582, + 0.15315121412277222, + 0.020117998123168945, + 0.14290480315685272, + 0.0035236680414527655, + 0.004602695815265179, + 0.003969997633248568, + 0.008545411750674248, + 0.004602852277457714, + 0.017730582505464554, + 0.005505354609340429, + 0.01398103404790163, + 0.0061580645851790905, + 0.005982063710689545, + 0.007596071343868971, + 0.006984068546444178, + 0.005820638965815306, + 0.020470889285206795, + 0.00911248754709959, + 0.01800578273832798, + 0.0029455972835421562, + 0.0035396625753492117, + 0.005810696631669998, + 0.0072533320635557175, + 0.002600145759060979, + 0.0077268159948289394, + 0.004967126064002514, + 0.007845855318009853, + 0.0007031793938949704, + 0.0008443177212029696, + 0.001272312249056995, + 0.001268238527700305, + 0.0008275927393697202, + 0.002138725481927395, + 0.001434542122296989, + 0.0013166950084269047, + 0.0016981663648039103, + 0.0015506490599364042, + 0.002714521950110793, + 0.002125590108335018, + 0.0011558271944522858, + 0.0019329697825014591, + 0.002016843296587467, + 0.0019772821106016636, + 0.007264423184096813, + 0.006738188210874796, + 0.005423156078904867, + 0.008164304308593273, + 0.006411979906260967, + 0.006876959931105375, + 0.008672185242176056, + 0.006805048324167728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.00958271138370037 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297035962343216 + ], + "opposite_logprob_tokens": [ + -2.2379703521728516 + ], + "logprob_reward": -0.11297035962343216, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297035962343216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11331339180469513, + 0.13431310653686523, + 0.04533388465642929, + 0.09907016158103943, + 0.08677323907613754, + 0.0879635289311409, + 0.034402888268232346, + 0.021681251004338264, + 0.04479547590017319, + 0.3323530852794647 + ], + "credit_map": [ + 0.012405936606228352, + 0.010281034745275974, + 0.012647733092308044, + 0.015003947541117668, + 0.005938354879617691, + 0.02082611247897148, + 0.010631607845425606, + 0.02557864598929882, + 0.011074273847043514, + 0.004954058211296797, + 0.006585093680769205, + 0.009408910758793354, + 0.004148965701460838, + 0.047278571873903275, + 0.005735787563025951, + 0.04512743651866913, + 0.003532394766807556, + 0.002521688351407647, + 0.004451061598956585, + 0.004019911400973797, + 0.004431599751114845, + 0.009072969667613506, + 0.005469722207635641, + 0.011834532022476196, + 0.0069155264645814896, + 0.011321593075990677, + 0.007356944028288126, + 0.01582012139260769, + 0.006922312080860138, + 0.021640902385115623, + 0.011614225804805756, + 0.01747853122651577, + 0.0024876422248780727, + 0.0031091475393623114, + 0.004516043234616518, + 0.006640758831053972, + 0.0026043392717838287, + 0.03358844295144081, + 0.005722787231206894, + 0.028104068711400032, + 0.007614582777023315, + 0.008535288274288177, + 0.010266918689012527, + 0.012318567372858524, + 0.009274656884372234, + 0.016057342290878296, + 0.011853118427097797, + 0.012043045833706856, + 0.0042874193750321865, + 0.0033489877823740244, + 0.004502328112721443, + 0.005984578747302294, + 0.003584812395274639, + 0.004865199327468872, + 0.0036779935471713543, + 0.00415156502276659, + 0.002071262104436755, + 0.00242661708034575, + 0.0030087835621088743, + 0.002823671093210578, + 0.0021522825118154287, + 0.003049671882763505, + 0.003718992695212364, + 0.002429969608783722, + 0.00495975511148572, + 0.005135145969688892, + 0.005442523397505283, + 0.007152126170694828, + 0.003739295294508338, + 0.005401699338108301, + 0.006693768780678511, + 0.006271157413721085, + 0.04963063821196556, + 0.03791424259543419, + 0.03151193633675575, + 0.048980463296175, + 0.038925155997276306, + 0.03429616242647171, + 0.05018964409828186, + 0.04090482369065285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.11297035962343216 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588964223861694 + ], + "opposite_logprob_tokens": [ + -3.8955888748168945 + ], + "logprob_reward": -0.020588964223861694, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588964223861694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09464564174413681, + 0.34212616086006165, + 0.08726990222930908, + 0.18145282566547394, + 0.03449124097824097, + 0.05625012889504433, + 0.027953386306762695, + 0.045958153903484344, + 0.036242228001356125, + 0.09361036866903305 + ], + "credit_map": [ + 0.01019586157053709, + 0.005798621103167534, + 0.01033193152397871, + 0.008561667986214161, + 0.007349680177867413, + 0.01656137965619564, + 0.006484710145741701, + 0.029361790046095848, + 0.02693021669983864, + 0.010368108749389648, + 0.018584227189421654, + 0.013192139565944672, + 0.012670210562646389, + 0.12036225199699402, + 0.015846479684114456, + 0.12417253851890564, + 0.008585760369896889, + 0.005125627387315035, + 0.012756707146763802, + 0.0074926018714904785, + 0.006953401956707239, + 0.013931039720773697, + 0.010413529351353645, + 0.022011231631040573, + 0.005549662280827761, + 0.0041285003535449505, + 0.0068314578384160995, + 0.007044846657663584, + 0.004573008511215448, + 0.07468146830797195, + 0.009565506130456924, + 0.06907838582992554, + 0.0021496217232197523, + 0.0019398233853280544, + 0.0028470263350754976, + 0.0032024725805968046, + 0.0033989017829298973, + 0.008382122963666916, + 0.00461822422221303, + 0.00795304961502552, + 0.0027208488900214434, + 0.0020864149555563927, + 0.006976759061217308, + 0.004379467107355595, + 0.003189637791365385, + 0.01672075130045414, + 0.005135390441864729, + 0.015040857717394829, + 0.0021637622267007828, + 0.0021023969165980816, + 0.005739363841712475, + 0.003764413297176361, + 0.0017308061942458153, + 0.004615094047039747, + 0.0033740014769136906, + 0.004463547840714455, + 0.004553667269647121, + 0.004045053385198116, + 0.0072362846694886684, + 0.006148783955723047, + 0.0041141207329928875, + 0.007691704668104649, + 0.005390553269535303, + 0.006777988281100988, + 0.003570152446627617, + 0.0025541153736412525, + 0.006466712802648544, + 0.004321185406297445, + 0.0028570243157446384, + 0.006817747373133898, + 0.0050310431979596615, + 0.00462424848228693, + 0.01321520283818245, + 0.010908136144280434, + 0.008626319468021393, + 0.013699675910174847, + 0.012766563333570957, + 0.00865218136459589, + 0.014650390483438969, + 0.011091894470155239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.020588964223861694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38708409667015076 + ], + "opposite_logprob_tokens": [ + -1.1370841264724731 + ], + "logprob_reward": -0.38708409667015076, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38708409667015076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.209548681974411, + 0.22463804483413696, + 0.04629022255539894, + 0.10365355014801025, + 0.11566666513681412, + 0.053743209689855576, + 0.03382362425327301, + 0.03301301598548889, + 0.07460660487413406, + 0.105016328394413 + ], + "credit_map": [ + 0.02420060895383358, + 0.019292933866381645, + 0.03170354291796684, + 0.029933178797364235, + 0.010397491976618767, + 0.026702040806412697, + 0.02033942937850952, + 0.04697944596409798, + 0.023410625755786896, + 0.011195937171578407, + 0.013187810778617859, + 0.016910074278712273, + 0.007815353572368622, + 0.076743945479393, + 0.012780836783349514, + 0.06259346753358841, + 0.004614588804543018, + 0.002936829812824726, + 0.005782801192253828, + 0.004526726435869932, + 0.004928270820528269, + 0.009030339308083057, + 0.006030476186424494, + 0.008440189994871616, + 0.005288869608193636, + 0.00736671919003129, + 0.006696502212435007, + 0.00796909537166357, + 0.004448528401553631, + 0.03347500413656235, + 0.009100040420889854, + 0.029308795928955078, + 0.005237945355474949, + 0.005064455792307854, + 0.008445038460195065, + 0.010618465021252632, + 0.004847890231758356, + 0.03587537631392479, + 0.017915209755301476, + 0.027662288397550583, + 0.005098174791783094, + 0.0049628326669335365, + 0.007376798894256353, + 0.00840748567134142, + 0.005322099197655916, + 0.00604707095772028, + 0.009238684549927711, + 0.0072900657542049885, + 0.0019400444580242038, + 0.002703192876651883, + 0.006678728386759758, + 0.004940950311720371, + 0.0019233634229749441, + 0.005245818756520748, + 0.005273898597806692, + 0.00511762872338295, + 0.0029429250862449408, + 0.0035077896900475025, + 0.004275020211935043, + 0.005304879974573851, + 0.002901141531765461, + 0.004740170203149319, + 0.005364880431443453, + 0.003976209554821253, + 0.009946119040250778, + 0.007909780368208885, + 0.008626212365925312, + 0.010321740992367268, + 0.00838462170213461, + 0.011967782862484455, + 0.008351479656994343, + 0.009098868817090988, + 0.014144951477646828, + 0.010302036069333553, + 0.009039989672601223, + 0.01738435961306095, + 0.009171877056360245, + 0.012101664207875729, + 0.01893998496234417, + 0.013931467197835445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.38708409667015076 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557969391345978 + ], + "opposite_logprob_tokens": [ + -1.6005797386169434 + ], + "logprob_reward": -0.22557969391345978, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557969391345978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10316949337720871, + 0.22819596529006958, + 0.03681422397494316, + 0.11724460124969482, + 0.1964898556470871, + 0.0486772283911705, + 0.038615040481090546, + 0.04687792807817459, + 0.0582234263420105, + 0.12569215893745422 + ], + "credit_map": [ + 0.012380740605294704, + 0.01223079115152359, + 0.01030851062387228, + 0.01968234032392502, + 0.005648121703416109, + 0.010874870233237743, + 0.012133881449699402, + 0.019910244271159172, + 0.01999005302786827, + 0.004591554403305054, + 0.008785166777670383, + 0.009977521374821663, + 0.006621141918003559, + 0.08925682306289673, + 0.007646909449249506, + 0.08132682740688324, + 0.0030249357223510742, + 0.002876955084502697, + 0.003555059665814042, + 0.004732443951070309, + 0.00507199065759778, + 0.004773333203047514, + 0.006858959794044495, + 0.005920551251620054, + 0.005054831504821777, + 0.0040949368849396706, + 0.00511223403736949, + 0.006609887816011906, + 0.0058579049073159695, + 0.043443795293569565, + 0.007329389452934265, + 0.0397416315972805, + 0.005724843125790358, + 0.00736826378852129, + 0.010046657174825668, + 0.011263567954301834, + 0.007715360727161169, + 0.07558506727218628, + 0.012924150563776493, + 0.06586196273565292, + 0.0040300022810697556, + 0.004415147937834263, + 0.005936114117503166, + 0.007890915498137474, + 0.00505808275192976, + 0.008623046800494194, + 0.004836295731365681, + 0.007887626998126507, + 0.0045382799580693245, + 0.00476827984675765, + 0.004407739732414484, + 0.006066374015063047, + 0.004592052195221186, + 0.0050168950110673904, + 0.004157152492552996, + 0.005068267695605755, + 0.0045653050765395164, + 0.005003287456929684, + 0.005150464829057455, + 0.0062581985257565975, + 0.00449404725804925, + 0.007786171976476908, + 0.007540350314229727, + 0.006080108229070902, + 0.0073827519081532955, + 0.007382475305348635, + 0.0062810443341732025, + 0.008357550948858261, + 0.006144220009446144, + 0.006006608717143536, + 0.007308844942599535, + 0.009359930641949177, + 0.012088004499673843, + 0.012412533164024353, + 0.009512331336736679, + 0.020541122183203697, + 0.014377475716173649, + 0.018113063648343086, + 0.02208191156387329, + 0.01656574383378029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.22557969391345978 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, shot at a high frame rate to create a breathtaking super-slow-motion effect, focuses on a crimson rose bud, about to bloom. In the soft, diffused, cool light of early morning, each velvety rose petal is clung to countless glistening, plump beads of morning dew. The background, a completely blurred, hazy emerald green, underscores the rose's aloofness and purity. As the flower is lifted upwards by an invisible force, the camera tilts slightly upward, capturing the subtle effects of gravity: The spherical dewdrops begin to lose their grip, stretching across the petal's smooth surface, forming tiny streaks before gracefully peeling off the edge, one by one, and sliding off. Captured in super-slow motion, each droplet, in its moment of descent, resembles a sparkling diamond, its surface shimmering, its form subtly shifting in mid-air before vanishing beyond the depth of field. The entire process is imbued with tranquility, poetry, and a fleeting, exquisite beauty.", + "seed": 11365, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_14/index_8471_seed_11365_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09554756917059422, + "vlm_logprob": -0.09554756917059422, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.09554756917059422, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016150010749697685 + ], + "opposite_logprob_tokens": [ + -4.266149997711182 + ], + "logprob_reward": -0.016150010749697685, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016150010749697685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11290482431650162, + 0.07528013736009598, + 0.042201872915029526, + 0.047055989503860474, + 0.19744224846363068, + 0.06377243995666504, + 0.09512794762849808, + 0.09407758712768555, + 0.17646288871765137, + 0.09567411988973618 + ], + "credit_map": [ + 0.009623445570468903, + 0.01052592508494854, + 0.010687362402677536, + 0.016484472900629044, + 0.011629458516836166, + 0.01417083665728569, + 0.019065722823143005, + 0.020717598497867584, + 0.006331979297101498, + 0.009723427705466747, + 0.00671151839196682, + 0.011050472036004066, + 0.00819072313606739, + 0.012547300197184086, + 0.008287986740469933, + 0.012436729855835438, + 0.004343239590525627, + 0.00505167106166482, + 0.0057714348658919334, + 0.005352206062525511, + 0.005104127340018749, + 0.006105573382228613, + 0.005335587076842785, + 0.005138034000992775, + 0.0053175706416368484, + 0.004319938365370035, + 0.005603461526334286, + 0.005989315453916788, + 0.004789883736521006, + 0.005418573971837759, + 0.008816826157271862, + 0.0068004196509718895, + 0.017091847956180573, + 0.03328351303935051, + 0.016746846958994865, + 0.020825153216719627, + 0.02502194233238697, + 0.024999570101499557, + 0.027330299839377403, + 0.03214306756854057, + 0.004526420496404171, + 0.006251232232898474, + 0.006652244366705418, + 0.00869152881205082, + 0.007080207113176584, + 0.007131424732506275, + 0.013148361817002296, + 0.01029102224856615, + 0.014387598261237144, + 0.019189754500985146, + 0.010973095893859863, + 0.009313940070569515, + 0.00974943209439516, + 0.010879754088819027, + 0.008580283261835575, + 0.012054086662828922, + 0.01112331636250019, + 0.017407633364200592, + 0.013326511718332767, + 0.010559938848018646, + 0.006884723901748657, + 0.0077306609600782394, + 0.014756730757653713, + 0.012288068421185017, + 0.02278575301170349, + 0.01751740090548992, + 0.022872911766171455, + 0.0200872290879488, + 0.017591726034879684, + 0.0322217233479023, + 0.016390053555369377, + 0.026996087282896042, + 0.018107958137989044, + 0.008786317892372608, + 0.01814916916191578, + 0.010458270087838173, + 0.006335551850497723, + 0.006630305200815201, + 0.014506191946566105, + 0.012700358405709267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.016150010749697685 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007646458223462105 + ], + "opposite_logprob_tokens": [ + -4.882646560668945 + ], + "logprob_reward": -0.007646458223462105, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646458223462105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14221630990505219, + 0.03541577234864235, + 0.03388156369328499, + 0.12765733897686005, + 0.05150832608342171, + 0.07303963601589203, + 0.027472879737615585, + 0.1155787780880928, + 0.2955988049507141, + 0.09763059765100479 + ], + "credit_map": [ + 0.013604390434920788, + 0.011092619970440865, + 0.01661258563399315, + 0.015365619212388992, + 0.01626921072602272, + 0.016058029606938362, + 0.026803355664014816, + 0.026410505175590515, + 0.0025567745324224234, + 0.003051700768992305, + 0.0034320920240134, + 0.00610730517655611, + 0.004522704519331455, + 0.004841810092329979, + 0.005535016302019358, + 0.005368370097130537, + 0.004455603659152985, + 0.004214114975184202, + 0.003556877840310335, + 0.004808684345334768, + 0.003607942955568433, + 0.00404326431453228, + 0.004880603868514299, + 0.004314471501857042, + 0.011594077572226524, + 0.00963748898357153, + 0.017518840730190277, + 0.012827746570110321, + 0.01387234777212143, + 0.015823720023036003, + 0.02959567867219448, + 0.016787420958280563, + 0.004560533910989761, + 0.005220956169068813, + 0.006920468993484974, + 0.008320653811097145, + 0.0060872165486216545, + 0.005638903938233852, + 0.006574653089046478, + 0.008184942416846752, + 0.005568636115640402, + 0.008242611773312092, + 0.00736379437148571, + 0.008360929787158966, + 0.012511600740253925, + 0.007033186964690685, + 0.012729942798614502, + 0.011228936724364758, + 0.002925344044342637, + 0.0034662482794374228, + 0.0037731833290308714, + 0.0032910204026848078, + 0.003842320293188095, + 0.0030375863425433636, + 0.004046855494379997, + 0.0030903220176696777, + 0.009782589040696621, + 0.010744154453277588, + 0.016696274280548096, + 0.015653569251298904, + 0.008687404915690422, + 0.01216983050107956, + 0.026938410475850105, + 0.014906544238328934, + 0.01922067254781723, + 0.024585379287600517, + 0.04638947546482086, + 0.03126896917819977, + 0.02420940063893795, + 0.042534299194812775, + 0.06904500722885132, + 0.03834560140967369, + 0.00878397561609745, + 0.00802554190158844, + 0.010337689891457558, + 0.015837274491786957, + 0.00907621905207634, + 0.010277358815073967, + 0.021937092766165733, + 0.013355453498661518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.007646458223462105 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347355917096138 + ], + "opposite_logprob_tokens": [ + -3.6513473987579346 + ], + "logprob_reward": -0.026347355917096138, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347355917096138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2231736034154892, + 0.3193289637565613, + 0.04574275016784668, + 0.06926649063825607, + 0.06010717153549194, + 0.07528191059827805, + 0.03468237444758415, + 0.03412901237607002, + 0.05919541418552399, + 0.07909227907657623 + ], + "credit_map": [ + 0.018199708312749863, + 0.015574868768453598, + 0.025897260755300522, + 0.022404422983527184, + 0.027028808370232582, + 0.030713750049471855, + 0.047014154493808746, + 0.03634063899517059, + 0.01991717703640461, + 0.022795481607317924, + 0.026384184136986732, + 0.03200644627213478, + 0.032847873866558075, + 0.08867428451776505, + 0.037934523075819016, + 0.058768972754478455, + 0.0039898185059428215, + 0.003976397216320038, + 0.006375482771545649, + 0.006669003050774336, + 0.00527826277539134, + 0.004751002881675959, + 0.00854191929101944, + 0.006160864606499672, + 0.005716071464121342, + 0.005954270716756582, + 0.009539690800011158, + 0.006937683559954166, + 0.007193253841251135, + 0.008448736742138863, + 0.016657739877700806, + 0.008819038979709148, + 0.005485302768647671, + 0.0107493307441473, + 0.005747322924435139, + 0.005903494078665972, + 0.008701132610440254, + 0.006390724331140518, + 0.007655239198356867, + 0.009474621154367924, + 0.006336865946650505, + 0.009457073174417019, + 0.010506900027394295, + 0.012345952913165092, + 0.009513546712696552, + 0.006598369684070349, + 0.011334524489939213, + 0.00918868463486433, + 0.0033960361033678055, + 0.0073155625723302364, + 0.003995727282017469, + 0.005021965131163597, + 0.0033959627617150545, + 0.003001031931489706, + 0.004209588747471571, + 0.0043465024791657925, + 0.0027369181625545025, + 0.003467149566859007, + 0.00473835738375783, + 0.00391598604619503, + 0.002775834407657385, + 0.0034105409868061543, + 0.008068651892244816, + 0.005015572067350149, + 0.005700758192688227, + 0.003902882570400834, + 0.010275737382471561, + 0.0068575055338442326, + 0.004014491103589535, + 0.006000044755637646, + 0.01457746047526598, + 0.007866536267101765, + 0.009773516096174717, + 0.007502045948058367, + 0.008597280830144882, + 0.011061028577387333, + 0.009717769920825958, + 0.007453572005033493, + 0.015297732315957546, + 0.009689333848655224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.026347355917096138 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34863486886024475 + ], + "opposite_logprob_tokens": [ + -1.2236348390579224 + ], + "logprob_reward": -0.34863486886024475, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34863486886024475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22657829523086548, + 0.07026504725217819, + 0.04445018246769905, + 0.056128960102796555, + 0.048988524824380875, + 0.13212116062641144, + 0.10892604291439056, + 0.08472802489995956, + 0.15962280333042145, + 0.06819089502096176 + ], + "credit_map": [ + 0.02044215425848961, + 0.021399803459644318, + 0.03489656001329422, + 0.023533590137958527, + 0.027525147423148155, + 0.02816496230661869, + 0.03279808908700943, + 0.037817999720573425, + 0.007211139891296625, + 0.008805842138826847, + 0.006640253588557243, + 0.007663704454898834, + 0.008337217383086681, + 0.013323348015546799, + 0.007124427240341902, + 0.011159117333590984, + 0.004363290965557098, + 0.007931777276098728, + 0.004495889879763126, + 0.003991389647126198, + 0.005743054673075676, + 0.00747101241722703, + 0.004684582818299532, + 0.005769188981503248, + 0.005071359686553478, + 0.007951571606099606, + 0.007093065418303013, + 0.007747269235551357, + 0.007805456407368183, + 0.007527163252234459, + 0.0060916561633348465, + 0.0068414220586419106, + 0.0064737615175545216, + 0.005768164526671171, + 0.004932784475386143, + 0.006478423718363047, + 0.005090297199785709, + 0.005895739886909723, + 0.006627601105719805, + 0.007721755653619766, + 0.010041403584182262, + 0.01709800958633423, + 0.009180211462080479, + 0.015477847307920456, + 0.025064820423722267, + 0.013820902444422245, + 0.02078987844288349, + 0.020648101344704628, + 0.009741261601448059, + 0.021779516711831093, + 0.011713945306837559, + 0.014990135096013546, + 0.012426287867128849, + 0.010842271149158478, + 0.013331831432878971, + 0.014100804924964905, + 0.007956201210618019, + 0.010350321419537067, + 0.009800498373806477, + 0.008807376027107239, + 0.00941886194050312, + 0.008023975417017937, + 0.019129132851958275, + 0.011241664178669453, + 0.021894583478569984, + 0.01890552043914795, + 0.02570246532559395, + 0.01938447169959545, + 0.010631760582327843, + 0.01771240308880806, + 0.024190442636609077, + 0.02120117098093033, + 0.014161958359181881, + 0.007672651205211878, + 0.010203326120972633, + 0.006482330616563559, + 0.004445409867912531, + 0.004490833263844252, + 0.012118355371057987, + 0.008616033010184765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.34863486886024475 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789591521024704 + ], + "opposite_logprob_tokens": [ + -2.5789592266082764 + ], + "logprob_reward": -0.0789591521024704, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789591521024704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16241924464702606, + 0.05581140145659447, + 0.05204698070883751, + 0.042417291551828384, + 0.08841047435998917, + 0.1348084658384323, + 0.09303291141986847, + 0.0770253911614418, + 0.1679837554693222, + 0.12604409456253052 + ], + "credit_map": [ + 0.01260376162827015, + 0.008757988922297955, + 0.025473173707723618, + 0.014275443740189075, + 0.018617838621139526, + 0.026098843663930893, + 0.031166188418865204, + 0.025426005944609642, + 0.004835009109228849, + 0.006371968425810337, + 0.00669061578810215, + 0.010133574716746807, + 0.007627754472196102, + 0.005198365077376366, + 0.006270395591855049, + 0.008683721534907818, + 0.005040150601416826, + 0.007929345592856407, + 0.004892535973340273, + 0.007580661680549383, + 0.006556904874742031, + 0.005156766157597303, + 0.008612941019237041, + 0.0062776776030659676, + 0.003880253527313471, + 0.004604242742061615, + 0.0028354141395539045, + 0.005951197352260351, + 0.004839450586587191, + 0.004735616967082024, + 0.009527984075248241, + 0.006043132394552231, + 0.0073132747784256935, + 0.01298227347433567, + 0.006792663596570492, + 0.01195827592164278, + 0.014055836014449596, + 0.010601529851555824, + 0.01084906142205, + 0.013857563026249409, + 0.010709918104112148, + 0.016055496409535408, + 0.015164271928369999, + 0.015660466626286507, + 0.01969439908862114, + 0.013919033110141754, + 0.023436937481164932, + 0.020167941227555275, + 0.011750508099794388, + 0.013965990394353867, + 0.014851454645395279, + 0.00934967864304781, + 0.011960220523178577, + 0.010384109802544117, + 0.0104741919785738, + 0.01029675081372261, + 0.006192113738507032, + 0.00901112426072359, + 0.009957489557564259, + 0.009400289505720139, + 0.006570607889443636, + 0.0068159401416778564, + 0.017467636615037918, + 0.011610185727477074, + 0.010794822126626968, + 0.015406929887831211, + 0.020134517922997475, + 0.019623037427663803, + 0.012980727478861809, + 0.0228017196059227, + 0.04376561939716339, + 0.022476378828287125, + 0.015718374401330948, + 0.014495491981506348, + 0.013452094048261642, + 0.018800457939505577, + 0.01570570468902588, + 0.0075613087974488735, + 0.025179138407111168, + 0.015131520107388496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.0789591521024704 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, shot at a high frame rate to create a breathtaking super-slow-motion effect, focuses on a crimson rose bud, about to bloom. In the soft, diffused, cool light of early morning, each velvety rose petal is clung to countless glistening, plump beads of morning dew. The background, a completely blurred, hazy emerald green, underscores the rose's aloofness and purity. As the flower is lifted upwards by an invisible force, the camera tilts slightly upward, capturing the subtle effects of gravity: The spherical dewdrops begin to lose their grip, stretching across the petal's smooth surface, forming tiny streaks before gracefully peeling off the edge, one by one, and sliding off. Captured in super-slow motion, each droplet, in its moment of descent, resembles a sparkling diamond, its surface shimmering, its form subtly shifting in mid-air before vanishing beyond the depth of field. The entire process is imbued with tranquility, poetry, and a fleeting, exquisite beauty.", + "seed": 11366, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_14/index_8471_seed_11366_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06643389724195004, + "vlm_logprob": -0.06643389724195004, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.06643389724195004, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04129825904965401 + ], + "opposite_logprob_tokens": [ + -3.2912981510162354 + ], + "logprob_reward": -0.04129825904965401, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04129825904965401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023331111297011375, + 0.012871799059212208, + 0.0027733859606087208, + 0.004655209369957447, + 0.016737092286348343, + 0.010407743975520134, + 0.002060485305264592, + 0.00897886324673891, + 0.0013615773059427738, + 0.9168227314949036 + ], + "credit_map": [ + 0.002813153201714158, + 0.0036015845835208893, + 0.0038536156062036753, + 0.0023256491404026747, + 0.0023018072824925184, + 0.0023670385126024485, + 0.0018034226959571242, + 0.004264840390533209, + 0.001156143145635724, + 0.000564814661629498, + 0.004625496454536915, + 0.0008511756313964725, + 0.0007534968899562955, + 0.0010535031324252486, + 0.0009214914171025157, + 0.002945677610114217, + 0.0002446061698719859, + 0.00039105830364860594, + 0.00047571322647854686, + 0.00028062050114385784, + 0.0002814649196807295, + 0.00034229832817800343, + 0.00030683312797918916, + 0.00045079117990098894, + 0.0005898564122617245, + 0.0005943096475675702, + 0.0006115628639236093, + 0.000539194094017148, + 0.0005395664484240115, + 0.0005717018502764404, + 0.0006607283721677959, + 0.0005482895066961646, + 0.0018699131906032562, + 0.0030348487198352814, + 0.0019220240646973252, + 0.002175770001485944, + 0.0020242291502654552, + 0.001681874506175518, + 0.002037354512140155, + 0.0019910784903913736, + 0.0010954716708511114, + 0.0012977164005860686, + 0.001667800359427929, + 0.0010178214870393276, + 0.0010652468772605062, + 0.001572472508996725, + 0.0016072331927716732, + 0.0010839817114174366, + 0.00014832279703114182, + 0.0005244540516287088, + 0.00024129418306984007, + 0.00020733314158860594, + 0.00019395130220800638, + 0.00016119949577841908, + 0.0002008187584578991, + 0.0003831115027423948, + 0.0006449762731790543, + 0.0011678134324029088, + 0.001578320749104023, + 0.0010736332042142749, + 0.0008270653197541833, + 0.000991741893813014, + 0.001273105968721211, + 0.0014222071040421724, + 0.0001330184459220618, + 0.0001698785781627521, + 0.0001941681548487395, + 0.00018570577958598733, + 0.00017027364810928702, + 0.00020047120051458478, + 0.00014938157983124256, + 0.00015867993352003396, + 0.02659117616713047, + 0.3066757619380951, + 0.07263799011707306, + 0.3376723825931549, + 0.07980886846780777, + 0.029262922704219818, + 0.04494597390294075, + 0.019227664917707443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.04129825904965401 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372080624103546 + ], + "opposite_logprob_tokens": [ + -3.4087207317352295 + ], + "logprob_reward": -0.03372080624103546, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372080624103546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013396179303526878, + 0.01092047244310379, + 0.0030524772591888905, + 0.004031742922961712, + 0.008945755660533905, + 0.0029254627879709005, + 0.004233677405864, + 0.013527920469641685, + 0.0011913689086213708, + 0.9377749562263489 + ], + "credit_map": [ + 0.001542797195725143, + 0.0019390106899663806, + 0.0021506822668015957, + 0.0016100105131044984, + 0.0017507723532617092, + 0.0013956428738310933, + 0.0014275378780439496, + 0.0015797242522239685, + 0.0009584326180629432, + 0.0006071343086659908, + 0.0038439598865807056, + 0.0007117518107406795, + 0.0008355612517334521, + 0.0009517019498161972, + 0.0007478494662791491, + 0.002264081733301282, + 0.00037322513526305556, + 0.00036656728480011225, + 0.0006062729516997933, + 0.0003795111842919141, + 0.00030342835816554725, + 0.0003460520238149911, + 0.00026193889789283276, + 0.000415481481468305, + 0.00044936995254829526, + 0.00037664989940822124, + 0.0006209574639797211, + 0.00041807210072875023, + 0.0003091649559792131, + 0.0005149318021722138, + 0.0008866806165315211, + 0.00045591621892526746, + 0.0009818265680223703, + 0.0015803334536030889, + 0.0010307814227417111, + 0.0011789402924478054, + 0.0009820344857871532, + 0.0009136107983067632, + 0.0011520721018314362, + 0.0011261560721322894, + 0.0003603835357353091, + 0.0003487821959424764, + 0.00041784823406487703, + 0.0002780788345262408, + 0.00045604139450006187, + 0.0002916488447226584, + 0.00045324378879740834, + 0.0003194360760971904, + 0.000312458403641358, + 0.001122700166888535, + 0.00037062683259136975, + 0.0003982703201472759, + 0.00046199955977499485, + 0.0003637121117208153, + 0.00036757957423105836, + 0.0008363303495571017, + 0.0010201508412137628, + 0.0016400187741965055, + 0.0025510082487016916, + 0.0015384119469672441, + 0.0011106343008577824, + 0.0014584185555577278, + 0.0018791290931403637, + 0.0023301485925912857, + 0.00011049954628106207, + 0.0001336545537924394, + 0.00018463717424310744, + 0.00014597333210986108, + 9.88463798421435e-05, + 0.00016591719759162515, + 0.00013468333054333925, + 0.00021715747425332665, + 0.027153024449944496, + 0.31372198462486267, + 0.07432740181684494, + 0.3452989459037781, + 0.08175762742757797, + 0.02996225841343403, + 0.04590582102537155, + 0.019647862762212753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.03372080624103546 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803412035107613 + ], + "opposite_logprob_tokens": [ + -3.529803514480591 + ], + "logprob_reward": -0.029803412035107613, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803412035107613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09863980859518051, + 0.1971776932477951, + 0.053435925394296646, + 0.038555413484573364, + 0.0955730751156807, + 0.02406102791428566, + 0.04590809717774391, + 0.03107963502407074, + 0.035202015191316605, + 0.38036730885505676 + ], + "credit_map": [ + 0.01037479005753994, + 0.017980046570301056, + 0.008485653437674046, + 0.009166926145553589, + 0.016209905967116356, + 0.012726940214633942, + 0.010528531856834888, + 0.013167016208171844, + 0.01792030781507492, + 0.01684805005788803, + 0.03017856739461422, + 0.022373024374246597, + 0.01692698337137699, + 0.019653113558888435, + 0.04567591845989227, + 0.027601731941103935, + 0.005763211753219366, + 0.00630714138969779, + 0.008473393507301807, + 0.006137573625892401, + 0.006709577050060034, + 0.008582722395658493, + 0.004849447403103113, + 0.006612855009734631, + 0.004146274644881487, + 0.005133859347552061, + 0.0052421786822378635, + 0.004096310120075941, + 0.004512114450335503, + 0.004064541310071945, + 0.005121159367263317, + 0.006238976493477821, + 0.009463404305279255, + 0.015535585582256317, + 0.012202037498354912, + 0.012446356937289238, + 0.009689987637102604, + 0.008173519745469093, + 0.01598871313035488, + 0.01207345724105835, + 0.0028675859794020653, + 0.0030751076992601156, + 0.004084086511284113, + 0.0023646375630050898, + 0.002278608502820134, + 0.003638454247266054, + 0.003178015584126115, + 0.0025745329912751913, + 0.0030649863183498383, + 0.015176566317677498, + 0.0024351179599761963, + 0.003702653804793954, + 0.004337318707257509, + 0.003089534817263484, + 0.0029541607946157455, + 0.011147754266858101, + 0.002533173654228449, + 0.003696053521707654, + 0.00601242994889617, + 0.004469042178243399, + 0.003174898447468877, + 0.002820938127115369, + 0.004722881596535444, + 0.0036502189468592405, + 0.002938247751444578, + 0.003279265481978655, + 0.004741869401186705, + 0.00581124285236001, + 0.0026686114724725485, + 0.0031905651558190584, + 0.005158058367669582, + 0.007414155639708042, + 0.011485626921057701, + 0.12723997235298157, + 0.030000966042280197, + 0.13916005194187164, + 0.033309388905763626, + 0.012174502946436405, + 0.018593020737171173, + 0.008403778076171875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.029803412035107613 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.127041757106781 + ], + "opposite_logprob_tokens": [ + -2.127041816711426 + ], + "logprob_reward": -0.127041757106781, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.127041757106781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0563923642039299, + 0.03438267484307289, + 0.06044652312994003, + 0.08710155636072159, + 0.06053532287478447, + 0.00805885624140501, + 0.00798068754374981, + 0.008652669377624989, + 0.010928729549050331, + 0.6655206680297852 + ], + "credit_map": [ + 0.006263989955186844, + 0.008763616904616356, + 0.006876977626234293, + 0.005880061537027359, + 0.006486440077424049, + 0.00795029941946268, + 0.0056522199884057045, + 0.008518755435943604, + 0.0036990344524383545, + 0.003052378073334694, + 0.006358901970088482, + 0.003489225171506405, + 0.0030041970312595367, + 0.0037614915054291487, + 0.00550009636208415, + 0.005517346318811178, + 0.007088296115398407, + 0.006527021061629057, + 0.00903445202857256, + 0.008494324050843716, + 0.0066924067214131355, + 0.008122279308736324, + 0.007627463433891535, + 0.006860276684165001, + 0.009578066878020763, + 0.008520277217030525, + 0.014140989631414413, + 0.008240927010774612, + 0.008280962705612183, + 0.00985900592058897, + 0.016068309545516968, + 0.01241301465779543, + 0.007083781063556671, + 0.01119597814977169, + 0.006017068400979042, + 0.00786063726991415, + 0.007428439799696207, + 0.006656246725469828, + 0.0073207663372159, + 0.006972397677600384, + 0.0011197339044883847, + 0.0013281007995828986, + 0.0012193038128316402, + 0.0008133449009619653, + 0.0009908706415444613, + 0.0007083293166942894, + 0.0011256086872890592, + 0.0007535639451816678, + 0.0007885306840762496, + 0.000888122885953635, + 0.0009698311914689839, + 0.0012031305814161897, + 0.0011045215651392937, + 0.001091552316211164, + 0.000968407082837075, + 0.0009665904799476266, + 0.0006802116404287517, + 0.0009386084857396781, + 0.0017010134179145098, + 0.0013403608463704586, + 0.0007996851927600801, + 0.0009042476303875446, + 0.0011195515980944037, + 0.0011689900420606136, + 0.0010617055231705308, + 0.0014160701539367437, + 0.00139232084620744, + 0.0027091328520327806, + 0.001228615641593933, + 0.0008389960275962949, + 0.0012086898786947131, + 0.0010731989750638604, + 0.01921217516064644, + 0.2225930243730545, + 0.052678048610687256, + 0.24529659748077393, + 0.05773768946528435, + 0.02135518379509449, + 0.03277973085641861, + 0.013868163339793682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.127041757106781 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030525177717209 + ], + "opposite_logprob_tokens": [ + -2.3503053188323975 + ], + "logprob_reward": -0.10030525177717209, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030525177717209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03353049233555794, + 0.015236147679388523, + 0.01487776730209589, + 0.02410058118402958, + 0.035526342689991, + 0.01027388870716095, + 0.01175695937126875, + 0.022001231089234352, + 0.0038279739674180746, + 0.8288686275482178 + ], + "credit_map": [ + 0.004705315921455622, + 0.004161129705607891, + 0.003985412418842316, + 0.003223385661840439, + 0.004141525365412235, + 0.004506255965679884, + 0.004533665254712105, + 0.004273795522749424, + 0.0014105165610089898, + 0.0012329998426139355, + 0.004395422525703907, + 0.001457023317925632, + 0.001206501037813723, + 0.0013195358915254474, + 0.0018899498973041773, + 0.002324196510016918, + 0.0017094191862270236, + 0.00177405821159482, + 0.002733878791332245, + 0.0019454184221103787, + 0.0019456661539152265, + 0.0017327736131846905, + 0.0012445468455553055, + 0.0017920045647770166, + 0.0022886493243277073, + 0.002126314677298069, + 0.0039132144302129745, + 0.0019699407275766134, + 0.001802685670554638, + 0.0038752120453864336, + 0.004286015406250954, + 0.0038385489024221897, + 0.003834210801869631, + 0.006182665005326271, + 0.004399199970066547, + 0.004723968915641308, + 0.003951400052756071, + 0.0032077047508209944, + 0.005003370810300112, + 0.004223817493766546, + 0.0010683841537684202, + 0.0011658682487905025, + 0.002072705887258053, + 0.0009374422952532768, + 0.0009056201670318842, + 0.001595748937688768, + 0.0014492310583591461, + 0.0010788867948576808, + 0.0007575445924885571, + 0.003209118964150548, + 0.0009416777756996453, + 0.0012023607268929482, + 0.0011460359673947096, + 0.0009280579979531467, + 0.0009656754555180669, + 0.0026064866688102484, + 0.001739888801239431, + 0.002726266859099269, + 0.004258513450622559, + 0.0026624712627381086, + 0.0018573412671685219, + 0.00202076300047338, + 0.003353586420416832, + 0.00338239804841578, + 0.00039711559657007456, + 0.00044949815492145717, + 0.00045980833237990737, + 0.000539556669536978, + 0.0004105469270143658, + 0.0006937366561032832, + 0.00037331494968384504, + 0.0005043962737545371, + 0.024005427956581116, + 0.2773517668247223, + 0.06540767848491669, + 0.3055640757083893, + 0.0720827728509903, + 0.026521138846874237, + 0.04062897339463234, + 0.01730673760175705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.10030525177717209 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, shot at a high frame rate to create a breathtaking super-slow-motion effect, focuses on a crimson rose bud, about to bloom. In the soft, diffused, cool light of early morning, each velvety rose petal is clung to countless glistening, plump beads of morning dew. The background, a completely blurred, hazy emerald green, underscores the rose's aloofness and purity. As the flower is lifted upwards by an invisible force, the camera tilts slightly upward, capturing the subtle effects of gravity: The spherical dewdrops begin to lose their grip, stretching across the petal's smooth surface, forming tiny streaks before gracefully peeling off the edge, one by one, and sliding off. Captured in super-slow motion, each droplet, in its moment of descent, resembles a sparkling diamond, its surface shimmering, its form subtly shifting in mid-air before vanishing beyond the depth of field. The entire process is imbued with tranquility, poetry, and a fleeting, exquisite beauty.", + "seed": 11367, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_14/index_8471_seed_11367_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.022332653775811194, + "vlm_logprob": -0.022332653775811194, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.022332653775811194, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010133367963135242 + ], + "opposite_logprob_tokens": [ + -4.885133266448975 + ], + "logprob_reward": -0.010133367963135242, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010133367963135242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17623735964298248, + 0.08160091936588287, + 0.058914992958307266, + 0.07398072630167007, + 0.08332502841949463, + 0.08920987695455551, + 0.07113560289144516, + 0.2228662520647049, + 0.037659041583538055, + 0.10507021099328995 + ], + "credit_map": [ + 0.009630721062421799, + 0.012902293354272842, + 0.02195269614458084, + 0.01895907334983349, + 0.018398035317659378, + 0.026742715388536453, + 0.021192407235503197, + 0.04645942151546478, + 0.004268660210072994, + 0.008738967590034008, + 0.006552144419401884, + 0.014027168042957783, + 0.010722761042416096, + 0.013187466189265251, + 0.007707570213824511, + 0.016396183520555496, + 0.004511734005063772, + 0.007344614714384079, + 0.006449008360505104, + 0.007591825444251299, + 0.01009612251073122, + 0.009011417627334595, + 0.006489936262369156, + 0.007420331705361605, + 0.006944484543055296, + 0.00913270190358162, + 0.00693553127348423, + 0.00897545088082552, + 0.013393501751124859, + 0.01140007097274065, + 0.009207759983837605, + 0.00799122080206871, + 0.005316561553627253, + 0.011337772011756897, + 0.006411602720618248, + 0.015398349612951279, + 0.012225658632814884, + 0.007728214375674725, + 0.01541028544306755, + 0.00949658453464508, + 0.008023166097700596, + 0.005476717837154865, + 0.02072904258966446, + 0.00977044552564621, + 0.0035170232877135277, + 0.00796444434672594, + 0.009018301032483578, + 0.024710729718208313, + 0.008902251720428467, + 0.007612619549036026, + 0.006703285500407219, + 0.006471740081906319, + 0.009798264130949974, + 0.009542460553348064, + 0.008087921887636185, + 0.014017060399055481, + 0.027585912495851517, + 0.009770522825419903, + 0.016487162560224533, + 0.013124101795256138, + 0.04932490363717079, + 0.06160975992679596, + 0.012379038147628307, + 0.03258486092090607, + 0.005062888842076063, + 0.004048666916787624, + 0.003856563474982977, + 0.006626899819821119, + 0.005652590189129114, + 0.003962282557040453, + 0.0044866325333714485, + 0.003962513525038958, + 0.015032337978482246, + 0.011824428103864193, + 0.01627463288605213, + 0.007902012206614017, + 0.0278826542198658, + 0.01475463341921568, + 0.005255394149571657, + 0.006144107785075903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.010133367963135242 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059635089710354805 + ], + "opposite_logprob_tokens": [ + -5.130963325500488 + ], + "logprob_reward": -0.0059635089710354805, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059635089710354805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08869590610265732, + 0.2787964642047882, + 0.0519573874771595, + 0.05922506004571915, + 0.11358707398176193, + 0.0501614511013031, + 0.08006677776575089, + 0.08987109363079071, + 0.15212176740169525, + 0.03551701456308365 + ], + "credit_map": [ + 0.00658909697085619, + 0.01233852468430996, + 0.008622273802757263, + 0.014003429561853409, + 0.014144179411232471, + 0.012153989635407925, + 0.00858366023749113, + 0.012260760180652142, + 0.013811823911964893, + 0.03826165944337845, + 0.018553733825683594, + 0.039341483265161514, + 0.042025577276945114, + 0.04410334676504135, + 0.026995759457349777, + 0.055703114718198776, + 0.004547263961285353, + 0.0050603486597537994, + 0.00524959247559309, + 0.007147450000047684, + 0.004696827847510576, + 0.007445196155458689, + 0.010620085522532463, + 0.00719063077121973, + 0.005803341511636972, + 0.01016609650105238, + 0.005656289402395487, + 0.008211694657802582, + 0.007826398126780987, + 0.006273531820625067, + 0.009414316155016422, + 0.005873396992683411, + 0.007978708483278751, + 0.01618296466767788, + 0.009259777143597603, + 0.01943906769156456, + 0.014179526828229427, + 0.009316960349678993, + 0.023275479674339294, + 0.013954592868685722, + 0.0077759260311722755, + 0.004587042611092329, + 0.005899358540773392, + 0.007213751785457134, + 0.00393199548125267, + 0.005344316828995943, + 0.008709841407835484, + 0.006699220277369022, + 0.009519238956272602, + 0.009518704377114773, + 0.007684024050831795, + 0.007524530403316021, + 0.009682062081992626, + 0.008916500955820084, + 0.007279061246663332, + 0.01994265615940094, + 0.013279477134346962, + 0.006210509222000837, + 0.010493403300642967, + 0.010370314121246338, + 0.008085941895842552, + 0.011077629402279854, + 0.008623640052974224, + 0.021730178967118263, + 0.017440076917409897, + 0.014378201216459274, + 0.013291725888848305, + 0.020434699952602386, + 0.030759887769818306, + 0.01725115068256855, + 0.016869686543941498, + 0.02169634960591793, + 0.004607577808201313, + 0.00421102624386549, + 0.005695790983736515, + 0.0050459629856050014, + 0.006947633344680071, + 0.0034574607852846384, + 0.0029352386482059956, + 0.002616327255964279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 4 + ], + "selected_reward": -0.0059635089710354805 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058335766196251 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058335766196251, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058335766196251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37906643748283386, + 0.15123164653778076, + 0.052822671830654144, + 0.0409749299287796, + 0.05908771976828575, + 0.036526430398225784, + 0.03772682324051857, + 0.1209632083773613, + 0.08215232938528061, + 0.03944779187440872 + ], + "credit_map": [ + 0.01722879707813263, + 0.0322333499789238, + 0.037899333983659744, + 0.04051725193858147, + 0.04843618720769882, + 0.062172167003154755, + 0.03582151234149933, + 0.10475784540176392, + 0.007918558083474636, + 0.014191006310284138, + 0.01355915516614914, + 0.020690059289336205, + 0.014051576144993305, + 0.023759625852108, + 0.018300864845514297, + 0.03876079246401787, + 0.0055948481895029545, + 0.005701557267457247, + 0.006177531089633703, + 0.00473048584535718, + 0.005158421117812395, + 0.012733285315334797, + 0.006381370592862368, + 0.006345172878354788, + 0.0038188996259123087, + 0.006958129350095987, + 0.005474284756928682, + 0.005575351882725954, + 0.004399959929287434, + 0.0037346009630709887, + 0.005995703395456076, + 0.00501799862831831, + 0.008457696065306664, + 0.006917004473507404, + 0.005410437472164631, + 0.011733340099453926, + 0.005494121462106705, + 0.005182298365980387, + 0.008643332868814468, + 0.0072494917549192905, + 0.004305354785174131, + 0.00399209326133132, + 0.007424322888255119, + 0.0046452926471829414, + 0.002084837295114994, + 0.003009926527738571, + 0.0037160285282880068, + 0.007348573766648769, + 0.005259530618786812, + 0.003136485582217574, + 0.0032034716568887234, + 0.0048625534400343895, + 0.00577592896297574, + 0.004736864939332008, + 0.0053222766146063805, + 0.005429713521152735, + 0.011481907218694687, + 0.005628567188978195, + 0.006549448240548372, + 0.006150222849100828, + 0.039057016372680664, + 0.03438376635313034, + 0.006958909798413515, + 0.010753368958830833, + 0.016486434265971184, + 0.008740581572055817, + 0.008539419621229172, + 0.008721781894564629, + 0.013057553209364414, + 0.012205038219690323, + 0.006351787131279707, + 0.008049728348851204, + 0.005772959440946579, + 0.003918382339179516, + 0.00853267777711153, + 0.004481008742004633, + 0.006126824766397476, + 0.0041128043085336685, + 0.002912449184805155, + 0.0035906832199543715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.02058335766196251 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864705726504326 + ], + "opposite_logprob_tokens": [ + -3.048647165298462 + ], + "logprob_reward": -0.04864705726504326, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864705726504326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10186650604009628, + 0.0644419938325882, + 0.1118537038564682, + 0.04355708882212639, + 0.0778864324092865, + 0.1882881373167038, + 0.10860242694616318, + 0.06352418661117554, + 0.2057490050792694, + 0.034230560064315796 + ], + "credit_map": [ + 0.005897197872400284, + 0.011775664053857327, + 0.013422817923128605, + 0.015593106858432293, + 0.018794097006320953, + 0.01468923594802618, + 0.010328126139938831, + 0.011366251856088638, + 0.0045072720386087894, + 0.00796901248395443, + 0.005365189164876938, + 0.010404638014733791, + 0.00943921972066164, + 0.00928869191557169, + 0.00687390984967351, + 0.010594058781862259, + 0.007962734438478947, + 0.013821510598063469, + 0.009679165668785572, + 0.01834178902208805, + 0.014060921035706997, + 0.018169337883591652, + 0.0169033482670784, + 0.012914895080029964, + 0.0034865473862737417, + 0.006426116451621056, + 0.003359228605404496, + 0.0066793253645300865, + 0.006610267795622349, + 0.007303705904632807, + 0.005357847549021244, + 0.004334053490310907, + 0.004948088899254799, + 0.013795637525618076, + 0.005791069939732552, + 0.012661763466894627, + 0.00980124156922102, + 0.005648139398545027, + 0.016657015308737755, + 0.008583472110331059, + 0.018062904477119446, + 0.01310407929122448, + 0.05077186971902847, + 0.01536722294986248, + 0.00695475609973073, + 0.01745317317545414, + 0.014615914784371853, + 0.051958225667476654, + 0.012149377726018429, + 0.011557068675756454, + 0.01289825513958931, + 0.012094041332602501, + 0.013083286583423615, + 0.01093412283807993, + 0.01655755005776882, + 0.019328726455569267, + 0.007217203266918659, + 0.004319406114518642, + 0.005354705732315779, + 0.005625982768833637, + 0.01909961737692356, + 0.011637524701654911, + 0.005681668873876333, + 0.004588072653859854, + 0.032745230942964554, + 0.02219555899500847, + 0.017178885638713837, + 0.021843884140253067, + 0.045954421162605286, + 0.027011007070541382, + 0.016244491562247276, + 0.022575514391064644, + 0.004823142196983099, + 0.003984409850090742, + 0.004588613752275705, + 0.0041117495857179165, + 0.006482847034931183, + 0.003394684987142682, + 0.0034202393144369125, + 0.003424872411414981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.04864705726504326 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crimson rose bud that is about to bloom?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the rose petals covered in countless glistening, plump beads of morning dew?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a completely blurred, hazy emerald green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt slightly upward as the flower is lifted upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dewdrops stretch into tiny streaks and peel off the petal edge one by one in super-slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026335977017879486 + ], + "opposite_logprob_tokens": [ + -3.6513359546661377 + ], + "logprob_reward": -0.026335977017879486, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026335977017879486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1849147379398346, + 0.03243830427527428, + 0.04576360061764717, + 0.051050830632448196, + 0.025607289746403694, + 0.08001360297203064, + 0.05977826938033104, + 0.17912770807743073, + 0.26997557282447815, + 0.07133002579212189 + ], + "credit_map": [ + 0.008731426671147346, + 0.017668411135673523, + 0.02128896862268448, + 0.022317953407764435, + 0.03317044675350189, + 0.02848084643483162, + 0.015609324909746647, + 0.03764735534787178, + 0.002409231849014759, + 0.004990426357835531, + 0.002870778786018491, + 0.00493355980142951, + 0.005292457994073629, + 0.0040665073320269585, + 0.0033974023535847664, + 0.004477940034121275, + 0.003743760986253619, + 0.005565986502915621, + 0.004754875786602497, + 0.006685787811875343, + 0.005151110701262951, + 0.008401316590607166, + 0.005807146430015564, + 0.005653615575283766, + 0.0037386142648756504, + 0.0097848204895854, + 0.0034810942597687244, + 0.007484393194317818, + 0.007646882440894842, + 0.005546364467591047, + 0.007937205024063587, + 0.005431455560028553, + 0.0017858354840427637, + 0.0031833297107368708, + 0.002613453194499016, + 0.004345694091171026, + 0.003050121944397688, + 0.0032889768481254578, + 0.003841281170025468, + 0.0034985982347279787, + 0.011144251562654972, + 0.00722713116556406, + 0.00891901645809412, + 0.011464149691164494, + 0.006752449087798595, + 0.0098572364076972, + 0.01340307667851448, + 0.011246291920542717, + 0.0075709936209023, + 0.006900928448885679, + 0.005038714036345482, + 0.005745019763708115, + 0.0073239742778241634, + 0.006377590820193291, + 0.004986894782632589, + 0.015834150835871696, + 0.020261025056242943, + 0.011446869000792503, + 0.012014204636216164, + 0.011756638996303082, + 0.06032077968120575, + 0.03482416272163391, + 0.013329783454537392, + 0.015174252912402153, + 0.03943583741784096, + 0.030739355832338333, + 0.02146795205771923, + 0.028972674161195755, + 0.06319202482700348, + 0.04204164445400238, + 0.024634387344121933, + 0.019491693004965782, + 0.009675712324678898, + 0.007620298303663731, + 0.010681458748877048, + 0.006795878522098064, + 0.017515046522021294, + 0.008286247961223125, + 0.005095574073493481, + 0.005659808870404959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.026335977017879486 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_15/index_tensor([7713, 7713])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_15/index_tensor([7713, 7713])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..73af0a4a445e30c8ddd14faf1179738c19ff44c6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_15/index_tensor([7713, 7713])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, shot from a slightly upward angle, focuses on a young man with delicate features. He wears a simple, blue-gray cotton and linen robe, the collar neatly folded across his chest. His long, jet-black hair is tied back with a simple wooden hairpin, a few strands falling around his pale cheeks. An unresolved sorrow clenched his brow, his lips pressed tightly together, attempting to suppress the immense grief within, but his trembling jaw betrays the suppressed emotion. A full, crystal-clear tear slowly falls from the corner of his slightly reddened right eye, leaving a wet, shimmering mark on his delicate skin, its surface reflecting faint light like a transparent gem. The background is a hazy, quaint study. A soft, slightly cool ray of natural light slants in through a carved wooden window, illuminating half of his cheek and making the teardrop stand out. It also casts a deep shadow on the other side of his face, intensifying the three-dimensional effect of the image. The main color tone of the entire picture is cold, creating an extremely peaceful yet repressed and sad atmosphere, full of restrained tragic beauty.", + "seed": 25770, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_15/index_7713_seed_25770_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16489602141082288, + "vlm_logprob": -0.16489602141082288, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.16489602141082288, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028571048751473427 + ], + "opposite_logprob_tokens": [ + -3.778571128845215 + ], + "logprob_reward": -0.028571048751473427, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.028571048751473427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2319031059741974, + 0.05242323502898216, + 0.04250074177980423, + 0.04413551464676857, + 0.12439887970685959, + 0.16621646285057068, + 0.18314221501350403, + 0.07578758895397186, + 0.01784217357635498, + 0.06165016070008278 + ], + "credit_map": [ + 0.016359087079763412, + 0.014755839481949806, + 0.03138027712702751, + 0.016322728246450424, + 0.03985464572906494, + 0.042684003710746765, + 0.022251585498452187, + 0.04829493165016174, + 0.004141249693930149, + 0.002643582411110401, + 0.005750108975917101, + 0.005091117229312658, + 0.012175023555755615, + 0.007552793249487877, + 0.008231709711253643, + 0.006837647408246994, + 0.0035889241844415665, + 0.0037901969626545906, + 0.004294621758162975, + 0.0045029716566205025, + 0.008168643340468407, + 0.006901825312525034, + 0.005692672915756702, + 0.0055608851835131645, + 0.0034305048175156116, + 0.003016865812242031, + 0.005318303592503071, + 0.0022123991511762142, + 0.008936269208788872, + 0.0044985138811171055, + 0.010911417193710804, + 0.005811244249343872, + 0.010315894149243832, + 0.007563651539385319, + 0.015552924014627934, + 0.008582024835050106, + 0.02248220145702362, + 0.014745011925697327, + 0.029609689489006996, + 0.015547489747405052, + 0.005751712713390589, + 0.00525398924946785, + 0.015943560749292374, + 0.007931607775390148, + 0.02649366483092308, + 0.04726603999733925, + 0.027004621922969818, + 0.03057127073407173, + 0.005014180205762386, + 0.004319412168115377, + 0.014635082334280014, + 0.006593981757760048, + 0.013853429816663265, + 0.06656436622142792, + 0.01220335066318512, + 0.05995841324329376, + 0.003911382053047419, + 0.00399482436478138, + 0.006056587677448988, + 0.00265892269089818, + 0.024088749662041664, + 0.006347334012389183, + 0.005997529719024897, + 0.02273225411772728, + 0.0017504087882116437, + 0.0017098127864301205, + 0.00232587568461895, + 0.002250140765681863, + 0.002083528321236372, + 0.002397133968770504, + 0.0028219041414558887, + 0.0025033699348568916, + 0.0023375973105430603, + 0.0016694109654054046, + 0.0039031202904880047, + 0.003348936792463064, + 0.007528120186179876, + 0.015642691403627396, + 0.008884656243026257, + 0.018335627391934395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.028571048751473427 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02061348780989647 + ], + "opposite_logprob_tokens": [ + -3.895613431930542 + ], + "logprob_reward": -0.02061348780989647, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061348780989647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24496100842952728, + 0.07835236936807632, + 0.07656166702508926, + 0.05476304516196251, + 0.09091503173112869, + 0.06757942587137222, + 0.13620725274085999, + 0.04566175490617752, + 0.04880589619278908, + 0.15619251132011414 + ], + "credit_map": [ + 0.01895274594426155, + 0.01643836684525013, + 0.030181515961885452, + 0.02383406087756157, + 0.04272157698869705, + 0.05027720332145691, + 0.02375509776175022, + 0.038800448179244995, + 0.007445030380040407, + 0.005047492682933807, + 0.008329443633556366, + 0.009816788136959076, + 0.015780126675963402, + 0.013197129592299461, + 0.009797721169888973, + 0.008938637562096119, + 0.004030948970466852, + 0.004878541454672813, + 0.008570603094995022, + 0.009843794628977776, + 0.015161029063165188, + 0.014687160030007362, + 0.01123097725212574, + 0.008158611133694649, + 0.003728611394762993, + 0.003987527918070555, + 0.006659429520368576, + 0.005894286558032036, + 0.007565479725599289, + 0.008807222358882427, + 0.009957528673112392, + 0.008162962272763252, + 0.014873120933771133, + 0.007505150977522135, + 0.011664479970932007, + 0.005890125408768654, + 0.012915685772895813, + 0.013389450497925282, + 0.011586204171180725, + 0.013090817257761955, + 0.006209562532603741, + 0.004254237748682499, + 0.011080939322710037, + 0.004243685398250818, + 0.00935761071741581, + 0.013304396532475948, + 0.011298945173621178, + 0.00783004891127348, + 0.00769929401576519, + 0.004542904440313578, + 0.012968447990715504, + 0.004999920725822449, + 0.007885957136750221, + 0.04241518676280975, + 0.007193172816187143, + 0.048502374440431595, + 0.0045312982983887196, + 0.003834915580227971, + 0.006182767450809479, + 0.0041744341142475605, + 0.006680459249764681, + 0.007265832275152206, + 0.007411236874759197, + 0.005580810829997063, + 0.006192534696310759, + 0.003362495917826891, + 0.006764029152691364, + 0.004948340822011232, + 0.006115586496889591, + 0.006402368191629648, + 0.006688469555228949, + 0.008332072757184505, + 0.013907141052186489, + 0.012672494165599346, + 0.009479749016463757, + 0.009518842212855816, + 0.018818823620676994, + 0.029906926676630974, + 0.024885274469852448, + 0.037003252655267715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.02061348780989647 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559425234794617 + ], + "opposite_logprob_tokens": [ + -1.6005942821502686 + ], + "logprob_reward": -0.22559425234794617, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559425234794617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14474506676197052, + 0.05226835608482361, + 0.09201473742723465, + 0.07412446290254593, + 0.07488264888525009, + 0.09847849607467651, + 0.15060044825077057, + 0.1933886706829071, + 0.06522922962903976, + 0.05426788702607155 + ], + "credit_map": [ + 0.015645800158381462, + 0.010247579775750637, + 0.020601078867912292, + 0.014762639999389648, + 0.01656326837837696, + 0.016416240483522415, + 0.015423282980918884, + 0.03508519008755684, + 0.004858412314206362, + 0.0039007021114230156, + 0.004688637796789408, + 0.005260668694972992, + 0.009701299481093884, + 0.006597092375159264, + 0.010493237525224686, + 0.006768305320292711, + 0.006471041589975357, + 0.005281401798129082, + 0.009384684264659882, + 0.00708999577909708, + 0.011922896839678288, + 0.01312890276312828, + 0.023758234456181526, + 0.014977582730352879, + 0.008735204115509987, + 0.006885108072310686, + 0.00872294045984745, + 0.006308975629508495, + 0.010235846973955631, + 0.011134077794849873, + 0.01107439212501049, + 0.011027914471924305, + 0.009410526603460312, + 0.004959206562489271, + 0.008717944845557213, + 0.005678641144186258, + 0.007473353296518326, + 0.010749475099146366, + 0.015694644302129745, + 0.012198854237794876, + 0.00754966028034687, + 0.006395808421075344, + 0.013131975196301937, + 0.007175135426223278, + 0.019556663930416107, + 0.011151961050927639, + 0.02213694155216217, + 0.011380352079868317, + 0.004441471770405769, + 0.0034429095685482025, + 0.013732991181313992, + 0.003926842473447323, + 0.008787745609879494, + 0.05640416219830513, + 0.008624287322163582, + 0.05124003812670708, + 0.010626757517457008, + 0.010050942189991474, + 0.015580270439386368, + 0.006092590745538473, + 0.06488684564828873, + 0.01226069126278162, + 0.009217187762260437, + 0.06467338651418686, + 0.006941079627722502, + 0.0045185452327132225, + 0.010022002272307873, + 0.006882970687001944, + 0.010117094032466412, + 0.008065180853009224, + 0.01260059978812933, + 0.006081757135689259, + 0.007838787510991096, + 0.007213657721877098, + 0.004704458173364401, + 0.005057319067418575, + 0.007549899164587259, + 0.005417779553681612, + 0.010602238588035107, + 0.005883747711777687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.22559425234794617 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026377543807029724 + ], + "opposite_logprob_tokens": [ + -3.6513774394989014 + ], + "logprob_reward": -0.026377543807029724, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026377543807029724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30839064717292786, + 0.054737430065870285, + 0.038514144718647, + 0.10210412740707397, + 0.03603446111083031, + 0.04167451336979866, + 0.024504654109477997, + 0.03151740878820419, + 0.031186101958155632, + 0.3313365578651428 + ], + "credit_map": [ + 0.029921425506472588, + 0.02579743228852749, + 0.06153356656432152, + 0.020408213138580322, + 0.01768612302839756, + 0.03938654437661171, + 0.013215843588113785, + 0.10044148564338684, + 0.009629395790398121, + 0.008383126929402351, + 0.006513399071991444, + 0.006297021172940731, + 0.0034787296317517757, + 0.00626478111371398, + 0.004001124761998653, + 0.010169853456318378, + 0.00461180554702878, + 0.004780637566000223, + 0.005424489267170429, + 0.004679257050156593, + 0.004162291064858437, + 0.005345414392650127, + 0.004497248213738203, + 0.00501300347968936, + 0.01763945259153843, + 0.016129285097122192, + 0.015434169210493565, + 0.012046776711940765, + 0.009144185110926628, + 0.010907234624028206, + 0.008122872561216354, + 0.012680146843194962, + 0.005995813757181168, + 0.0050780512392520905, + 0.004869504366070032, + 0.0035576338414102793, + 0.003973695915192366, + 0.0036826913710683584, + 0.004111040849238634, + 0.004766029305756092, + 0.003258270677179098, + 0.002888392424210906, + 0.00726232910528779, + 0.0035585914738476276, + 0.00477939797565341, + 0.010790427215397358, + 0.003951809834688902, + 0.005185293965041637, + 0.002124055055901408, + 0.001388539676554501, + 0.002548473421484232, + 0.001296939211897552, + 0.00309703778475523, + 0.008098685182631016, + 0.0022446289658546448, + 0.003706294810399413, + 0.003740459680557251, + 0.0034723677672445774, + 0.005826360546052456, + 0.003340622177347541, + 0.004019563086330891, + 0.0040145330131053925, + 0.003581543220207095, + 0.003521958366036415, + 0.0030421935953199863, + 0.0026500930543988943, + 0.004157265182584524, + 0.0029388393741101027, + 0.00330531457439065, + 0.005497119389474392, + 0.0049007488414645195, + 0.004694527480751276, + 0.020761167630553246, + 0.015643520280718803, + 0.026386717334389687, + 0.011833910830318928, + 0.03187836706638336, + 0.09691891074180603, + 0.021603289991617203, + 0.10631067305803299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.026377543807029724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233237743377686 + ], + "opposite_logprob_tokens": [ + -0.8983237743377686 + ], + "logprob_reward": -0.5233237743377686, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233237743377686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12782010436058044, + 0.031769100576639175, + 0.02502087876200676, + 0.020224856212735176, + 0.04497481882572174, + 0.08355225622653961, + 0.4065689444541931, + 0.11925144493579865, + 0.04292816296219826, + 0.09788944572210312 + ], + "credit_map": [ + 0.013316669501364231, + 0.008411092683672905, + 0.024932926520705223, + 0.011101135984063148, + 0.008079205639660358, + 0.01526772789657116, + 0.008198277093470097, + 0.03851306438446045, + 0.0024018902331590652, + 0.001773722586221993, + 0.004213240463286638, + 0.0032621980644762516, + 0.004897322040051222, + 0.0038951532915234566, + 0.004686201456934214, + 0.006639373488724232, + 0.001992035424336791, + 0.0016617490909993649, + 0.002583072753623128, + 0.002557217376306653, + 0.0038589590694755316, + 0.003096929518505931, + 0.005514268297702074, + 0.0037566462997347116, + 0.001386633375659585, + 0.00139694195240736, + 0.0026098862290382385, + 0.0018126490758731961, + 0.0036588581278920174, + 0.0024327996652573347, + 0.004356528632342815, + 0.0025705595035105944, + 0.004807253833860159, + 0.0029221053700894117, + 0.006481213495135307, + 0.003422481706365943, + 0.007804980035871267, + 0.005824512802064419, + 0.007993096485733986, + 0.0057191746309399605, + 0.003774880664423108, + 0.003098314395174384, + 0.01289176195859909, + 0.0037895662244409323, + 0.00957325380295515, + 0.03111352026462555, + 0.0070062424056231976, + 0.012304713018238544, + 0.008579429239034653, + 0.007271287031471729, + 0.0355745293200016, + 0.011406748555600643, + 0.022487560287117958, + 0.15946133434772491, + 0.018463479354977608, + 0.1433245688676834, + 0.004780446644872427, + 0.004178288858383894, + 0.010791564360260963, + 0.006304593291133642, + 0.03730672597885132, + 0.012394369579851627, + 0.012393198907375336, + 0.031102260574698448, + 0.003308293642476201, + 0.0034282749984413385, + 0.004574927501380444, + 0.005512327887117863, + 0.0045092287473380566, + 0.005211026407778263, + 0.009454871527850628, + 0.006929208058863878, + 0.003384161042049527, + 0.0035180114209651947, + 0.00409495597705245, + 0.005656139459460974, + 0.011342997662723064, + 0.025037245824933052, + 0.01859734207391739, + 0.026258595287799835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.5233237743377686 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly upward angle, focuses on a young man with delicate features. He wears a simple, blue-gray cotton and linen robe, the collar neatly folded across his chest. His long, jet-black hair is tied back with a simple wooden hairpin, a few strands falling around his pale cheeks. An unresolved sorrow clenched his brow, his lips pressed tightly together, attempting to suppress the immense grief within, but his trembling jaw betrays the suppressed emotion. A full, crystal-clear tear slowly falls from the corner of his slightly reddened right eye, leaving a wet, shimmering mark on his delicate skin, its surface reflecting faint light like a transparent gem. The background is a hazy, quaint study. A soft, slightly cool ray of natural light slants in through a carved wooden window, illuminating half of his cheek and making the teardrop stand out. It also casts a deep shadow on the other side of his face, intensifying the three-dimensional effect of the image. The main color tone of the entire picture is cold, creating an extremely peaceful yet repressed and sad atmosphere, full of restrained tragic beauty.", + "seed": 25771, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_15/index_7713_seed_25771_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1657445777207613, + "vlm_logprob": -0.1657445777207613, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.1657445777207613, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014456884935498238 + ], + "opposite_logprob_tokens": [ + -4.639456748962402 + ], + "logprob_reward": -0.014456884935498238, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.014456884935498238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17398419976234436, + 0.05869290605187416, + 0.045930370688438416, + 0.03649578616023064, + 0.2977830767631531, + 0.07942189276218414, + 0.053734928369522095, + 0.041429661214351654, + 0.029364151880145073, + 0.18316295742988586 + ], + "credit_map": [ + 0.015812980011105537, + 0.011597705073654652, + 0.028353575617074966, + 0.018602576106786728, + 0.029495181515812874, + 0.01604282669723034, + 0.03426039591431618, + 0.019818991422653198, + 0.006072693038731813, + 0.0042924098670482635, + 0.00889087188988924, + 0.0061116390861570835, + 0.010083101689815521, + 0.007977860048413277, + 0.008235999383032322, + 0.007028334308415651, + 0.004238004796206951, + 0.0024996851570904255, + 0.008255166932940483, + 0.004932865034788847, + 0.007604209706187248, + 0.003246589098125696, + 0.008728026412427425, + 0.006425830069929361, + 0.0037233366165310144, + 0.0035006944090127945, + 0.00538026075810194, + 0.002393110189586878, + 0.006112630013376474, + 0.006986015476286411, + 0.0045402683317661285, + 0.003859474090859294, + 0.04892713204026222, + 0.014358323998749256, + 0.09630641341209412, + 0.015595883131027222, + 0.05187327787280083, + 0.021962864324450493, + 0.01977529376745224, + 0.02898392453789711, + 0.009998603723943233, + 0.006926089059561491, + 0.012605378404259682, + 0.006186466198414564, + 0.011834091506898403, + 0.015190514735877514, + 0.008509152568876743, + 0.008171608671545982, + 0.0038337931036949158, + 0.005304722115397453, + 0.008420375175774097, + 0.00463977362960577, + 0.010643848218023777, + 0.004596050828695297, + 0.00986518245190382, + 0.006431189365684986, + 0.007863047532737255, + 0.002858453430235386, + 0.005824057385325432, + 0.00324828433804214, + 0.005271558184176683, + 0.007293341215699911, + 0.0050702267326414585, + 0.004000695887953043, + 0.0031876133289188147, + 0.0026735065039247274, + 0.004795296583324671, + 0.0025400337763130665, + 0.005488598253577948, + 0.0029580702539533377, + 0.004747885745018721, + 0.0029731509275734425, + 0.03933696821331978, + 0.013029810041189194, + 0.02704869583249092, + 0.01121246162801981, + 0.021035896614193916, + 0.03625229373574257, + 0.017101095989346504, + 0.018145747482776642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.014456884935498238 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018206151202321053 + ], + "opposite_logprob_tokens": [ + -4.0182061195373535 + ], + "logprob_reward": -0.018206151202321053, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018206151202321053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2638032138347626, + 0.16833680868148804, + 0.0444224588572979, + 0.10336669534444809, + 0.21954119205474854, + 0.07247235625982285, + 0.04440978914499283, + 0.017375167459249496, + 0.021260669454932213, + 0.045011572539806366 + ], + "credit_map": [ + 0.02494928054511547, + 0.024932174012064934, + 0.04016435146331787, + 0.025557231158018112, + 0.04591069743037224, + 0.02292560786008835, + 0.0433034747838974, + 0.03606041893362999, + 0.01501520536839962, + 0.013421067968010902, + 0.02700181119143963, + 0.016353119164705276, + 0.02682510018348694, + 0.025285618379712105, + 0.018885955214500427, + 0.025548934936523438, + 0.0036203742492944, + 0.0033057716209441423, + 0.007233678828924894, + 0.003857514588162303, + 0.007592088542878628, + 0.006321044638752937, + 0.0062751127406954765, + 0.006216876674443483, + 0.008850017562508583, + 0.010648657567799091, + 0.014983857981860638, + 0.006484914571046829, + 0.013844241388142109, + 0.02661118470132351, + 0.00849844329059124, + 0.013445388525724411, + 0.020201900973916054, + 0.02170814573764801, + 0.03908095136284828, + 0.02041107416152954, + 0.03371959924697876, + 0.015294625423848629, + 0.02699998766183853, + 0.0421249121427536, + 0.008671456016600132, + 0.006367248948663473, + 0.011476801708340645, + 0.005122630391269922, + 0.009148718789219856, + 0.014987759292125702, + 0.007704701274633408, + 0.008993040770292282, + 0.005301042925566435, + 0.005303720012307167, + 0.006603220943361521, + 0.003909350838512182, + 0.006814482621848583, + 0.006252695806324482, + 0.006584925577044487, + 0.003640356007963419, + 0.0017724315403029323, + 0.001590550527907908, + 0.002783084288239479, + 0.0022880262695252895, + 0.0028183951508253813, + 0.0013261858839541674, + 0.0030611413531005383, + 0.0017353546572849154, + 0.003320651361718774, + 0.0015766983851790428, + 0.0030307690612971783, + 0.001481813844293356, + 0.00337006151676178, + 0.003213182557374239, + 0.003201803658157587, + 0.0020656920969486237, + 0.008282016962766647, + 0.0031625390984117985, + 0.007995499297976494, + 0.003148385090753436, + 0.005922721698880196, + 0.008020224049687386, + 0.004194173496216536, + 0.004286016803234816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.018206151202321053 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252125084400177 + ], + "opposite_logprob_tokens": [ + -1.5021250247955322 + ], + "logprob_reward": -0.252125084400177, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252125084400177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15917368233203888, + 0.0456102229654789, + 0.05348857492208481, + 0.06103847548365593, + 0.3954058289527893, + 0.07465990632772446, + 0.0427587628364563, + 0.02496003359556198, + 0.03079795464873314, + 0.1121065765619278 + ], + "credit_map": [ + 0.020148223266005516, + 0.013429977931082249, + 0.024274179711937904, + 0.017960648983716965, + 0.02097412571310997, + 0.010991415940225124, + 0.037020064890384674, + 0.014375047758221626, + 0.008139267563819885, + 0.0030903471633791924, + 0.007718604989349842, + 0.004419498611241579, + 0.005252602510154247, + 0.006046110764145851, + 0.0048066116869449615, + 0.006137177813798189, + 0.0037745381705462933, + 0.003960298839956522, + 0.008723042905330658, + 0.004697008058428764, + 0.00797724723815918, + 0.011136746034026146, + 0.007029744330793619, + 0.006189951207488775, + 0.005871120374649763, + 0.006186988670378923, + 0.00861688144505024, + 0.004550507292151451, + 0.00827461015433073, + 0.014857574366033077, + 0.0055503626354038715, + 0.007130431942641735, + 0.065902940928936, + 0.015223770402371883, + 0.12924739718437195, + 0.02199164777994156, + 0.06362877041101456, + 0.03124306909739971, + 0.025979606434702873, + 0.042188625782728195, + 0.015749618411064148, + 0.005597737152129412, + 0.009751037694513798, + 0.006596997380256653, + 0.009860983118414879, + 0.012008561752736568, + 0.007096424233168364, + 0.007998545654118061, + 0.0027007509488612413, + 0.003468302544206381, + 0.007011604029685259, + 0.003840710036456585, + 0.009481859393417835, + 0.004129739478230476, + 0.007410282734781504, + 0.004715512506663799, + 0.0033117844723165035, + 0.0020718160085380077, + 0.0035299332812428474, + 0.0016128929564729333, + 0.0043237777426838875, + 0.0032396111637353897, + 0.004381046164780855, + 0.0024891714565455914, + 0.005844992119818926, + 0.0025893154088407755, + 0.004348959308117628, + 0.001987937605008483, + 0.0042862300761044025, + 0.005229316186159849, + 0.0035746688954532146, + 0.002936534583568573, + 0.023899570107460022, + 0.0077634891495108604, + 0.017109299078583717, + 0.007036190014332533, + 0.01325635053217411, + 0.021074805408716202, + 0.010261706076562405, + 0.011705175042152405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.252125084400177 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020601576194167137 + ], + "opposite_logprob_tokens": [ + -3.895601511001587 + ], + "logprob_reward": -0.020601576194167137, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020601576194167137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08371735364198685, + 0.02901231311261654, + 0.05649140104651451, + 0.05795632675290108, + 0.0276118703186512, + 0.1889200061559677, + 0.027640191838145256, + 0.032224640250205994, + 0.026089895516633987, + 0.47033607959747314 + ], + "credit_map": [ + 0.011766932904720306, + 0.007026893086731434, + 0.013394476845860481, + 0.00945475697517395, + 0.011743255890905857, + 0.007184999529272318, + 0.013118266128003597, + 0.010027768090367317, + 0.004987653810530901, + 0.0021362295374274254, + 0.004842944908887148, + 0.002681528450921178, + 0.004924270790070295, + 0.0024849397595971823, + 0.0037146590184420347, + 0.003240083809942007, + 0.008808276616036892, + 0.0033226669766008854, + 0.010819555260241032, + 0.004689405672252178, + 0.006666080094873905, + 0.009151246398687363, + 0.006674281321465969, + 0.006359882652759552, + 0.005484006833285093, + 0.004571012686938047, + 0.00900352094322443, + 0.005549308378249407, + 0.00814186129719019, + 0.010361510328948498, + 0.007132374681532383, + 0.007712727878242731, + 0.003225464839488268, + 0.001684698392637074, + 0.006721664220094681, + 0.0021616423036903143, + 0.005167458206415176, + 0.0022001899778842926, + 0.003229189198464155, + 0.003221560502424836, + 0.024211185052990913, + 0.014267001301050186, + 0.03557203710079193, + 0.014162213541567326, + 0.024873433634638786, + 0.03940518572926521, + 0.018374836072325706, + 0.01805408112704754, + 0.002851111115887761, + 0.002812998602166772, + 0.004226853605359793, + 0.0028293936047703028, + 0.0047307852655649185, + 0.003284946782514453, + 0.004414461087435484, + 0.0024896403774619102, + 0.005443906411528587, + 0.002950706286355853, + 0.005503166001290083, + 0.0023696308489888906, + 0.005141734145581722, + 0.003934049513190985, + 0.0040327152237296104, + 0.0028487287927418947, + 0.004419561475515366, + 0.0020503029227256775, + 0.0033670980483293533, + 0.001973642036318779, + 0.004230353981256485, + 0.0045805382542312145, + 0.0030977630522102118, + 0.0023706313222646713, + 0.10679832100868225, + 0.03184204921126366, + 0.06608673930168152, + 0.02901517227292061, + 0.05448295921087265, + 0.09093010425567627, + 0.043848998844623566, + 0.04733166843652725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.020601576194167137 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233331918716431 + ], + "opposite_logprob_tokens": [ + -0.8983331918716431 + ], + "logprob_reward": -0.5233331918716431, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233331918716431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13887444138526917, + 0.04951002821326256, + 0.03544922545552254, + 0.06926603615283966, + 0.30170920491218567, + 0.08178462088108063, + 0.10423313081264496, + 0.0979020893573761, + 0.08007455617189407, + 0.04119665548205376 + ], + "credit_map": [ + 0.02200530655682087, + 0.011325824074447155, + 0.020135311409831047, + 0.010169868357479572, + 0.019485924392938614, + 0.01972121000289917, + 0.017551712691783905, + 0.018479283899068832, + 0.0054418477229774, + 0.0035944224800914526, + 0.00795488990843296, + 0.004918099846690893, + 0.00844220444560051, + 0.004885447211563587, + 0.0068331630900502205, + 0.007439952809363604, + 0.0049606803804636, + 0.003479451173916459, + 0.005771128460764885, + 0.00233889976516366, + 0.006075928453356028, + 0.004861896391957998, + 0.003850298235192895, + 0.00411094818264246, + 0.005408565513789654, + 0.006418668664991856, + 0.010671288706362247, + 0.0054870652966201305, + 0.00987421814352274, + 0.01529247872531414, + 0.0061607444658875465, + 0.009953013621270657, + 0.04835963994264603, + 0.01137767918407917, + 0.09881032258272171, + 0.0158450398594141, + 0.05561087653040886, + 0.021356094628572464, + 0.02093425951898098, + 0.02941528707742691, + 0.01769539713859558, + 0.006247953046113253, + 0.011807204224169254, + 0.005844686646014452, + 0.010106178000569344, + 0.014876670204102993, + 0.006767177022993565, + 0.008439356461167336, + 0.009893748909235, + 0.010613306425511837, + 0.016469204798340797, + 0.012363017536699772, + 0.019198479130864143, + 0.011274570599198341, + 0.013634412549436092, + 0.010786395519971848, + 0.019284965470433235, + 0.006600644439458847, + 0.014230103231966496, + 0.0064756060019135475, + 0.012087547220289707, + 0.01673404686152935, + 0.01270252000540495, + 0.00978667289018631, + 0.012674863450229168, + 0.0056231957860291, + 0.01107617188245058, + 0.005098077934235334, + 0.012851376086473465, + 0.012940162792801857, + 0.012873168103396893, + 0.006937551777809858, + 0.007208375725895166, + 0.003129472490400076, + 0.0080942427739501, + 0.003073664614930749, + 0.005550182890146971, + 0.006289580836892128, + 0.003972885198891163, + 0.003878253046423197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.5233331918716431 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly upward angle, focuses on a young man with delicate features. He wears a simple, blue-gray cotton and linen robe, the collar neatly folded across his chest. His long, jet-black hair is tied back with a simple wooden hairpin, a few strands falling around his pale cheeks. An unresolved sorrow clenched his brow, his lips pressed tightly together, attempting to suppress the immense grief within, but his trembling jaw betrays the suppressed emotion. A full, crystal-clear tear slowly falls from the corner of his slightly reddened right eye, leaving a wet, shimmering mark on his delicate skin, its surface reflecting faint light like a transparent gem. The background is a hazy, quaint study. A soft, slightly cool ray of natural light slants in through a carved wooden window, illuminating half of his cheek and making the teardrop stand out. It also casts a deep shadow on the other side of his face, intensifying the three-dimensional effect of the image. The main color tone of the entire picture is cold, creating an extremely peaceful yet repressed and sad atmosphere, full of restrained tragic beauty.", + "seed": 25772, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_15/index_7713_seed_25772_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18596163876354693, + "vlm_logprob": -0.18596163876354693, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.18596163876354693, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03654026612639427 + ], + "opposite_logprob_tokens": [ + -3.5365402698516846 + ], + "logprob_reward": -0.03654026612639427, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03654026612639427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29531994462013245, + 0.1110161542892456, + 0.06280884146690369, + 0.03172722086310387, + 0.03316102549433708, + 0.08621425926685333, + 0.1024361252784729, + 0.07442189753055573, + 0.1255747675895691, + 0.07731983065605164 + ], + "credit_map": [ + 0.019510015845298767, + 0.011255437508225441, + 0.06599600613117218, + 0.01667363941669464, + 0.0514896921813488, + 0.0811673253774643, + 0.020631738007068634, + 0.028596071526408195, + 0.021612176671624184, + 0.010352532379329205, + 0.012106168083846569, + 0.012725010514259338, + 0.02085551619529724, + 0.015445077791810036, + 0.007712787948548794, + 0.010206880047917366, + 0.008509463630616665, + 0.00574134849011898, + 0.006800632458180189, + 0.006921310443431139, + 0.01281585730612278, + 0.009622265584766865, + 0.005463168956339359, + 0.006934789940714836, + 0.002759696450084448, + 0.0023499727249145508, + 0.004135109484195709, + 0.0026105374563485384, + 0.005545278079807758, + 0.004366651643067598, + 0.005320712924003601, + 0.004639257676899433, + 0.004401795566082001, + 0.003692499827593565, + 0.003650289261713624, + 0.0032741555478423834, + 0.004164147190749645, + 0.0038115628995001316, + 0.004183588083833456, + 0.005982982460409403, + 0.01306139212101698, + 0.004702069330960512, + 0.030093131586909294, + 0.007373076863586903, + 0.006739663425832987, + 0.01409635879099369, + 0.0031674636993557215, + 0.006981086917221546, + 0.008038252592086792, + 0.0030295681208372116, + 0.039756227284669876, + 0.008323830552399158, + 0.007996469736099243, + 0.011917042545974255, + 0.0036847074516117573, + 0.019690020009875298, + 0.007930015213787556, + 0.005210464354604483, + 0.011191458441317081, + 0.009281391277909279, + 0.013024176470935345, + 0.008815603330731392, + 0.010944376699626446, + 0.008024402894079685, + 0.010359418578445911, + 0.014228283427655697, + 0.016082897782325745, + 0.013277935795485973, + 0.024921158328652382, + 0.017416441813111305, + 0.013895618729293346, + 0.015392993576824665, + 0.0044128610752522945, + 0.0051675657741725445, + 0.009867068380117416, + 0.008800618350505829, + 0.01647069863975048, + 0.011927539482712746, + 0.009628468193113804, + 0.011045004241168499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.03654026612639427 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014223858714103699 + ], + "opposite_logprob_tokens": [ + -4.264224052429199 + ], + "logprob_reward": -0.014223858714103699, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014223858714103699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3350350558757782, + 0.09876825660467148, + 0.07107523828744888, + 0.06279131025075912, + 0.048553064465522766, + 0.05460421368479729, + 0.05942520871758461, + 0.11867973953485489, + 0.09454856812953949, + 0.056519344449043274 + ], + "credit_map": [ + 0.029468677937984467, + 0.014884641394019127, + 0.04917394742369652, + 0.021771086379885674, + 0.04602883383631706, + 0.10381677001714706, + 0.026700273156166077, + 0.04319079965353012, + 0.020522208884358406, + 0.009508041664958, + 0.01079234853386879, + 0.011666073463857174, + 0.015011008828878403, + 0.012599554844200611, + 0.00826612301170826, + 0.01040289644151926, + 0.012036306783556938, + 0.006468357052654028, + 0.0077363778837025166, + 0.00826960988342762, + 0.010981959290802479, + 0.009347585029900074, + 0.007018882315605879, + 0.009216159582138062, + 0.004567397758364677, + 0.0036011827178299427, + 0.015348429791629314, + 0.006516170687973499, + 0.009483257308602333, + 0.0063355411402881145, + 0.007197837345302105, + 0.009741494432091713, + 0.007782439701259136, + 0.0044951974414289, + 0.011593376286327839, + 0.005183326080441475, + 0.0057433289475739, + 0.004475498106330633, + 0.004134257789701223, + 0.005145643372088671, + 0.005108848214149475, + 0.0029014130122959614, + 0.014527578838169575, + 0.0038715957198292017, + 0.0045486134476959705, + 0.005374397151172161, + 0.0031716427765786648, + 0.015100127086043358, + 0.005778573453426361, + 0.0021563181653618813, + 0.022837840020656586, + 0.0047074672766029835, + 0.005321727599948645, + 0.008450274355709553, + 0.003100245725363493, + 0.00707275839522481, + 0.021081313490867615, + 0.01410506758838892, + 0.021082257851958275, + 0.014078156091272831, + 0.013908714056015015, + 0.011806313879787922, + 0.010584397241473198, + 0.012033519335091114, + 0.010196126066148281, + 0.012353155761957169, + 0.012106507085263729, + 0.015379991382360458, + 0.013621732592582703, + 0.011026362888514996, + 0.0084489481523633, + 0.011415744200348854, + 0.0054437946528196335, + 0.00404627900570631, + 0.00891293678432703, + 0.009279821999371052, + 0.008031300269067287, + 0.006036554463207722, + 0.006941149942576885, + 0.00782750640064478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.014223858714103699 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256011962890625 + ], + "opposite_logprob_tokens": [ + -1.6006011962890625 + ], + "logprob_reward": -0.2256011962890625, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256011962890625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5505322813987732, + 0.09026607871055603, + 0.05538652092218399, + 0.053590912371873856, + 0.0225184615701437, + 0.03409696742892265, + 0.06455601006746292, + 0.05712578445672989, + 0.0453595295548439, + 0.026567475870251656 + ], + "credit_map": [ + 0.038736455142498016, + 0.010770019143819809, + 0.19717763364315033, + 0.024794159457087517, + 0.03984154388308525, + 0.18265661597251892, + 0.022471336647868156, + 0.03408448398113251, + 0.014244470745325089, + 0.005857523996382952, + 0.013847695663571358, + 0.008917579427361488, + 0.01204293966293335, + 0.009419288486242294, + 0.013987754471600056, + 0.011948819272220135, + 0.006187588442116976, + 0.0034302494022995234, + 0.00653029652312398, + 0.004930711816996336, + 0.008775342255830765, + 0.007260187063366175, + 0.008058367297053337, + 0.010213769041001797, + 0.009911635890603065, + 0.0040928698144853115, + 0.009735103696584702, + 0.004059305414557457, + 0.007665352430194616, + 0.005449381656944752, + 0.004750380292534828, + 0.007926877588033676, + 0.002413833048194647, + 0.0014096183003857732, + 0.002951056230813265, + 0.0017434107139706612, + 0.004691469483077526, + 0.0031501944176852703, + 0.0026216881815344095, + 0.0035371887497603893, + 0.0017782741924747825, + 0.0013606853317469358, + 0.009333477355539799, + 0.0032538753002882004, + 0.0049360874108970165, + 0.0028811937663704157, + 0.003342272015288472, + 0.0072110965847969055, + 0.00490144407376647, + 0.0020471469033509493, + 0.022609516978263855, + 0.005158499348908663, + 0.005633482709527016, + 0.00619221106171608, + 0.003399450099095702, + 0.014614248648285866, + 0.007832660339772701, + 0.0053516048938035965, + 0.007799279410392046, + 0.009120607748627663, + 0.007418654393404722, + 0.005924096796661615, + 0.007593600079417229, + 0.006085274275392294, + 0.004259746056050062, + 0.004877409897744656, + 0.0055802264250814915, + 0.008854338899254799, + 0.0065650613978505135, + 0.00570738036185503, + 0.00478837126865983, + 0.0047269947826862335, + 0.002466656267642975, + 0.0014425056288018823, + 0.006097788922488689, + 0.0026954300701618195, + 0.0043085552752017975, + 0.003253309754654765, + 0.0029359515756368637, + 0.003367275930941105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.2256011962890625 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020609749481081963 + ], + "opposite_logprob_tokens": [ + -3.8956098556518555 + ], + "logprob_reward": -0.020609749481081963, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020609749481081963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5482025742530823, + 0.07892218977212906, + 0.04014255478978157, + 0.01944880560040474, + 0.05980667471885681, + 0.03594047576189041, + 0.019449135288596153, + 0.045986924320459366, + 0.12625961005687714, + 0.025841085240244865 + ], + "credit_map": [ + 0.046615369617938995, + 0.009696855209767818, + 0.2017567753791809, + 0.0249603521078825, + 0.03952888771891594, + 0.17445282638072968, + 0.020727334544062614, + 0.03046419657766819, + 0.013075768947601318, + 0.011019189842045307, + 0.012367534451186657, + 0.01235181838274002, + 0.007408066652715206, + 0.009289304725825787, + 0.005545904394239187, + 0.00786459818482399, + 0.007241683546453714, + 0.0053166961297392845, + 0.004854373633861542, + 0.0056076813489198685, + 0.004959944635629654, + 0.004257704131305218, + 0.003511538729071617, + 0.004392929375171661, + 0.002386868465691805, + 0.0015978302108123899, + 0.004265831783413887, + 0.0021407257299870253, + 0.0031830864027142525, + 0.002196072367951274, + 0.0016933403676375747, + 0.001985049806535244, + 0.010338911786675453, + 0.005988658405840397, + 0.01730462722480297, + 0.006008302792906761, + 0.005603233817964792, + 0.005883768200874329, + 0.003210061462596059, + 0.005469114985316992, + 0.003334019100293517, + 0.0025932786520570517, + 0.009272085502743721, + 0.0025622143875807524, + 0.0024720646906644106, + 0.0032910797744989395, + 0.0018674604361876845, + 0.010548274032771587, + 0.002486146055161953, + 0.001731328316964209, + 0.0037063593044877052, + 0.0024319689255207777, + 0.0022575235925614834, + 0.002524363109841943, + 0.0017298415768891573, + 0.0025816052220761776, + 0.006336455699056387, + 0.003875058377161622, + 0.00618162052705884, + 0.008618572726845741, + 0.007173425983637571, + 0.005861749406903982, + 0.0030987965874373913, + 0.0048412419855594635, + 0.021834298968315125, + 0.015482254326343536, + 0.02018674835562706, + 0.01934223063290119, + 0.01620975136756897, + 0.011592837050557137, + 0.008230308070778847, + 0.01338118128478527, + 0.004403936676681042, + 0.003213191172108054, + 0.003196642966940999, + 0.004651789553463459, + 0.0030513349920511246, + 0.002642192644998431, + 0.002333339536562562, + 0.002348656766116619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.020609749481081963 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328331232070923 + ], + "opposite_logprob_tokens": [ + -0.7578331232070923 + ], + "logprob_reward": -0.6328331232070923, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328331232070923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20066925883293152, + 0.0468515083193779, + 0.04579200968146324, + 0.04818453639745712, + 0.043341804295778275, + 0.09563837945461273, + 0.3441590666770935, + 0.07554549723863602, + 0.07324178516864777, + 0.026576165109872818 + ], + "credit_map": [ + 0.019950218498706818, + 0.009025885723531246, + 0.06749415397644043, + 0.01377134770154953, + 0.025425985455513, + 0.03970688208937645, + 0.010369028896093369, + 0.014925745315849781, + 0.005410685203969479, + 0.0033172359690070152, + 0.006258546840399504, + 0.0054247681982815266, + 0.010820004157721996, + 0.005233024712651968, + 0.004380353260785341, + 0.006006889045238495, + 0.008236329071223736, + 0.00552273029461503, + 0.004193862900137901, + 0.0056133633479475975, + 0.007551089394837618, + 0.004547088406980038, + 0.004694034345448017, + 0.005433513317257166, + 0.006158475764095783, + 0.00377840269356966, + 0.007467256858944893, + 0.004521485418081284, + 0.006863915361464024, + 0.0050778696313500404, + 0.008737628348171711, + 0.005579506512731314, + 0.0074546681717038155, + 0.0031388187780976295, + 0.011820245534181595, + 0.003301524091511965, + 0.005366457160562277, + 0.003967533819377422, + 0.0028362444136291742, + 0.005456310231238604, + 0.01369315292686224, + 0.006428019609302282, + 0.022535283118486404, + 0.008943631313741207, + 0.0072334082797169685, + 0.015935420989990234, + 0.005420902743935585, + 0.015448560006916523, + 0.03114321269094944, + 0.010235882364213467, + 0.1457187384366989, + 0.025494661182165146, + 0.026757268235087395, + 0.05528365448117256, + 0.008464181795716286, + 0.041061483323574066, + 0.005714534316211939, + 0.009335620328783989, + 0.009257938712835312, + 0.011805891990661621, + 0.013083431869745255, + 0.00823140051215887, + 0.01052568294107914, + 0.007591004483401775, + 0.003891492262482643, + 0.0034595071338117123, + 0.014305260963737965, + 0.004656495526432991, + 0.01623701862990856, + 0.010666318237781525, + 0.009009440429508686, + 0.011016251519322395, + 0.001835320144891739, + 0.002076179953292012, + 0.004017911851406097, + 0.0022552513983100653, + 0.0055999052710831165, + 0.003823965322226286, + 0.0026563473511487246, + 0.004311284516006708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.6328331232070923 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a slightly upward angle, focuses on a young man with delicate features. He wears a simple, blue-gray cotton and linen robe, the collar neatly folded across his chest. His long, jet-black hair is tied back with a simple wooden hairpin, a few strands falling around his pale cheeks. An unresolved sorrow clenched his brow, his lips pressed tightly together, attempting to suppress the immense grief within, but his trembling jaw betrays the suppressed emotion. A full, crystal-clear tear slowly falls from the corner of his slightly reddened right eye, leaving a wet, shimmering mark on his delicate skin, its surface reflecting faint light like a transparent gem. The background is a hazy, quaint study. A soft, slightly cool ray of natural light slants in through a carved wooden window, illuminating half of his cheek and making the teardrop stand out. It also casts a deep shadow on the other side of his face, intensifying the three-dimensional effect of the image. The main color tone of the entire picture is cold, creating an extremely peaceful yet repressed and sad atmosphere, full of restrained tragic beauty.", + "seed": 25773, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_15/index_7713_seed_25773_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13246027678251265, + "vlm_logprob": -0.13246027678251265, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.13246027678251265, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013312267139554024 + ], + "opposite_logprob_tokens": [ + -4.763312339782715 + ], + "logprob_reward": -0.013312267139554024, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.013312267139554024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2747970223426819, + 0.07691685110330582, + 0.1283390372991562, + 0.051580481231212616, + 0.15167894959449768, + 0.08405475318431854, + 0.03596092388033867, + 0.046654537320137024, + 0.03939938172698021, + 0.11061803251504898 + ], + "credit_map": [ + 0.03862061724066734, + 0.028500961139798164, + 0.05661894008517265, + 0.030272550880908966, + 0.024920456111431122, + 0.04252360388636589, + 0.027570847421884537, + 0.025769036263227463, + 0.010172701440751553, + 0.008960106410086155, + 0.008908017538487911, + 0.013007009401917458, + 0.0075842030346393585, + 0.01007826253771782, + 0.009465371258556843, + 0.008741175755858421, + 0.010760908015072346, + 0.011715211905539036, + 0.02123415842652321, + 0.013850562274456024, + 0.019069457426667213, + 0.01512570958584547, + 0.022316575050354004, + 0.01426644530147314, + 0.006160155404359102, + 0.00697740726172924, + 0.006525783333927393, + 0.006407455075532198, + 0.004815266001969576, + 0.007932716980576515, + 0.006671630311757326, + 0.006090066395699978, + 0.00782165490090847, + 0.011682596057653427, + 0.03811913728713989, + 0.04235540330410004, + 0.009458531625568867, + 0.016215497627854347, + 0.011609429493546486, + 0.01441669836640358, + 0.007166826631873846, + 0.008844895288348198, + 0.01115511916577816, + 0.011513623408973217, + 0.013816077262163162, + 0.010507343336939812, + 0.010807104408740997, + 0.010243764147162437, + 0.0020527623128145933, + 0.0017985414015129209, + 0.00880604237318039, + 0.006306422408670187, + 0.004824589937925339, + 0.003126049879938364, + 0.006133481860160828, + 0.0029130326583981514, + 0.0026952424086630344, + 0.0028670290485024452, + 0.010107893496751785, + 0.007183062378317118, + 0.005036918446421623, + 0.006176010239869356, + 0.005925026722252369, + 0.006663352716714144, + 0.002841450972482562, + 0.0027218987233936787, + 0.006923318840563297, + 0.005529453977942467, + 0.004311032127588987, + 0.0038227513432502747, + 0.005303164944052696, + 0.007946309633553028, + 0.004844327457249165, + 0.0057018171064555645, + 0.016726266592741013, + 0.017175691202282906, + 0.020640790462493896, + 0.011360562406480312, + 0.02291739732027054, + 0.011251175776124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.013312267139554024 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012563253752887249 + ], + "opposite_logprob_tokens": [ + -4.387563228607178 + ], + "logprob_reward": -0.012563253752887249, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012563253752887249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3775192201137543, + 0.07024560868740082, + 0.10403468459844589, + 0.05323333665728569, + 0.05558840557932854, + 0.08488074690103531, + 0.0347457081079483, + 0.04100543633103371, + 0.035667870193719864, + 0.14307893812656403 + ], + "credit_map": [ + 0.05093177780508995, + 0.04095866158604622, + 0.06455837935209274, + 0.045088984072208405, + 0.032891895622015, + 0.060914672911167145, + 0.04103921726346016, + 0.041135646402835846, + 0.01040408294647932, + 0.00821197684854269, + 0.008902022615075111, + 0.010859422385692596, + 0.008821489289402962, + 0.005137936677783728, + 0.009246928617358208, + 0.008661753498017788, + 0.009186923503875732, + 0.009757592342793941, + 0.018981225788593292, + 0.0155804343521595, + 0.013204757124185562, + 0.010691839270293713, + 0.017493071034550667, + 0.009138849563896656, + 0.006273739971220493, + 0.010053299367427826, + 0.00779520720243454, + 0.008004947565495968, + 0.0035757243167608976, + 0.006117840297520161, + 0.005962034221738577, + 0.005450542084872723, + 0.002746402285993099, + 0.0026959434617310762, + 0.01205204427242279, + 0.014008202590048313, + 0.0059476252645254135, + 0.004705716855823994, + 0.006802283227443695, + 0.006630188785493374, + 0.006825744174420834, + 0.007451293058693409, + 0.01094035618007183, + 0.013169076293706894, + 0.010491203516721725, + 0.012495674192905426, + 0.012483112514019012, + 0.0110242934897542, + 0.002555418061092496, + 0.002165037440136075, + 0.007386162411421537, + 0.005795641336590052, + 0.005127396434545517, + 0.00350378779694438, + 0.004515138454735279, + 0.0036971282679587603, + 0.004056920763105154, + 0.0034896698780357838, + 0.007030919194221497, + 0.007079288363456726, + 0.004310466814786196, + 0.003826688276603818, + 0.005647990852594376, + 0.00556349428370595, + 0.0030340533703565598, + 0.0029881035443395376, + 0.005893426947295666, + 0.004732127301394939, + 0.005330952350050211, + 0.0033591247629374266, + 0.004313248209655285, + 0.006016833707690239, + 0.008219152688980103, + 0.008672662079334259, + 0.02586476132273674, + 0.022403966635465622, + 0.02246607281267643, + 0.013470500707626343, + 0.01958215981721878, + 0.0223996601998806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.012563253752887249 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157910883426666 + ], + "opposite_logprob_tokens": [ + -1.7015790939331055 + ], + "logprob_reward": -0.20157910883426666, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157910883426666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15922442078590393, + 0.16749654710292816, + 0.16160312294960022, + 0.20336486399173737, + 0.058712877333164215, + 0.07789978384971619, + 0.024488480761647224, + 0.027470437809824944, + 0.037597302347421646, + 0.0821421891450882 + ], + "credit_map": [ + 0.025657914578914642, + 0.013953804969787598, + 0.02475532703101635, + 0.022520385682582855, + 0.02024875581264496, + 0.018716169521212578, + 0.018327148631215096, + 0.015044896863400936, + 0.01975337415933609, + 0.01523189339786768, + 0.02090151235461235, + 0.034406956285238266, + 0.01710592582821846, + 0.016760647296905518, + 0.024487504735589027, + 0.01884874515235424, + 0.013064681552350521, + 0.01755782589316368, + 0.02976471744477749, + 0.013266473077237606, + 0.02584877610206604, + 0.02419697307050228, + 0.020788323134183884, + 0.017115335911512375, + 0.01000195648521185, + 0.013692605309188366, + 0.0319114550948143, + 0.03576534241437912, + 0.023560356348752975, + 0.03533003479242325, + 0.026982227340340614, + 0.026120895519852638, + 0.006497091148048639, + 0.00759146036580205, + 0.010512658394873142, + 0.008649629540741444, + 0.00662498502060771, + 0.004740624222904444, + 0.007523234933614731, + 0.006573195569217205, + 0.008305542171001434, + 0.008755885064601898, + 0.015035071410238743, + 0.01547092106193304, + 0.006566653959453106, + 0.005999915301799774, + 0.010211720131337643, + 0.007554073818027973, + 0.0013832947006449103, + 0.0024509343784302473, + 0.004844222217798233, + 0.004563166759908199, + 0.002623555948957801, + 0.0026859387289732695, + 0.003421227680519223, + 0.0025161413941532373, + 0.0019792013335973024, + 0.0024459874257445335, + 0.004764013458043337, + 0.0048027122393250465, + 0.00294533371925354, + 0.002733270637691021, + 0.0047951797023415565, + 0.0030047402251511812, + 0.0020981808193027973, + 0.0026880926452577114, + 0.005189249292016029, + 0.00665705930441618, + 0.00419690040871501, + 0.005693661514669657, + 0.0068389191292226315, + 0.004235236439853907, + 0.004904850851744413, + 0.005026143975555897, + 0.011947600170969963, + 0.013807038776576519, + 0.011692050844430923, + 0.011230082251131535, + 0.012277155183255672, + 0.011257261969149113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.20157910883426666 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005960072390735149 + ], + "opposite_logprob_tokens": [ + -5.130959987640381 + ], + "logprob_reward": -0.005960072390735149, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960072390735149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2941027283668518, + 0.11683157086372375, + 0.0823902115225792, + 0.06284879893064499, + 0.14292003214359283, + 0.10123870521783829, + 0.04581088572740555, + 0.053168460726737976, + 0.04035422205924988, + 0.060334376990795135 + ], + "credit_map": [ + 0.03907744958996773, + 0.03270174562931061, + 0.05453132465481758, + 0.03211610019207001, + 0.027714945375919342, + 0.04907326027750969, + 0.029368363320827484, + 0.029519570991396904, + 0.01731276512145996, + 0.011403912678360939, + 0.017584025859832764, + 0.023636015132069588, + 0.013889946974813938, + 0.011703741736710072, + 0.01207994855940342, + 0.009221216663718224, + 0.0078887939453125, + 0.00679782684892416, + 0.012461287900805473, + 0.008836108259856701, + 0.009725276380777359, + 0.013167263008654118, + 0.011919957585632801, + 0.011593691073358059, + 0.007442202884703875, + 0.008171546272933483, + 0.006132614333182573, + 0.01319981086999178, + 0.006149004679173231, + 0.005322879645973444, + 0.007423377595841885, + 0.009007361717522144, + 0.017997484654188156, + 0.015633849427103996, + 0.04051680490374565, + 0.031546734273433685, + 0.008655684068799019, + 0.005164256785064936, + 0.011215925216674805, + 0.012189283967018127, + 0.010200037620961666, + 0.012638921849429607, + 0.017414240166544914, + 0.022390902042388916, + 0.00776701420545578, + 0.011078815907239914, + 0.009587441571056843, + 0.010161331854760647, + 0.004900240805000067, + 0.0045522903092205524, + 0.008969583548605442, + 0.009194117039442062, + 0.00538913207128644, + 0.003892542328685522, + 0.005254015792161226, + 0.0036589629016816616, + 0.005076928995549679, + 0.00431639701128006, + 0.009413006715476513, + 0.009740966372191906, + 0.007325607817620039, + 0.004969592671841383, + 0.007115374784916639, + 0.005210589151829481, + 0.004409668501466513, + 0.004539960529655218, + 0.006104056723415852, + 0.0065789190120995045, + 0.006331915035843849, + 0.0037663194816559553, + 0.00459918100386858, + 0.004024200141429901, + 0.005790320225059986, + 0.006928722839802504, + 0.008186019025743008, + 0.012490343302488327, + 0.007073554210364819, + 0.003982772119343281, + 0.007502303924411535, + 0.0083803441375494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.005960072390735149 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on a young man with long, jet-black hair tied back by a wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a blue-gray cotton and linen robe with a neatly folded collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a full, crystal-clear tear visible falling from the corner of his slightly reddened right eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a hazy, quaint study with a soft ray of natural light slanting through a carved wooden window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42888668179512024 + ], + "opposite_logprob_tokens": [ + -1.0538866519927979 + ], + "logprob_reward": -0.42888668179512024, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42888668179512024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20461249351501465, + 0.10376560688018799, + 0.07685283571481705, + 0.07412680983543396, + 0.12819361686706543, + 0.09574231505393982, + 0.042432039976119995, + 0.08383902162313461, + 0.03898942470550537, + 0.15144579112529755 + ], + "credit_map": [ + 0.02644621953368187, + 0.01651294156908989, + 0.030849352478981018, + 0.0392167828977108, + 0.0322054959833622, + 0.01600244641304016, + 0.026029620319604874, + 0.017349641770124435, + 0.015086827799677849, + 0.01021756511181593, + 0.013994782231748104, + 0.02115350216627121, + 0.0077004460617899895, + 0.010640086606144905, + 0.013243895024061203, + 0.011728505603969097, + 0.006651442032307386, + 0.007743366993963718, + 0.011318463832139969, + 0.00881612952798605, + 0.0114913834258914, + 0.007683668751269579, + 0.014839619398117065, + 0.008308757096529007, + 0.005666201002895832, + 0.008034208789467812, + 0.013210533186793327, + 0.016683226451277733, + 0.006940948776900768, + 0.008156679570674896, + 0.008622645400464535, + 0.006812361069023609, + 0.010310127399861813, + 0.009949752129614353, + 0.01388704963028431, + 0.018557103350758553, + 0.01350143738090992, + 0.022527581080794334, + 0.02381438948214054, + 0.015646183863282204, + 0.005522780120372772, + 0.007234313990920782, + 0.019594188779592514, + 0.01299160998314619, + 0.012497475370764732, + 0.010635145008563995, + 0.012430870905518532, + 0.01483593974262476, + 0.0054649836383759975, + 0.005106405820697546, + 0.0070978919975459576, + 0.008612445555627346, + 0.0036120025906711817, + 0.003898514900356531, + 0.004644774831831455, + 0.003995020408183336, + 0.003393336432054639, + 0.00432246271520853, + 0.017175499349832535, + 0.01374443992972374, + 0.011613037437200546, + 0.006807058118283749, + 0.013525379821658134, + 0.01325780525803566, + 0.0021765390411019325, + 0.003073708154261112, + 0.007077065296471119, + 0.007296079304069281, + 0.006084240507334471, + 0.0033163283951580524, + 0.0054496233351528645, + 0.004515840671956539, + 0.007838230580091476, + 0.009458739310503006, + 0.020811261609196663, + 0.015537995845079422, + 0.02022547647356987, + 0.020592689514160156, + 0.027056816965341568, + 0.029924582690000534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.42888668179512024 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_2/index_tensor([442, 442])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_2/index_tensor([442, 442])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cc4e948622d480df2ffed1074381b0d5af0f9c69 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_2/index_tensor([442, 442])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic overhead shot zooms in, focusing on a warrior clad in black armor. He stands alone in the center of a dilapidated ancient battlefield. Broken weapons and shattered stone statues litter the ground. In the distance, steep mountains rise and fall, shrouded in inky mist. The warrior, a massive figure, clad in a thick, battle-scarred suit of black iron armor, engraved with intricate ancient cloud-thunder patterns. The armor reflects a cool metallic luster in the dim light of eternal twilight. In his right hand, he clutches a spear, also crafted from black iron and emblazoned with intricate dragon motifs. Steadily and deliberately, he executes a motion reminiscent of drawing a sword from its sheath, his body lowering slightly, his arm muscles tensing. At that moment, the dragon motif on the spear suddenly ignites, a fiery crimson glow emanating from the cracks in the carving. This immense energy rapidly condenses, stretches, and takes shape, instantly transforming into the soul of a vast, scaly, and formidable Eastern dragon. The crimson dragon spirit spirals upward from the spear's tip, opening its fanged maw and unleashing a silent yet earth-shattering roar. The entire scene is presented in an epic, dark mythology art style, dominated by dark grey, black, and ochre. The only strong light source is the destructive crimson dragon spirit, which sharply outlines the warrior's silhouette and casts a massive, distorted shadow behind him, creating an atmosphere of power, mystery, and impending devastating battle.", + "seed": 18004, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_2/index_442_seed_18004_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10605476526543497, + "vlm_logprob": -0.10605476526543497, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.10605476526543497, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3967103064060211 + ], + "opposite_logprob_tokens": [ + -1.1467102766036987 + ], + "logprob_reward": -0.3967103064060211, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3967103064060211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1342894732952118, + 0.28207457065582275, + 0.02798032946884632, + 0.050597429275512695, + 0.031762465834617615, + 0.3430652320384979, + 0.01351633109152317, + 0.08816062659025192, + 0.022374490275979042, + 0.0061790961772203445 + ], + "credit_map": [ + 0.0345807820558548, + 0.007096773479133844, + 0.009890693239867687, + 0.005766666494309902, + 0.006667216774076223, + 0.0440840944647789, + 0.016869422048330307, + 0.009333822876214981, + 0.012690417468547821, + 0.006977367214858532, + 0.017146404832601547, + 0.015451929531991482, + 0.011917177587747574, + 0.011871212162077427, + 0.09940879791975021, + 0.10661126673221588, + 0.0015050808433443308, + 0.0020847690757364035, + 0.0031195420306175947, + 0.003709940705448389, + 0.002920042257755995, + 0.00482498062774539, + 0.0064957174472510815, + 0.0033202606718987226, + 0.007568940985947847, + 0.0020384504459798336, + 0.003594569396227598, + 0.0042725736275315285, + 0.004307099152356386, + 0.006634874269366264, + 0.017645733430981636, + 0.004535188432782888, + 0.003699538065120578, + 0.002479379065334797, + 0.0039499737322330475, + 0.004835064522922039, + 0.0039926269091665745, + 0.004105128813534975, + 0.004083752166479826, + 0.004617002326995134, + 0.02720595896244049, + 0.005756822414696217, + 0.14689850807189941, + 0.014948260970413685, + 0.023197093978524208, + 0.016350898891687393, + 0.09877415001392365, + 0.009933538734912872, + 0.0010007109958678484, + 0.001261828700080514, + 0.0018212952418252826, + 0.001655864529311657, + 0.0017634431133046746, + 0.0018704051617532969, + 0.0020163855515420437, + 0.0021263991948217154, + 0.0235337782651186, + 0.006903285626322031, + 0.026105765253305435, + 0.006741012446582317, + 0.0036338113714009523, + 0.007841851562261581, + 0.008263412863016129, + 0.00513771316036582, + 0.004537089727818966, + 0.0011716955341398716, + 0.006741358898580074, + 0.0024618504103273153, + 0.0019081536447629333, + 0.0018947310745716095, + 0.0018296571215614676, + 0.0018299547955393791, + 0.0011619485449045897, + 0.0004671069618780166, + 0.0013805418275296688, + 0.0007521030493080616, + 0.0005362981464713812, + 0.00038826707168482244, + 0.0010647798189893365, + 0.0004280511348042637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.3967103064060211 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016086198389530182 + ], + "opposite_logprob_tokens": [ + -4.141086101531982 + ], + "logprob_reward": -0.016086198389530182, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016086198389530182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04631518945097923, + 0.1409820020198822, + 0.041267842054367065, + 0.046417564153671265, + 0.039107684046030045, + 0.20684993267059326, + 0.032748639583587646, + 0.4010632634162903, + 0.018504973500967026, + 0.026742931455373764 + ], + "credit_map": [ + 0.004897326696664095, + 0.003329358994960785, + 0.00354307284578681, + 0.005464587826281786, + 0.002869140822440386, + 0.013441071845591068, + 0.0060578035190701485, + 0.00671282596886158, + 0.006813548039644957, + 0.005393177270889282, + 0.01224291417747736, + 0.009612408466637135, + 0.00917782261967659, + 0.00588257284834981, + 0.05305223539471626, + 0.03880733251571655, + 0.005464204121381044, + 0.004614426754415035, + 0.008027992211282253, + 0.004109473433345556, + 0.002959311241284013, + 0.006005549803376198, + 0.0042488230392336845, + 0.00583806075155735, + 0.004504572134464979, + 0.002279323060065508, + 0.007781692780554295, + 0.004238757770508528, + 0.004291939549148083, + 0.011169536970555782, + 0.007969658821821213, + 0.004182080272585154, + 0.0043414924293756485, + 0.00252877757884562, + 0.008594222366809845, + 0.006381492596119642, + 0.0025442412588745356, + 0.005680324975401163, + 0.00446508452296257, + 0.004572048783302307, + 0.014578074216842651, + 0.004370616748929024, + 0.08744237571954727, + 0.0094542745500803, + 0.015068991109728813, + 0.010288244113326073, + 0.05999755859375, + 0.005649789236485958, + 0.001982706366106868, + 0.0032957603689283133, + 0.0036789497826248407, + 0.004313585814088583, + 0.006066338159143925, + 0.0032761050388216972, + 0.007131507154554129, + 0.0030036892276257277, + 0.12143287062644958, + 0.0297416839748621, + 0.15532848238945007, + 0.016684917733073235, + 0.011536991223692894, + 0.03468994423747063, + 0.012953458353877068, + 0.018694935366511345, + 0.0029444743413478136, + 0.0012587738456204534, + 0.003040278097614646, + 0.002053098054602742, + 0.003716373350471258, + 0.0019079140620306134, + 0.0016858897870406508, + 0.0018981738248839974, + 0.004363521933555603, + 0.0024217746686190367, + 0.007286445237696171, + 0.0029749134555459023, + 0.0020067482255399227, + 0.002138479147106409, + 0.002763095311820507, + 0.0027879541739821434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.016086198389530182 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003015860915184 + ], + "opposite_logprob_tokens": [ + -2.350301504135132 + ], + "logprob_reward": -0.1003015860915184, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003015860915184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24928037822246552, + 0.5879158973693848, + 0.014697452075779438, + 0.020474983379244804, + 0.026001842692494392, + 0.03456601873040199, + 0.012024547904729843, + 0.028084101155400276, + 0.01957646943628788, + 0.0073783546686172485 + ], + "credit_map": [ + 0.0747784897685051, + 0.009629080072045326, + 0.015278216451406479, + 0.009068774990737438, + 0.011468839831650257, + 0.0838305875658989, + 0.021244153380393982, + 0.0239822156727314, + 0.0230975691229105, + 0.013200906105339527, + 0.032726436853408813, + 0.030461035668849945, + 0.02467268705368042, + 0.023699285462498665, + 0.22024302184581757, + 0.2198149710893631, + 0.0008331091958098114, + 0.001322590745985508, + 0.00211218255572021, + 0.0018823390128090978, + 0.0016566828126087785, + 0.0022566583938896656, + 0.002494719112291932, + 0.0021391708869487047, + 0.0026958216913044453, + 0.0012550139799714088, + 0.002684278879314661, + 0.0017250494565814734, + 0.0023802376817911863, + 0.002887572394683957, + 0.004970576614141464, + 0.0018764337291941047, + 0.0030750262085348368, + 0.002305439440533519, + 0.0035053943283855915, + 0.003588214283809066, + 0.0037356915418058634, + 0.0028527609538286924, + 0.003913325257599354, + 0.0030259923078119755, + 0.002535030012950301, + 0.0027084313333034515, + 0.005483050365000963, + 0.004185730125755072, + 0.003427645657211542, + 0.005332382395863533, + 0.0053702183067798615, + 0.005523530766367912, + 0.0011742733186110854, + 0.0007829563692212105, + 0.001399789471179247, + 0.0009579107281751931, + 0.002282924484461546, + 0.001713374862447381, + 0.0023078201338648796, + 0.0014054974308237433, + 0.0065633514896035194, + 0.0018314365297555923, + 0.008264713920652866, + 0.002140461467206478, + 0.0028733538929373026, + 0.0028237588703632355, + 0.0019994941540062428, + 0.0015875305980443954, + 0.003040145616978407, + 0.0012699090875685215, + 0.0034563804510980844, + 0.002388251945376396, + 0.003623650409281254, + 0.0021266955882310867, + 0.002055179560557008, + 0.0016162552637979388, + 0.001443609595298767, + 0.0005123575683683157, + 0.001455430407077074, + 0.0006749812164343894, + 0.0007997274515219033, + 0.0006774226203560829, + 0.001339404727332294, + 0.00047542087850160897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.1003015860915184 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637439269572496 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.004637439269572496, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637439269572496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15325020253658295, + 0.4688184857368469, + 0.017287535592913628, + 0.009563025087118149, + 0.011390257626771927, + 0.02374127134680748, + 0.012379460036754608, + 0.22447219491004944, + 0.04355515539646149, + 0.035542428493499756 + ], + "credit_map": [ + 0.044935207813978195, + 0.007521421182900667, + 0.009346526116132736, + 0.004901006817817688, + 0.005196300335228443, + 0.05675741285085678, + 0.013215241953730583, + 0.011377084068953991, + 0.018818186596035957, + 0.013089929707348347, + 0.03708716481924057, + 0.029647963121533394, + 0.021130908280611038, + 0.02075911872088909, + 0.1641061156988144, + 0.16417911648750305, + 0.001171695883385837, + 0.0017483483534306288, + 0.0031011265236884356, + 0.002309581730514765, + 0.0016627077711746097, + 0.002198999747633934, + 0.0017432990716770291, + 0.003351777559146285, + 0.0009165326482616365, + 0.0009198775514960289, + 0.0012386278249323368, + 0.0008905683062039316, + 0.0014188905479386449, + 0.0018511571688577533, + 0.0015438536647707224, + 0.0007835180731490254, + 0.0014405339024960995, + 0.0007372223190031946, + 0.0027406199369579554, + 0.0017046368448063731, + 0.0010921156499534845, + 0.001298390096053481, + 0.001280353986658156, + 0.0010963847162202, + 0.0019345200853422284, + 0.0010481164790689945, + 0.008259325288236141, + 0.0017874980112537742, + 0.0020955409854650497, + 0.0017644409090280533, + 0.005526803433895111, + 0.001325025805272162, + 0.0012189402477815747, + 0.0006746696890331805, + 0.00179323460906744, + 0.0016855044523254037, + 0.0015786892035976052, + 0.0016016765730455518, + 0.0014787386171519756, + 0.002348006470128894, + 0.07014412432909012, + 0.017771927639842033, + 0.0881388932466507, + 0.009312666952610016, + 0.005544153042137623, + 0.01881222426891327, + 0.004870238713920116, + 0.009877968579530716, + 0.006177529692649841, + 0.0023677905555814505, + 0.007099732290953398, + 0.004168643616139889, + 0.008129735477268696, + 0.006695700343698263, + 0.004974100738763809, + 0.00394192012026906, + 0.00492880679666996, + 0.003782313084229827, + 0.003932852763682604, + 0.004668721463531256, + 0.0057188062928617, + 0.0034257841762155294, + 0.00598470913246274, + 0.0031004329212009907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.004637439269572496 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538296170532703 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538296170532703, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538296170532703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3462637662887573, + 0.247269406914711, + 0.013438615947961807, + 0.023646391928195953, + 0.008579028770327568, + 0.16154931485652924, + 0.009740324690937996, + 0.14018148183822632, + 0.02533002197742462, + 0.024001706391572952 + ], + "credit_map": [ + 0.10212012380361557, + 0.016175512224435806, + 0.021666117012500763, + 0.011496058665215969, + 0.011059590615332127, + 0.1270236223936081, + 0.030689766630530357, + 0.026032982394099236, + 0.010222839191555977, + 0.006292855367064476, + 0.015901291742920876, + 0.01453832071274519, + 0.010112208314239979, + 0.010533425025641918, + 0.0875505730509758, + 0.09211788326501846, + 0.0009183702059090137, + 0.0012149213580414653, + 0.002673372160643339, + 0.0015729119768366218, + 0.0012492062523961067, + 0.001953636994585395, + 0.001562274293974042, + 0.00229392247274518, + 0.0028039172757416964, + 0.0014182712184265256, + 0.0039712064899504185, + 0.002040331484749913, + 0.002255336381494999, + 0.005683590658009052, + 0.0031598603818565607, + 0.002313877223059535, + 0.001267512678168714, + 0.0005990599165670574, + 0.0010517460759729147, + 0.0013933664886280894, + 0.0008880930254235864, + 0.0013810384552925825, + 0.0009550705435685813, + 0.0010431422851979733, + 0.01219291239976883, + 0.0027771135792136192, + 0.06936535239219666, + 0.007268599234521389, + 0.010377868078649044, + 0.007501053623855114, + 0.047290220856666565, + 0.004776190035045147, + 0.0007537594647146761, + 0.0008899520034901798, + 0.0012206443352624774, + 0.0014747001696377993, + 0.0017075807554647326, + 0.001130025484599173, + 0.001436373800970614, + 0.0011272885603830218, + 0.04314076527953148, + 0.010911552235484123, + 0.054220378398895264, + 0.0062817479483783245, + 0.003740457119420171, + 0.011960593052208424, + 0.003523900406435132, + 0.006402085535228252, + 0.004334100987762213, + 0.0016085399547591805, + 0.0044201877899467945, + 0.0032284362241625786, + 0.003816421376541257, + 0.0022015906870365143, + 0.0034370615612715483, + 0.0022836821153759956, + 0.003229763126000762, + 0.0024580166209489107, + 0.004923456814140081, + 0.002973756520077586, + 0.003543427912518382, + 0.0019555557519197464, + 0.0030617995653301477, + 0.0018559303134679794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.012538296170532703 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic overhead shot zooms in, focusing on a warrior clad in black armor. He stands alone in the center of a dilapidated ancient battlefield. Broken weapons and shattered stone statues litter the ground. In the distance, steep mountains rise and fall, shrouded in inky mist. The warrior, a massive figure, clad in a thick, battle-scarred suit of black iron armor, engraved with intricate ancient cloud-thunder patterns. The armor reflects a cool metallic luster in the dim light of eternal twilight. In his right hand, he clutches a spear, also crafted from black iron and emblazoned with intricate dragon motifs. Steadily and deliberately, he executes a motion reminiscent of drawing a sword from its sheath, his body lowering slightly, his arm muscles tensing. At that moment, the dragon motif on the spear suddenly ignites, a fiery crimson glow emanating from the cracks in the carving. This immense energy rapidly condenses, stretches, and takes shape, instantly transforming into the soul of a vast, scaly, and formidable Eastern dragon. The crimson dragon spirit spirals upward from the spear's tip, opening its fanged maw and unleashing a silent yet earth-shattering roar. The entire scene is presented in an epic, dark mythology art style, dominated by dark grey, black, and ochre. The only strong light source is the destructive crimson dragon spirit, which sharply outlines the warrior's silhouette and casts a massive, distorted shadow behind him, creating an atmosphere of power, mystery, and impending devastating battle.", + "seed": 18005, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_2/index_442_seed_18005_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1530098281800747, + "vlm_logprob": -0.1530098281800747, + "vlm_yes_no_margin": 2.949999976158142, + "vlm_token_credit": -0.1530098281800747, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5894227027893066 + ], + "opposite_logprob_tokens": [ + -0.8394227027893066 + ], + "logprob_reward": -0.5894227027893066, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5894227027893066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13875335454940796, + 0.1538040190935135, + 0.2079690396785736, + 0.07324374467134476, + 0.08397047966718674, + 0.07153979688882828, + 0.0466187410056591, + 0.11816953122615814, + 0.0349084846675396, + 0.07102278620004654 + ], + "credit_map": [ + 0.01667836122214794, + 0.019569627940654755, + 0.024105707183480263, + 0.010130713693797588, + 0.016162458807229996, + 0.02110479585826397, + 0.016824757680296898, + 0.014176950789988041, + 0.04433244466781616, + 0.0067719630897045135, + 0.05247931554913521, + 0.015695340931415558, + 0.007502415683120489, + 0.00840917881578207, + 0.008108128793537617, + 0.01050523016601801, + 0.035623133182525635, + 0.013801918365061283, + 0.03140466660261154, + 0.03746965527534485, + 0.019278360530734062, + 0.014151161536574364, + 0.02944711595773697, + 0.02679305337369442, + 0.015069929882884026, + 0.003654171945527196, + 0.019335802644491196, + 0.010445436462759972, + 0.004482340533286333, + 0.005319525487720966, + 0.007229701150208712, + 0.007706841453909874, + 0.01125351432710886, + 0.006573426071554422, + 0.01173277385532856, + 0.01606708951294422, + 0.013097232207655907, + 0.008296779356896877, + 0.00822782889008522, + 0.008721844293177128, + 0.009761704131960869, + 0.004965068306773901, + 0.007818851619958878, + 0.012456894852221012, + 0.0073593053966760635, + 0.004641604609787464, + 0.013187956996262074, + 0.01134841050952673, + 0.005742882844060659, + 0.0050123510882258415, + 0.006463176570832729, + 0.0066747525706887245, + 0.004087594337761402, + 0.004299541935324669, + 0.007891528308391571, + 0.006446915678679943, + 0.0184425450861454, + 0.01532701961696148, + 0.015778806060552597, + 0.01238150056451559, + 0.0141631243750453, + 0.008853091858327389, + 0.015175793319940567, + 0.018047649413347244, + 0.0051956018432974815, + 0.0028538363985717297, + 0.007820334285497665, + 0.004608582239598036, + 0.0024855260271579027, + 0.002718435600399971, + 0.005246584303677082, + 0.003979585599154234, + 0.010507922619581223, + 0.008155752904713154, + 0.007075837813317776, + 0.007774206344038248, + 0.012812883593142033, + 0.012578283436596394, + 0.0070325881242752075, + 0.005085323471575975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.5894227027893066 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023313812911510468 + ], + "opposite_logprob_tokens": [ + -3.7733137607574463 + ], + "logprob_reward": -0.023313812911510468, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023313812911510468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1927715390920639, + 0.34056875109672546, + 0.08366226404905319, + 0.034614384174346924, + 0.05844702944159508, + 0.038750626146793365, + 0.028349274769425392, + 0.08359092473983765, + 0.09662780165672302, + 0.042617395520210266 + ], + "credit_map": [ + 0.034946028143167496, + 0.02003105729818344, + 0.031031662598252296, + 0.01751687005162239, + 0.01282647904008627, + 0.040968213230371475, + 0.014448123052716255, + 0.021003104746341705, + 0.08411923050880432, + 0.01876019313931465, + 0.10496014356613159, + 0.036714818328619, + 0.019932109862565994, + 0.024440160021185875, + 0.02207079529762268, + 0.029571285471320152, + 0.01160137914121151, + 0.006822475232183933, + 0.014733998104929924, + 0.014868528582155704, + 0.005811204668134451, + 0.005756079219281673, + 0.012036551721394062, + 0.01203205343335867, + 0.00439076591283083, + 0.0022607112769037485, + 0.004346420988440514, + 0.007600622251629829, + 0.004176346119493246, + 0.002719331067055464, + 0.005237597040832043, + 0.003882589051499963, + 0.011789178475737572, + 0.002908761380240321, + 0.014568930491805077, + 0.008346593007445335, + 0.005264580715447664, + 0.004370668437331915, + 0.005654136650264263, + 0.005544179584830999, + 0.00614666473120451, + 0.0033351934980601072, + 0.0048444499261677265, + 0.005453716032207012, + 0.003300582990050316, + 0.003162310691550374, + 0.005262832157313824, + 0.007244873326271772, + 0.0033154431730508804, + 0.0021077669225633144, + 0.003956913482397795, + 0.0036057536490261555, + 0.003345917444676161, + 0.0018366248114034534, + 0.0058968402445316315, + 0.004284017253667116, + 0.018041672185063362, + 0.00791156105697155, + 0.016062026843428612, + 0.0083418358117342, + 0.006351153366267681, + 0.008722464554011822, + 0.011028748005628586, + 0.007131465710699558, + 0.008694845251739025, + 0.015142478980123997, + 0.011624090373516083, + 0.015725506469607353, + 0.008089717477560043, + 0.006779811345040798, + 0.01872689463198185, + 0.011844453401863575, + 0.004824153613299131, + 0.006646560505032539, + 0.004869928117841482, + 0.006645261775702238, + 0.005427365656942129, + 0.0027979519218206406, + 0.006202565971761942, + 0.00520360516384244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.023313812911510468 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703976035118103 + ], + "opposite_logprob_tokens": [ + -2.127039670944214 + ], + "logprob_reward": -0.12703976035118103, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12703976035118103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08323819190263748, + 0.1296730786561966, + 0.060988254845142365, + 0.10624690353870392, + 0.1065007820725441, + 0.08590419590473175, + 0.05145454779267311, + 0.20982979238033295, + 0.1004461795091629, + 0.06571805477142334 + ], + "credit_map": [ + 0.010455409064888954, + 0.008338319137692451, + 0.012682951055467129, + 0.009767287410795689, + 0.010191167704761028, + 0.007701292634010315, + 0.014299288392066956, + 0.009802477434277534, + 0.026620913296937943, + 0.0069563258439302444, + 0.03589322790503502, + 0.012886900454759598, + 0.007810715585947037, + 0.010033455677330494, + 0.020606672391295433, + 0.008864867500960827, + 0.01642756536602974, + 0.0035659363493323326, + 0.012307994067668915, + 0.005679319612681866, + 0.005948319565504789, + 0.00472440617159009, + 0.007524796761572361, + 0.004809915088117123, + 0.02257852628827095, + 0.006917798426002264, + 0.027328407391905785, + 0.015953391790390015, + 0.008003713563084602, + 0.004597191233187914, + 0.012741952203214169, + 0.008125919848680496, + 0.01689467579126358, + 0.010110757313668728, + 0.016259117051959038, + 0.021078187972307205, + 0.009914355352520943, + 0.008297646418213844, + 0.012581122107803822, + 0.011364932172000408, + 0.01136906910687685, + 0.005542045459151268, + 0.009775837883353233, + 0.011997481808066368, + 0.011487680487334728, + 0.004476440139114857, + 0.0168578140437603, + 0.014397826977074146, + 0.008646373637020588, + 0.004311278462409973, + 0.006729927845299244, + 0.007659035734832287, + 0.00479571707546711, + 0.0055431826040148735, + 0.008288880810141563, + 0.005480150226503611, + 0.041029393672943115, + 0.023318897932767868, + 0.027866117656230927, + 0.029187597334384918, + 0.016928568482398987, + 0.01560670230537653, + 0.03096018359065056, + 0.02493232861161232, + 0.013115649111568928, + 0.010608132928609848, + 0.016067733988165855, + 0.013789040967822075, + 0.010797077789902687, + 0.007107883226126432, + 0.016337202861905098, + 0.012623454444110394, + 0.009759251959621906, + 0.0071044983342289925, + 0.007111117243766785, + 0.012419942766427994, + 0.0066992901265621185, + 0.006602558773010969, + 0.008704839274287224, + 0.0073165553621947765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.12703976035118103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107556652277708 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.01107556652277708, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107556652277708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06895014643669128, + 0.18284334242343903, + 0.0349486842751503, + 0.05084426328539848, + 0.09115616977214813, + 0.027515577152371407, + 0.059125714004039764, + 0.2089281678199768, + 0.21307425200939178, + 0.06261374056339264 + ], + "credit_map": [ + 0.011205112561583519, + 0.007352103013545275, + 0.011193415150046349, + 0.006173575296998024, + 0.007527131121605635, + 0.01110561192035675, + 0.005753147881478071, + 0.00864004623144865, + 0.045552145689725876, + 0.009351563639938831, + 0.06024374067783356, + 0.020326916128396988, + 0.0073745111003518105, + 0.01429771352559328, + 0.010990524664521217, + 0.014706209301948547, + 0.007886732928454876, + 0.0023763279896229506, + 0.004937837831676006, + 0.003930095117539167, + 0.004912500735372305, + 0.003061215626075864, + 0.004448410123586655, + 0.003395564155653119, + 0.012669206596910954, + 0.0024537371937185526, + 0.016960659995675087, + 0.004561471752822399, + 0.0027490996289998293, + 0.0024741841480135918, + 0.004065323155373335, + 0.004910578951239586, + 0.02424781396985054, + 0.005412336904555559, + 0.02744979038834572, + 0.009097740985453129, + 0.004599383100867271, + 0.004742929246276617, + 0.00739759998396039, + 0.008208571001887321, + 0.0035587723832577467, + 0.0023971518967300653, + 0.002899118233472109, + 0.0031970511190593243, + 0.005486709997057915, + 0.0022043532226234674, + 0.004215939901769161, + 0.0035564808640629053, + 0.0067930035293102264, + 0.005258803255856037, + 0.0072761713527143, + 0.008311568759381771, + 0.006278646178543568, + 0.004090188071131706, + 0.013662966899573803, + 0.007454365957528353, + 0.0496339313685894, + 0.020639125257730484, + 0.03645237535238266, + 0.01820240542292595, + 0.013201535679399967, + 0.025457322597503662, + 0.027396026998758316, + 0.017945457249879837, + 0.026808416470885277, + 0.02324005775153637, + 0.03364110738039017, + 0.024242913350462914, + 0.01856493018567562, + 0.023740513250231743, + 0.037126120179891586, + 0.025710195302963257, + 0.007161223795264959, + 0.00939841940999031, + 0.005994486156851053, + 0.01039299089461565, + 0.006053839810192585, + 0.006122179329395294, + 0.009325282648205757, + 0.008165315724909306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.01107556652277708 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419729832559824 + ], + "opposite_logprob_tokens": [ + -4.26419734954834 + ], + "logprob_reward": -0.01419729832559824, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419729832559824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11491400748491287, + 0.11444602161645889, + 0.051225725561380386, + 0.07169375568628311, + 0.04447916895151138, + 0.03948841616511345, + 0.0547245517373085, + 0.08770709484815598, + 0.33606430888175964, + 0.08525686711072922 + ], + "credit_map": [ + 0.015154572203755379, + 0.016137391328811646, + 0.02377965673804283, + 0.009361496195197105, + 0.012069101445376873, + 0.018665827810764313, + 0.010071633383631706, + 0.009674334898591042, + 0.023497004061937332, + 0.007184484042227268, + 0.02570640668272972, + 0.017713135108351707, + 0.009011187590658665, + 0.01228010468184948, + 0.008800845593214035, + 0.010252860374748707, + 0.006836917717009783, + 0.0037319380789995193, + 0.010207965038716793, + 0.006591860670596361, + 0.006996086798608303, + 0.004282134119421244, + 0.005718162748962641, + 0.006860661320388317, + 0.0099521204829216, + 0.005356071516871452, + 0.01197507232427597, + 0.006793695967644453, + 0.0178521778434515, + 0.004660679493099451, + 0.008757077157497406, + 0.006346868816763163, + 0.007146694231778383, + 0.0028493429999798536, + 0.006047557573765516, + 0.005218988750129938, + 0.011334773153066635, + 0.003038821741938591, + 0.004897167440503836, + 0.003945828881114721, + 0.0050083426758646965, + 0.0023727728985249996, + 0.004444267600774765, + 0.006511406507343054, + 0.006178978830575943, + 0.002348256064578891, + 0.006197665818035603, + 0.006426729261875153, + 0.007718463893979788, + 0.0043546599335968494, + 0.006806761026382446, + 0.008297158405184746, + 0.005246983375400305, + 0.0051557947881519794, + 0.010534326545894146, + 0.006610411684960127, + 0.018935954198241234, + 0.009629803709685802, + 0.012039796449244022, + 0.010376372374594212, + 0.006980182137340307, + 0.008506711572408676, + 0.011163407005369663, + 0.010074879974126816, + 0.05691156163811684, + 0.02981104888021946, + 0.048005908727645874, + 0.05016026645898819, + 0.019788868725299835, + 0.033799927681684494, + 0.0502038449048996, + 0.04738292470574379, + 0.008005142211914062, + 0.01655196212232113, + 0.009597215801477432, + 0.013952299952507019, + 0.010923247784376144, + 0.004451556596904993, + 0.009735199622809887, + 0.012040246278047562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.01419729832559824 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic overhead shot zooms in, focusing on a warrior clad in black armor. He stands alone in the center of a dilapidated ancient battlefield. Broken weapons and shattered stone statues litter the ground. In the distance, steep mountains rise and fall, shrouded in inky mist. The warrior, a massive figure, clad in a thick, battle-scarred suit of black iron armor, engraved with intricate ancient cloud-thunder patterns. The armor reflects a cool metallic luster in the dim light of eternal twilight. In his right hand, he clutches a spear, also crafted from black iron and emblazoned with intricate dragon motifs. Steadily and deliberately, he executes a motion reminiscent of drawing a sword from its sheath, his body lowering slightly, his arm muscles tensing. At that moment, the dragon motif on the spear suddenly ignites, a fiery crimson glow emanating from the cracks in the carving. This immense energy rapidly condenses, stretches, and takes shape, instantly transforming into the soul of a vast, scaly, and formidable Eastern dragon. The crimson dragon spirit spirals upward from the spear's tip, opening its fanged maw and unleashing a silent yet earth-shattering roar. The entire scene is presented in an epic, dark mythology art style, dominated by dark grey, black, and ochre. The only strong light source is the destructive crimson dragon spirit, which sharply outlines the warrior's silhouette and casts a massive, distorted shadow behind him, creating an atmosphere of power, mystery, and impending devastating battle.", + "seed": 18006, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_2/index_442_seed_18006_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1409922677092254, + "vlm_logprob": -0.1409922677092254, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.1409922677092254, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6453551650047302 + ], + "opposite_logprob_tokens": [ + -0.7703551650047302 + ], + "logprob_reward": -0.6453551650047302, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6453551650047302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15380790829658508, + 0.08722282946109772, + 0.05708272382616997, + 0.07872775942087173, + 0.10867246240377426, + 0.1056264191865921, + 0.2437051385641098, + 0.0680265799164772, + 0.05515822023153305, + 0.041969988495111465 + ], + "credit_map": [ + 0.02186151035130024, + 0.018783634528517723, + 0.02529114857316017, + 0.015413057990372181, + 0.017531011253595352, + 0.021312713623046875, + 0.014607406221330166, + 0.01900741644203663, + 0.010913019068539143, + 0.013463829644024372, + 0.01232047937810421, + 0.011573650874197483, + 0.008910030126571655, + 0.01075312402099371, + 0.011056219227612019, + 0.008232463151216507, + 0.006657162215560675, + 0.01495361514389515, + 0.005329564213752747, + 0.004957755096256733, + 0.006597047671675682, + 0.007649746257811785, + 0.006676706485450268, + 0.004261120222508907, + 0.01132109109312296, + 0.00840547401458025, + 0.011320831254124641, + 0.00767394807189703, + 0.01199751440435648, + 0.009901571087539196, + 0.009283885359764099, + 0.00882344413548708, + 0.008211003616452217, + 0.007881691679358482, + 0.0129766296595335, + 0.025224575772881508, + 0.008401699364185333, + 0.018042976036667824, + 0.01641344279050827, + 0.011520439758896828, + 0.00853098463267088, + 0.008356435224413872, + 0.01608872413635254, + 0.008974439464509487, + 0.010857714340090752, + 0.01551614049822092, + 0.02161276340484619, + 0.01568921096622944, + 0.02184355817735195, + 0.040522150695323944, + 0.030767707154154778, + 0.027864841744303703, + 0.0409165620803833, + 0.034829117357730865, + 0.026290735229849815, + 0.020670456811785698, + 0.0044282362796366215, + 0.004617475904524326, + 0.010858421213924885, + 0.004545176867395639, + 0.004924681968986988, + 0.021850448101758957, + 0.01135911326855421, + 0.0054430183954536915, + 0.006378432270139456, + 0.006799081806093454, + 0.010283339768648148, + 0.008970165625214577, + 0.003868557745590806, + 0.00946365762501955, + 0.003784686792641878, + 0.005610298365354538, + 0.006969609763473272, + 0.005929357837885618, + 0.005241276230663061, + 0.005023850593715906, + 0.005054174922406673, + 0.004848501645028591, + 0.004926319699734449, + 0.003976892679929733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.6453551650047302 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009801003150641918 + ], + "opposite_logprob_tokens": [ + -4.634800910949707 + ], + "logprob_reward": -0.009801003150641918, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009801003150641918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18859238922595978, + 0.11590708047151566, + 0.10026033967733383, + 0.041584815829992294, + 0.15812252461910248, + 0.1593921184539795, + 0.09198638796806335, + 0.03717851638793945, + 0.057070598006248474, + 0.049905188381671906 + ], + "credit_map": [ + 0.016920942813158035, + 0.046052608639001846, + 0.014082036912441254, + 0.012635101564228535, + 0.015379427000880241, + 0.05623437836766243, + 0.012640026398003101, + 0.014647874981164932, + 0.01126366388052702, + 0.016125012189149857, + 0.015689948573708534, + 0.014159082435071468, + 0.013193938881158829, + 0.013507361523807049, + 0.015478209592401981, + 0.016489868983626366, + 0.012370424345135689, + 0.0241745226085186, + 0.014896017499268055, + 0.0063621592707931995, + 0.011314270086586475, + 0.017173178493976593, + 0.008319901302456856, + 0.005649874918162823, + 0.0047631761990487576, + 0.003950228448957205, + 0.006759102921932936, + 0.007304548751562834, + 0.0034264270216226578, + 0.005284737795591354, + 0.004894676618278027, + 0.005201923195272684, + 0.018900208175182343, + 0.009186402894556522, + 0.032276567071676254, + 0.021462850272655487, + 0.010364500805735588, + 0.023467110469937325, + 0.019460562616586685, + 0.02300434187054634, + 0.00840659812092781, + 0.012011871673166752, + 0.022925592958927155, + 0.022770706564188004, + 0.013985518366098404, + 0.02108169160783291, + 0.042297426611185074, + 0.0159127339720726, + 0.009857629425823689, + 0.01066119596362114, + 0.011662330478429794, + 0.012043947353959084, + 0.013983985409140587, + 0.014501413330435753, + 0.009421966969966888, + 0.00985393114387989, + 0.004296407103538513, + 0.002925647422671318, + 0.006451769266277552, + 0.0025969515554606915, + 0.003011209424585104, + 0.007334857247769833, + 0.007485562469810247, + 0.0030761128291487694, + 0.008439809083938599, + 0.007009606342762709, + 0.01258009858429432, + 0.008089524693787098, + 0.004101838916540146, + 0.006067517213523388, + 0.0045841834507882595, + 0.006198028102517128, + 0.006743318401277065, + 0.007167861331254244, + 0.006265569478273392, + 0.006876311730593443, + 0.005803998094052076, + 0.00479775620624423, + 0.005805306136608124, + 0.006445069797337055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.009801003150641918 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327992208302021 + ], + "opposite_logprob_tokens": [ + -3.773279905319214 + ], + "logprob_reward": -0.02327992208302021, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327992208302021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17224572598934174, + 0.09234104305505753, + 0.06386686116456985, + 0.16717121005058289, + 0.2287394106388092, + 0.08892641961574554, + 0.05568310618400574, + 0.07590445131063461, + 0.03538482263684273, + 0.019736960530281067 + ], + "credit_map": [ + 0.017317159101366997, + 0.026579104363918304, + 0.018960775807499886, + 0.014667226001620293, + 0.028583362698554993, + 0.031028887256979942, + 0.020171931013464928, + 0.014937272295355797, + 0.010450947098433971, + 0.010622449219226837, + 0.010633615776896477, + 0.010869275778532028, + 0.013794034719467163, + 0.010519381612539291, + 0.013575064018368721, + 0.01187627948820591, + 0.005559231154620647, + 0.014285117387771606, + 0.007317611947655678, + 0.0065168277360498905, + 0.008233752101659775, + 0.006681868340820074, + 0.008520415052771568, + 0.0067520346492528915, + 0.024168359115719795, + 0.014898826368153095, + 0.02802128903567791, + 0.015427772887051105, + 0.015315968543291092, + 0.019416064023971558, + 0.02374565601348877, + 0.026177290827035904, + 0.017490673810243607, + 0.016560973599553108, + 0.030999938026070595, + 0.024145139381289482, + 0.025387847796082497, + 0.03272237256169319, + 0.038015592843294144, + 0.04341688007116318, + 0.007823985069990158, + 0.00474193412810564, + 0.01830567605793476, + 0.01258939690887928, + 0.009396924637258053, + 0.013341464102268219, + 0.01330952811986208, + 0.009417511522769928, + 0.004998131189495325, + 0.009902063757181168, + 0.00579040264710784, + 0.00636790506541729, + 0.010064318776130676, + 0.007889525964856148, + 0.0051843165419995785, + 0.005486443173140287, + 0.004139814991503954, + 0.003573250724002719, + 0.011282660067081451, + 0.003515439573675394, + 0.004292835481464863, + 0.027851851657032967, + 0.01593339443206787, + 0.005315197631716728, + 0.007184464018791914, + 0.0048796553164720535, + 0.006352921016514301, + 0.004990364424884319, + 0.0021555640269070864, + 0.0028606883715838194, + 0.0024760179221630096, + 0.004485145676881075, + 0.0025390291120857, + 0.002708851359784603, + 0.0023133864160627127, + 0.002788443583995104, + 0.002327255206182599, + 0.001965524861589074, + 0.0019402564503252506, + 0.003154214471578598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.02327992208302021 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946444813162088 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946444813162088, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946444813162088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29246073961257935, + 0.054442837834358215, + 0.05876091122627258, + 0.06473289430141449, + 0.07553009688854218, + 0.029406461864709854, + 0.06685072183609009, + 0.10671932995319366, + 0.15949144959449768, + 0.0916045606136322 + ], + "credit_map": [ + 0.02239464409649372, + 0.08415327221155167, + 0.01522978488355875, + 0.01243780367076397, + 0.01749473810195923, + 0.11101815849542618, + 0.009676466695964336, + 0.020055867731571198, + 0.006252110470086336, + 0.005617228802293539, + 0.009385725483298302, + 0.0065623014234006405, + 0.0052369930781424046, + 0.007535202894359827, + 0.006633648183196783, + 0.007219634484499693, + 0.00815432146191597, + 0.01608112081885338, + 0.007823496125638485, + 0.003971586935222149, + 0.005481180734932423, + 0.00738518824800849, + 0.005586071405559778, + 0.004277947824448347, + 0.00585972098633647, + 0.005652975291013718, + 0.009943707846105099, + 0.0070708259008824825, + 0.010377069935202599, + 0.007525062654167414, + 0.009231635369360447, + 0.00907189678400755, + 0.008815481327474117, + 0.006616729311645031, + 0.012200333178043365, + 0.01022647600620985, + 0.005693934857845306, + 0.012706523761153221, + 0.0075010997243225574, + 0.01176952663809061, + 0.0022503803484141827, + 0.002876592567190528, + 0.004396611358970404, + 0.003224835032597184, + 0.004925351124256849, + 0.003955729305744171, + 0.003873520065099001, + 0.003903445787727833, + 0.006778727751225233, + 0.009080946445465088, + 0.008481667377054691, + 0.00873807817697525, + 0.00985450018197298, + 0.00956067442893982, + 0.0072523984126746655, + 0.00710372906178236, + 0.011045818217098713, + 0.008495824411511421, + 0.0176983792334795, + 0.009344377554953098, + 0.008973686955869198, + 0.02548382803797722, + 0.016241107136011124, + 0.009436320513486862, + 0.030910296365618706, + 0.02298833057284355, + 0.0309620164334774, + 0.023255543783307076, + 0.010054601356387138, + 0.011269401758909225, + 0.012103007175028324, + 0.017948251217603683, + 0.01353633776307106, + 0.015218039974570274, + 0.012623575516045094, + 0.01263523381203413, + 0.008272645063698292, + 0.006889939773827791, + 0.01094347145408392, + 0.011485325172543526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.005946444813162088 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057880349457264 + ], + "opposite_logprob_tokens": [ + -3.8955788612365723 + ], + "logprob_reward": -0.02057880349457264, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057880349457264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36016350984573364, + 0.06742019206285477, + 0.05039703845977783, + 0.09341953694820404, + 0.04660842940211296, + 0.03344271704554558, + 0.13911496102809906, + 0.07781358808279037, + 0.08142104744911194, + 0.050198994576931 + ], + "credit_map": [ + 0.027354387566447258, + 0.10761810094118118, + 0.017865853384137154, + 0.01520337350666523, + 0.026597999036312103, + 0.12943024933338165, + 0.012822822667658329, + 0.023270748555660248, + 0.006936034187674522, + 0.00814090110361576, + 0.008839807473123074, + 0.00832420028746128, + 0.009226077236235142, + 0.01167463231831789, + 0.008343118242919445, + 0.005935422610491514, + 0.006261386908590794, + 0.01212170161306858, + 0.007788621820509434, + 0.004046039190143347, + 0.004515710286796093, + 0.005964150186628103, + 0.005128076765686274, + 0.00457135122269392, + 0.010677308775484562, + 0.007281900849193335, + 0.01711096242070198, + 0.016088496893644333, + 0.00695697171613574, + 0.01106481533497572, + 0.011528775095939636, + 0.012710309587419033, + 0.0033403837587684393, + 0.00252381619066, + 0.006102059967815876, + 0.004947660956531763, + 0.005109983496367931, + 0.006497740279883146, + 0.011009234003722668, + 0.0070775519125163555, + 0.002250904915854335, + 0.002443623961880803, + 0.00515885790809989, + 0.004774110391736031, + 0.003413887694478035, + 0.004706820007413626, + 0.0075982725247740746, + 0.0030962377786636353, + 0.012701420113444328, + 0.02308783493936062, + 0.016665169969201088, + 0.019091110676527023, + 0.02426881156861782, + 0.01940123923122883, + 0.01330525241792202, + 0.010594123974442482, + 0.012964894995093346, + 0.008127126842737198, + 0.015391016378998756, + 0.009033206850290298, + 0.0067518893629312515, + 0.011830124072730541, + 0.007932338863611221, + 0.0057829893194139, + 0.021833887323737144, + 0.017840612679719925, + 0.00864022970199585, + 0.0076838149689137936, + 0.0042181676253676414, + 0.006434453185647726, + 0.006531373597681522, + 0.008238505572080612, + 0.006414591800421476, + 0.007207907736301422, + 0.006495160516351461, + 0.006634101737290621, + 0.006508493330329657, + 0.0038157496601343155, + 0.005487400572746992, + 0.007635591086000204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.02057880349457264 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic overhead shot zooms in, focusing on a warrior clad in black armor. He stands alone in the center of a dilapidated ancient battlefield. Broken weapons and shattered stone statues litter the ground. In the distance, steep mountains rise and fall, shrouded in inky mist. The warrior, a massive figure, clad in a thick, battle-scarred suit of black iron armor, engraved with intricate ancient cloud-thunder patterns. The armor reflects a cool metallic luster in the dim light of eternal twilight. In his right hand, he clutches a spear, also crafted from black iron and emblazoned with intricate dragon motifs. Steadily and deliberately, he executes a motion reminiscent of drawing a sword from its sheath, his body lowering slightly, his arm muscles tensing. At that moment, the dragon motif on the spear suddenly ignites, a fiery crimson glow emanating from the cracks in the carving. This immense energy rapidly condenses, stretches, and takes shape, instantly transforming into the soul of a vast, scaly, and formidable Eastern dragon. The crimson dragon spirit spirals upward from the spear's tip, opening its fanged maw and unleashing a silent yet earth-shattering roar. The entire scene is presented in an epic, dark mythology art style, dominated by dark grey, black, and ochre. The only strong light source is the destructive crimson dragon spirit, which sharply outlines the warrior's silhouette and casts a massive, distorted shadow behind him, creating an atmosphere of power, mystery, and impending devastating battle.", + "seed": 18007, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_2/index_442_seed_18007_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12936107045970857, + "vlm_logprob": -0.12936107045970857, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.12936107045970857, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5911837816238403 + ], + "opposite_logprob_tokens": [ + -0.8411837816238403 + ], + "logprob_reward": -0.5911837816238403, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5911837816238403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17373469471931458, + 0.08260277658700943, + 0.03390698879957199, + 0.20608092844486237, + 0.05295487493276596, + 0.062191907316446304, + 0.05943083018064499, + 0.07774724811315536, + 0.09122774004936218, + 0.16012197732925415 + ], + "credit_map": [ + 0.021805593743920326, + 0.022529298439621925, + 0.024436745792627335, + 0.020681845024228096, + 0.01856313832104206, + 0.02179248072206974, + 0.0211523175239563, + 0.022773288190364838, + 0.010387352667748928, + 0.008926164358854294, + 0.011141980066895485, + 0.010895675048232079, + 0.011424641124904156, + 0.00857978779822588, + 0.012518931180238724, + 0.008728248998522758, + 0.005353496875613928, + 0.0025434386916458607, + 0.004124316852539778, + 0.00506256939843297, + 0.003350392449647188, + 0.003569498658180237, + 0.004758649040013552, + 0.005144631024450064, + 0.012958877719938755, + 0.021030675619840622, + 0.012273120693862438, + 0.06685269623994827, + 0.020244289189577103, + 0.011295110918581486, + 0.05069837346673012, + 0.01072780229151249, + 0.0036222231574356556, + 0.005752294324338436, + 0.0049258205108344555, + 0.011284139938652515, + 0.005513520911335945, + 0.004597770515829325, + 0.011654322035610676, + 0.005604790989309549, + 0.004982701502740383, + 0.009051754139363766, + 0.006992440205067396, + 0.010377286933362484, + 0.0056112464517354965, + 0.005836097523570061, + 0.008155431598424911, + 0.01118496060371399, + 0.006185914855450392, + 0.007971055805683136, + 0.008681058883666992, + 0.01134465727955103, + 0.005327627994120121, + 0.00466896640136838, + 0.009388973005115986, + 0.005862582474946976, + 0.009870661422610283, + 0.00461272057145834, + 0.009847748093307018, + 0.013934849761426449, + 0.004837407264858484, + 0.01358471717685461, + 0.0047111790627241135, + 0.016347970813512802, + 0.010232272557914257, + 0.01061661820858717, + 0.01580079458653927, + 0.014229700900614262, + 0.009969272650778294, + 0.011824178509414196, + 0.010396027006208897, + 0.008158891461789608, + 0.01207429077476263, + 0.02476711943745613, + 0.02024143561720848, + 0.022426458075642586, + 0.023928843438625336, + 0.013374877162277699, + 0.024612607434391975, + 0.018696364015340805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.5911837816238403 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008656949736177921 + ], + "opposite_logprob_tokens": [ + -4.758656978607178 + ], + "logprob_reward": -0.008656949736177921, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008656949736177921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16712673008441925, + 0.1472306102514267, + 0.07759226858615875, + 0.13223406672477722, + 0.0744827538728714, + 0.1218971237540245, + 0.05297819897532463, + 0.054117072373628616, + 0.11428268998861313, + 0.05805853381752968 + ], + "credit_map": [ + 0.029340345412492752, + 0.016168741509318352, + 0.02398773655295372, + 0.0181047935038805, + 0.01597093604505062, + 0.029718196019530296, + 0.015554708428680897, + 0.018281254917383194, + 0.016690917313098907, + 0.027569392696022987, + 0.022227415814995766, + 0.017138421535491943, + 0.013635790906846523, + 0.01477288082242012, + 0.021883230656385422, + 0.013312545605003834, + 0.009185350500047207, + 0.007864048704504967, + 0.00991699006408453, + 0.01873299665749073, + 0.006581129506230354, + 0.0054132589139044285, + 0.01302158273756504, + 0.006876903120428324, + 0.011076312512159348, + 0.01260671392083168, + 0.01073192898184061, + 0.03752811998128891, + 0.014213647693395615, + 0.009657075628638268, + 0.02632690593600273, + 0.01009334996342659, + 0.005301015451550484, + 0.01314281951636076, + 0.007608422543853521, + 0.015169369988143444, + 0.007737493608146906, + 0.0052267275750637054, + 0.01479180809110403, + 0.005505082197487354, + 0.009799526073038578, + 0.01637835055589676, + 0.016379456967115402, + 0.025820713490247726, + 0.012414646334946156, + 0.00889936275780201, + 0.021602973341941833, + 0.010602077469229698, + 0.004724330268800259, + 0.010921286419034004, + 0.006420197896659374, + 0.007961888797581196, + 0.006871731951832771, + 0.004549307748675346, + 0.006631539668887854, + 0.004897912498563528, + 0.007376540917903185, + 0.005562826991081238, + 0.006604737602174282, + 0.00880309846252203, + 0.004454274661839008, + 0.006033722776919603, + 0.007620350923389196, + 0.007661512121558189, + 0.009877532720565796, + 0.018145974725484848, + 0.020390303805470467, + 0.0169124323874712, + 0.012437952682375908, + 0.010643874295055866, + 0.01299599651247263, + 0.012878617271780968, + 0.0041838595643639565, + 0.00848626159131527, + 0.010241210460662842, + 0.009972987696528435, + 0.005381557159125805, + 0.003549442859366536, + 0.011922705918550491, + 0.004320500884205103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.008656949736177921 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336940661072731 + ], + "opposite_logprob_tokens": [ + -3.40869402885437 + ], + "logprob_reward": -0.0336940661072731, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336940661072731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15526461601257324, + 0.06156453117728233, + 0.07789309322834015, + 0.05984950810670853, + 0.12304805219173431, + 0.1250450164079666, + 0.13323305547237396, + 0.0795341357588768, + 0.12666086852550507, + 0.05790717154741287 + ], + "credit_map": [ + 0.023056169971823692, + 0.019642136991024017, + 0.018542809411883354, + 0.019726866856217384, + 0.01940026879310608, + 0.01620951108634472, + 0.02470885030925274, + 0.013978001661598682, + 0.007039832416921854, + 0.01259792223572731, + 0.005171273835003376, + 0.006609431933611631, + 0.008935432881116867, + 0.004507787525653839, + 0.009397979825735092, + 0.007304871920496225, + 0.010407128371298313, + 0.007509530987590551, + 0.010865754447877407, + 0.01837724633514881, + 0.0076083107851445675, + 0.008636835031211376, + 0.009704981930553913, + 0.004783306270837784, + 0.006408702582120895, + 0.0071241650730371475, + 0.007285487372428179, + 0.01169828325510025, + 0.006932191085070372, + 0.006752294022589922, + 0.006993738003075123, + 0.00665464298799634, + 0.01046731136739254, + 0.01592748425900936, + 0.020112134516239166, + 0.013487547636032104, + 0.012737615965306759, + 0.016649872064590454, + 0.01574968360364437, + 0.01791640743613243, + 0.012799982912838459, + 0.00826218817383051, + 0.011078757233917713, + 0.027907194569706917, + 0.009500126354396343, + 0.009105843491852283, + 0.036304354667663574, + 0.010086574591696262, + 0.013634014874696732, + 0.017268650233745575, + 0.013256634585559368, + 0.018699467182159424, + 0.01964307390153408, + 0.013703734613955021, + 0.023013325408101082, + 0.01401415467262268, + 0.007443840615451336, + 0.006723964586853981, + 0.008461406454443932, + 0.01243300549685955, + 0.006678038276731968, + 0.014141043648123741, + 0.009734776802361012, + 0.013918058015406132, + 0.012286395765841007, + 0.018330717459321022, + 0.022463981062173843, + 0.019362162798643112, + 0.01145714707672596, + 0.017304524779319763, + 0.014666032046079636, + 0.010789903812110424, + 0.003930683713406324, + 0.011181083507835865, + 0.005103085655719042, + 0.00774156441912055, + 0.011627130210399628, + 0.0053285835310816765, + 0.007056414615362883, + 0.005938627291470766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.0336940661072731 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002196520334109664 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.002196520334109664, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002196520334109664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1107088178396225, + 0.049432650208473206, + 0.042656753212213516, + 0.032914839684963226, + 0.04741827771067619, + 0.07017993927001953, + 0.0607762448489666, + 0.1342593878507614, + 0.3798206150531769, + 0.07183242589235306 + ], + "credit_map": [ + 0.014909910969436169, + 0.01006383541971445, + 0.01525072194635868, + 0.012892994098365307, + 0.011232404038310051, + 0.016982875764369965, + 0.011698178946971893, + 0.017677899450063705, + 0.005081202369183302, + 0.0072431000880897045, + 0.005878225900232792, + 0.005589500069618225, + 0.007271960377693176, + 0.005141015630215406, + 0.007651655934751034, + 0.005575992166996002, + 0.004824223928153515, + 0.004869887605309486, + 0.005176298785954714, + 0.010863668285310268, + 0.003864216385409236, + 0.002809603698551655, + 0.0058588553220033646, + 0.0043900045566260815, + 0.0024987405631691217, + 0.005104497540742159, + 0.004201323725283146, + 0.005739554762840271, + 0.003565598279237747, + 0.004197103902697563, + 0.005015952046960592, + 0.002592068864032626, + 0.0035331472754478455, + 0.00547429034486413, + 0.005378495901823044, + 0.009938938543200493, + 0.004096814431250095, + 0.005359559785574675, + 0.01003800518810749, + 0.0035990304313600063, + 0.00792494136840105, + 0.007633185479789972, + 0.010289632715284824, + 0.011649939231574535, + 0.004630756564438343, + 0.008030104450881481, + 0.011664669960737228, + 0.008356711827218533, + 0.005723846144974232, + 0.01026842836290598, + 0.006112827453762293, + 0.009092183783650398, + 0.008151663467288017, + 0.006452341563999653, + 0.009186367504298687, + 0.0057885851711034775, + 0.016063250601291656, + 0.008235344663262367, + 0.01307227835059166, + 0.02286437898874283, + 0.009351889602839947, + 0.020443836227059364, + 0.020760878920555115, + 0.0234675332903862, + 0.030504480004310608, + 0.03964410722255707, + 0.05188822001218796, + 0.054633673280477524, + 0.0514250174164772, + 0.056241247802972794, + 0.0411517508327961, + 0.054332129657268524, + 0.00549196545034647, + 0.008784536272287369, + 0.008407332003116608, + 0.012228602543473244, + 0.008176677860319614, + 0.007385056000202894, + 0.013279839418828487, + 0.008078419603407383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.002196520334109664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that zooms in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a warrior wearing black iron armor with cloud-thunder patterns standing in the center of a dilapidated battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the warrior execute a deliberate motion of drawing a spear from its sheath while lowering his body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crimson dragon spirit ignite from the spear, spiral upward, and unleash a roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by dark grey, black, and ochre colors with the crimson dragon spirit as the only strong light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074034497141838 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011074034497141838, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074034497141838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11860956996679306, + 0.08469714969396591, + 0.056975286453962326, + 0.04121176153421402, + 0.0934082418680191, + 0.09716357290744781, + 0.05012872442603111, + 0.11412867903709412, + 0.2804982364177704, + 0.06317883729934692 + ], + "credit_map": [ + 0.01994303986430168, + 0.009054194204509258, + 0.013810582458972931, + 0.012895739637315273, + 0.008736821822822094, + 0.020871322602033615, + 0.011355209164321423, + 0.02194264344871044, + 0.00899309292435646, + 0.007772880140691996, + 0.016136132180690765, + 0.006732028443366289, + 0.005581006407737732, + 0.017462696880102158, + 0.013601744547486305, + 0.008417556993663311, + 0.006792648229748011, + 0.006353318691253662, + 0.008035359904170036, + 0.013125614263117313, + 0.0067209238186478615, + 0.004969138652086258, + 0.007174449507147074, + 0.0038038277998566628, + 0.004730152897536755, + 0.006197252310812473, + 0.0051096538081765175, + 0.004809210076928139, + 0.005623133387416601, + 0.005420983303338289, + 0.00540733290836215, + 0.003914037719368935, + 0.0064859045669436455, + 0.009695112705230713, + 0.012374104000627995, + 0.01716999150812626, + 0.007399044465273619, + 0.011175570078194141, + 0.01704370602965355, + 0.012064796872437, + 0.011810430325567722, + 0.008213548921048641, + 0.018615631386637688, + 0.016225649043917656, + 0.006446117535233498, + 0.009620307944715023, + 0.017441680654883385, + 0.008790191262960434, + 0.004497230518609285, + 0.0058340891264379025, + 0.005584350787103176, + 0.006632120348513126, + 0.006010073237121105, + 0.005960728973150253, + 0.007774563506245613, + 0.007835564203560352, + 0.011501851491630077, + 0.00890345312654972, + 0.013072502799332142, + 0.02051355317234993, + 0.01071970071643591, + 0.01305173896253109, + 0.01653510518372059, + 0.01983075588941574, + 0.026326965540647507, + 0.03613416850566864, + 0.04947826266288757, + 0.04121885076165199, + 0.037389613687992096, + 0.022712819278240204, + 0.04360542073845863, + 0.023632105439901352, + 0.006042547523975372, + 0.00741643225774169, + 0.007966126315295696, + 0.009830418974161148, + 0.007100981660187244, + 0.00755619490519166, + 0.011178326793015003, + 0.006087800953537226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.011074034497141838 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_3/index_tensor([5685, 5685])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_3/index_tensor([5685, 5685])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..698b0439cc886aa077b389208f684cbeb9fb9f4f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_3/index_tensor([5685, 5685])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion camera, filmed from a slightly upward angle, slowly approaches a massive white tiger, crouching in a languid yet powerful pose deep within an ancient, damp forest. Beneath it rests a massive, weathered rock, coated in a thick, moist, emerald green moss. In the background, layers of ferns and tangled vines loom in the light mist. The tiger's thick, snow-white fur gleams with a healthy sheen in the soft, diffused forest light, its inky black stripes flowing from head to tail outlining its muscular build. Its massive claws rest on one another, their sharp nails retracted. Its chest rises and falls gently with each deep, steady breath, its ears occasionally shifting alertly, picking up the subtlest sounds of the forest. The mottled sunlight penetrates the dense treetops, casting dynamic spots of light on it, and its sharp amber eyes are gazing majestically and quietly into the distance, as if in deep thought. The whole picture is filled with a tranquil and solemn atmosphere, showing a primitive and inviolable kingly demeanor.", + "seed": 10520, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_3/index_5685_seed_10520_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05863974429666996, + "vlm_logprob": -0.05863974429666996, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.05863974429666996, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00936602521687746 + ], + "opposite_logprob_tokens": [ + -5.259366035461426 + ], + "logprob_reward": -0.00936602521687746, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00936602521687746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08176624774932861, + 0.14758659899234772, + 0.04961543530225754, + 0.18672603368759155, + 0.25912415981292725, + 0.09695018827915192, + 0.03416122496128082, + 0.07483542710542679, + 0.03127321973443031, + 0.0379614531993866 + ], + "credit_map": [ + 0.009868362918496132, + 0.00622865604236722, + 0.013777375221252441, + 0.011830027215182781, + 0.008982326835393906, + 0.011246653273701668, + 0.011298062279820442, + 0.00853478442877531, + 0.009921363554894924, + 0.02143053524196148, + 0.02344454824924469, + 0.01905645616352558, + 0.03119245171546936, + 0.01982901245355606, + 0.013219494372606277, + 0.00949273630976677, + 0.005779844708740711, + 0.008124522864818573, + 0.0061103580519557, + 0.0073915934190154076, + 0.005518678575754166, + 0.005539142061024904, + 0.006281916983425617, + 0.004869379103183746, + 0.018010294064879417, + 0.02469358779489994, + 0.024041863158345222, + 0.017908388748764992, + 0.041748370975255966, + 0.016881652176380157, + 0.027990614995360374, + 0.015451266430318356, + 0.013861369341611862, + 0.011650445871055126, + 0.0873202532529831, + 0.014293013140559196, + 0.01196026336401701, + 0.009318151511251926, + 0.09730583429336548, + 0.01341486070305109, + 0.014934828504920006, + 0.013873573392629623, + 0.010855965316295624, + 0.011888659559190273, + 0.011372845619916916, + 0.013602822087705135, + 0.008698063902556896, + 0.011723444797098637, + 0.004668113309890032, + 0.004092318005859852, + 0.004027824383229017, + 0.004612477496266365, + 0.0037531426642090082, + 0.003762282198294997, + 0.0036615373101085424, + 0.005583532154560089, + 0.01168909389525652, + 0.01026550866663456, + 0.00923128891736269, + 0.009959317743778229, + 0.007041019853204489, + 0.005078534130007029, + 0.009388338774442673, + 0.012182328850030899, + 0.004200552124530077, + 0.0027980857994407415, + 0.0033968822099268436, + 0.00354453781619668, + 0.003991137258708477, + 0.004966307897120714, + 0.004998976364731789, + 0.003376740263774991, + 0.003235206240788102, + 0.003294892143458128, + 0.005272189155220985, + 0.0046399617567658424, + 0.0028661207761615515, + 0.004228015895932913, + 0.008876488544046879, + 0.005548582877963781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.00936602521687746 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157228410243988 + ], + "opposite_logprob_tokens": [ + -1.701572299003601 + ], + "logprob_reward": -0.20157228410243988, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157228410243988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06869150698184967, + 0.06090591102838516, + 0.0619829036295414, + 0.19597715139389038, + 0.3643134534358978, + 0.05564076080918312, + 0.046244557946920395, + 0.06581243127584457, + 0.04476745426654816, + 0.03566384315490723 + ], + "credit_map": [ + 0.006461811251938343, + 0.008632074110209942, + 0.011104041710495949, + 0.008854472078382969, + 0.008308962918817997, + 0.007973086088895798, + 0.0112794553861022, + 0.006077608559280634, + 0.0043375687673687935, + 0.008362437598407269, + 0.00840244721621275, + 0.011930500157177448, + 0.008214299567043781, + 0.0071039809845387936, + 0.007093466818332672, + 0.005461215041577816, + 0.004999070893973112, + 0.006215258035808802, + 0.008536930195987225, + 0.007518945261836052, + 0.011784383095800877, + 0.008006604388356209, + 0.009582918137311935, + 0.005338801071047783, + 0.024272579699754715, + 0.02199457585811615, + 0.02920490689575672, + 0.019242463633418083, + 0.03825455904006958, + 0.020146075636148453, + 0.02693907916545868, + 0.01592293381690979, + 0.0247009489685297, + 0.03006342425942421, + 0.10216446220874786, + 0.026080675423145294, + 0.029995400458574295, + 0.023318495601415634, + 0.10685104131698608, + 0.021139048039913177, + 0.006190599408000708, + 0.007116471417248249, + 0.005988103337585926, + 0.006321226712316275, + 0.00567151652649045, + 0.010662836953997612, + 0.006767288316041231, + 0.006922723259776831, + 0.004475143738090992, + 0.00437413714826107, + 0.008961742743849754, + 0.006011218298226595, + 0.002979052485898137, + 0.0029653585515916348, + 0.009369335137307644, + 0.007108572870492935, + 0.0114030996337533, + 0.008998227305710316, + 0.009265700355172157, + 0.007180768996477127, + 0.005625686142593622, + 0.004782851319760084, + 0.008807038888335228, + 0.009749066084623337, + 0.004018131177872419, + 0.0027420634869486094, + 0.008274995721876621, + 0.0054130698554217815, + 0.0036948174238204956, + 0.004309541080147028, + 0.009536137804389, + 0.006778699811547995, + 0.003870789660140872, + 0.002476567169651389, + 0.005411973688751459, + 0.0029476757626980543, + 0.0033363248221576214, + 0.0036471406929194927, + 0.009006375446915627, + 0.004966997541487217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.20157228410243988 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625637110322714 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.007625637110322714, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625637110322714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15281479060649872, + 0.07020162791013718, + 0.08156245946884155, + 0.18624207377433777, + 0.2502894103527069, + 0.07635863125324249, + 0.04938415437936783, + 0.029617976397275925, + 0.0845227763056755, + 0.019006067886948586 + ], + "credit_map": [ + 0.014908555895090103, + 0.013693238608539104, + 0.02711649239063263, + 0.023305179551243782, + 0.016082556918263435, + 0.021964149549603462, + 0.018125979229807854, + 0.017618626356124878, + 0.007004571612924337, + 0.013100807555019855, + 0.01020108349621296, + 0.008419839665293694, + 0.013533387333154678, + 0.005692837294191122, + 0.0072934324853122234, + 0.004955665674060583, + 0.006848971359431744, + 0.0091982651501894, + 0.010945051908493042, + 0.01227525994181633, + 0.012116639874875546, + 0.009757237508893013, + 0.011241909116506577, + 0.009179126471281052, + 0.02175566554069519, + 0.03558860719203949, + 0.02017023228108883, + 0.024289876222610474, + 0.03092270717024803, + 0.015395663678646088, + 0.021458351984620094, + 0.016660964116454124, + 0.011845269240438938, + 0.00932250265032053, + 0.08238991349935532, + 0.013808980584144592, + 0.01065139751881361, + 0.009476217441260815, + 0.09481092542409897, + 0.017984217032790184, + 0.007024722173810005, + 0.008101671934127808, + 0.009066502563655376, + 0.005518820136785507, + 0.006891536992043257, + 0.015641778707504272, + 0.015032289549708366, + 0.009081315249204636, + 0.004418594762682915, + 0.0033866027370095253, + 0.009089150466024876, + 0.00522766076028347, + 0.0038096350617706776, + 0.004436151590198278, + 0.010185418650507927, + 0.008830943144857883, + 0.0032276439014822245, + 0.0028021857142448425, + 0.0057707251980900764, + 0.0037116259336471558, + 0.0029648758936673403, + 0.0023317812010645866, + 0.004499954637140036, + 0.004309182986617088, + 0.0048535531386733055, + 0.005209543742239475, + 0.015363887883722782, + 0.007651172112673521, + 0.005603704135864973, + 0.006560683716088533, + 0.02471902035176754, + 0.014561209827661514, + 0.002140931086614728, + 0.0013370625674724579, + 0.003160193096846342, + 0.001922291936352849, + 0.0014564882731065154, + 0.0013156321365386248, + 0.004839048255234957, + 0.0028344211168587208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.007625637110322714 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635184537619352 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.004635184537619352, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635184537619352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09265957027673721, + 0.12342675030231476, + 0.14101803302764893, + 0.30056115984916687, + 0.133979931473732, + 0.10901577770709991, + 0.024934975430369377, + 0.032885462045669556, + 0.022434117272496223, + 0.019084282219409943 + ], + "credit_map": [ + 0.007770756725221872, + 0.00826116744428873, + 0.014111977070569992, + 0.017253022640943527, + 0.011510864831507206, + 0.010087459348142147, + 0.013578993268311024, + 0.01008532289415598, + 0.008318732492625713, + 0.016135524958372116, + 0.01824081875383854, + 0.021528301760554314, + 0.023606572300195694, + 0.014398050494492054, + 0.01212285365909338, + 0.0090758902952075, + 0.014300959184765816, + 0.0163882989436388, + 0.01869826391339302, + 0.018267277628183365, + 0.020891886204481125, + 0.016471993178129196, + 0.02081662230193615, + 0.01518271118402481, + 0.04167560860514641, + 0.06070992723107338, + 0.033589862287044525, + 0.04056532308459282, + 0.04397672414779663, + 0.027647927403450012, + 0.03175520524382591, + 0.020640572533011436, + 0.0072260769084095955, + 0.00619716988876462, + 0.04016799107193947, + 0.009544098749756813, + 0.007736336439847946, + 0.005970441736280918, + 0.04820326715707779, + 0.008934530429542065, + 0.006935260724276304, + 0.0069200326688587666, + 0.015358404256403446, + 0.008766216225922108, + 0.010886081494390965, + 0.02234162762761116, + 0.02524743042886257, + 0.012560712173581123, + 0.0024290417786687613, + 0.0022915001027286053, + 0.004236551932990551, + 0.002553808968514204, + 0.0026811985298991203, + 0.0016772227827459574, + 0.005071690306067467, + 0.003993958700448275, + 0.002257469343021512, + 0.002031569601967931, + 0.004994748160243034, + 0.00335326767526567, + 0.002338564954698086, + 0.001826310297474265, + 0.01030800398439169, + 0.005775522906333208, + 0.0023761633783578873, + 0.0015651518478989601, + 0.0034200074151158333, + 0.0022259652614593506, + 0.002086447784677148, + 0.0028683971613645554, + 0.004824785050004721, + 0.0030671979766339064, + 0.002351114060729742, + 0.0017778786132112145, + 0.002498844638466835, + 0.0023860703222453594, + 0.0014683724148198962, + 0.0012200771598145366, + 0.004835870582610369, + 0.0025460522156208754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.004635184537619352 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0699995905160904 + ], + "opposite_logprob_tokens": [ + -2.6949996948242188 + ], + "logprob_reward": -0.0699995905160904, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0699995905160904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0799654871225357, + 0.19184011220932007, + 0.1116613894701004, + 0.06597709655761719, + 0.42455363273620605, + 0.028126588091254234, + 0.03693779557943344, + 0.013963976874947548, + 0.03134489059448242, + 0.015628982335329056 + ], + "credit_map": [ + 0.00753050809726119, + 0.013973967172205448, + 0.011761477217078209, + 0.007277944125235081, + 0.013002722524106503, + 0.007688663434237242, + 0.012703907676041126, + 0.006026305723935366, + 0.013356605544686317, + 0.02825087495148182, + 0.029925916343927383, + 0.02847510389983654, + 0.038846369832754135, + 0.023218797519803047, + 0.018367193639278412, + 0.011399263516068459, + 0.010312962345778942, + 0.01021424401551485, + 0.017555607482790947, + 0.0114461500197649, + 0.01899658516049385, + 0.012783595360815525, + 0.01691593788564205, + 0.013436319306492805, + 0.009279713965952396, + 0.013448704034090042, + 0.006818546447902918, + 0.009332080371677876, + 0.010102545842528343, + 0.006112045608460903, + 0.005699612200260162, + 0.005183848086744547, + 0.022088265046477318, + 0.02053235098719597, + 0.15347427129745483, + 0.021281655877828598, + 0.01939510926604271, + 0.01565505750477314, + 0.15150222182273865, + 0.020624728873372078, + 0.0034359158016741276, + 0.004332368262112141, + 0.0033091779332607985, + 0.0033923883456736803, + 0.002552272751927376, + 0.0033710303250700235, + 0.004149485845118761, + 0.0035839497577399015, + 0.003672743448987603, + 0.004120876081287861, + 0.004922687541693449, + 0.00517871743068099, + 0.004150810185819864, + 0.0026863287203013897, + 0.006796734407544136, + 0.005408897530287504, + 0.0013092255685478449, + 0.0010994705371558666, + 0.0021800759714096785, + 0.0022663543932139874, + 0.0011178840650245547, + 0.0013240447733551264, + 0.0027069959323853254, + 0.0019599259831011295, + 0.004491027444601059, + 0.001995599362999201, + 0.0045099868439137936, + 0.0034738227259367704, + 0.003009225009009242, + 0.00397944962605834, + 0.0057588014751672745, + 0.004126979038119316, + 0.001378765213303268, + 0.001044039730913937, + 0.002778033260256052, + 0.002132328925654292, + 0.0010282077128067613, + 0.0013044803636148572, + 0.003803687635809183, + 0.0021594420541077852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.0699995905160904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion camera, filmed from a slightly upward angle, slowly approaches a massive white tiger, crouching in a languid yet powerful pose deep within an ancient, damp forest. Beneath it rests a massive, weathered rock, coated in a thick, moist, emerald green moss. In the background, layers of ferns and tangled vines loom in the light mist. The tiger's thick, snow-white fur gleams with a healthy sheen in the soft, diffused forest light, its inky black stripes flowing from head to tail outlining its muscular build. Its massive claws rest on one another, their sharp nails retracted. Its chest rises and falls gently with each deep, steady breath, its ears occasionally shifting alertly, picking up the subtlest sounds of the forest. The mottled sunlight penetrates the dense treetops, casting dynamic spots of light on it, and its sharp amber eyes are gazing majestically and quietly into the distance, as if in deep thought. The whole picture is filled with a tranquil and solemn atmosphere, showing a primitive and inviolable kingly demeanor.", + "seed": 10521, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_3/index_5685_seed_10521_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.038842771342024206, + "vlm_logprob": -0.038842771342024206, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.038842771342024206, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01928461529314518 + ], + "opposite_logprob_tokens": [ + -4.144284725189209 + ], + "logprob_reward": -0.01928461529314518, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01928461529314518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04556792974472046, + 0.06494268029928207, + 0.25152337551116943, + 0.1269206553697586, + 0.3760070502758026, + 0.010892417281866074, + 0.01677693799138069, + 0.06180720403790474, + 0.020738784223794937, + 0.024822957813739777 + ], + "credit_map": [ + 0.006816729437559843, + 0.003393585095182061, + 0.0035867660772055387, + 0.005212002899497747, + 0.006849884986877441, + 0.006731981877237558, + 0.007068437058478594, + 0.005908544175326824, + 0.013704463839530945, + 0.0027027782052755356, + 0.005018820986151695, + 0.003054798813536763, + 0.0072899674996733665, + 0.01809394173324108, + 0.004830047022551298, + 0.010247870348393917, + 0.031525518745183945, + 0.011297534219920635, + 0.06353233754634857, + 0.016946783289313316, + 0.016033800318837166, + 0.054438237100839615, + 0.021010542288422585, + 0.03673863410949707, + 0.021961119025945663, + 0.006012209691107273, + 0.03484826907515526, + 0.009678248316049576, + 0.013220017775893211, + 0.017062349244952202, + 0.00876404345035553, + 0.015374386683106422, + 0.051707565784454346, + 0.028104061260819435, + 0.08643747121095657, + 0.030999774113297462, + 0.05523598566651344, + 0.0453086756169796, + 0.031149592250585556, + 0.04706393554806709, + 0.0011911336332559586, + 0.0015951234381645918, + 0.0010617191437631845, + 0.0014729853719472885, + 0.0014221675228327513, + 0.0014030211605131626, + 0.001202338607981801, + 0.0015439275885000825, + 0.002845808630809188, + 0.0014588452177122235, + 0.002659323625266552, + 0.0015417475951835513, + 0.0020230377558618784, + 0.0021112491376698017, + 0.002025179797783494, + 0.0021117469295859337, + 0.012220419943332672, + 0.0040883030742406845, + 0.007600255776196718, + 0.005641831085085869, + 0.007231688592582941, + 0.011932790279388428, + 0.00510985916480422, + 0.007982056587934494, + 0.0027822263073176146, + 0.0016292011132463813, + 0.0027563495095819235, + 0.002548615448176861, + 0.0036593559198081493, + 0.0024997082073241472, + 0.0020534421782940626, + 0.002809886122122407, + 0.002548754680901766, + 0.003104408271610737, + 0.0020616662222892046, + 0.0040214224718511105, + 0.004338642116636038, + 0.002973397960886359, + 0.002738530747592449, + 0.0030361367389559746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.01928461529314518 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900475710630417 + ], + "opposite_logprob_tokens": [ + -2.5790047645568848 + ], + "logprob_reward": -0.07900475710630417, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900475710630417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13757982850074768, + 0.22489039599895477, + 0.07246674597263336, + 0.04398481920361519, + 0.11084963381290436, + 0.043976616114377975, + 0.03783036023378372, + 0.10370876640081406, + 0.12944577634334564, + 0.09526705741882324 + ], + "credit_map": [ + 0.024816468358039856, + 0.011457406915724277, + 0.011117402464151382, + 0.01068964321166277, + 0.019555704668164253, + 0.024749593809247017, + 0.015926165506243706, + 0.019267451018095016, + 0.05263657495379448, + 0.008094669319689274, + 0.019921785220503807, + 0.013131246902048588, + 0.02132142148911953, + 0.05898038297891617, + 0.012455055490136147, + 0.03834925591945648, + 0.012604261748492718, + 0.005845000967383385, + 0.00784318521618843, + 0.006838249508291483, + 0.009355423972010612, + 0.014021423645317554, + 0.006193456705659628, + 0.00976574420928955, + 0.006222550757229328, + 0.0037964279763400555, + 0.006980346515774727, + 0.0032035200856626034, + 0.005880347918719053, + 0.0074735176749527454, + 0.005632378160953522, + 0.004795731510967016, + 0.017584754154086113, + 0.0071597034111619, + 0.02981490269303322, + 0.007651594467461109, + 0.010871886275708675, + 0.01425402145832777, + 0.009507096372544765, + 0.014005673117935658, + 0.004557468928396702, + 0.005999982822686434, + 0.0035779320169240236, + 0.005051980726420879, + 0.006551644299179316, + 0.005127310287207365, + 0.007180382031947374, + 0.005929914303123951, + 0.005274146795272827, + 0.003042943775653839, + 0.00794848520308733, + 0.0033641597256064415, + 0.003401519265025854, + 0.004422315396368504, + 0.005146164447069168, + 0.005230626557022333, + 0.025414831936359406, + 0.007505468092858791, + 0.011508376337587833, + 0.007780071813613176, + 0.008053229190409184, + 0.021215077489614487, + 0.006840849760919809, + 0.015390858985483646, + 0.0225385632365942, + 0.010183911770582199, + 0.02426222153007984, + 0.012548555620014668, + 0.011652439832687378, + 0.018872030079364777, + 0.013465914875268936, + 0.015922125428915024, + 0.011800611391663551, + 0.009519001469016075, + 0.014807166531682014, + 0.014243853278458118, + 0.012394854798913002, + 0.012197865173220634, + 0.009735023602843285, + 0.010568681173026562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.07900475710630417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820087829604745 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.002820087829604745, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820087829604745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11958096921443939, + 0.026480671018362045, + 0.04465628042817116, + 0.38503795862197876, + 0.026873352006077766, + 0.0235424991697073, + 0.044377174228429794, + 0.0996408686041832, + 0.036426715552806854, + 0.19338352978229523 + ], + "credit_map": [ + 0.028143884614109993, + 0.00618613138794899, + 0.016510605812072754, + 0.01245822198688984, + 0.007306565996259451, + 0.021840309724211693, + 0.0076118228025734425, + 0.019523421302437782, + 0.0028227996081113815, + 0.0018706271657720208, + 0.0035618802066892385, + 0.002821829169988632, + 0.004205416422337294, + 0.004616116173565388, + 0.0035285060293972492, + 0.003053495427593589, + 0.007630451582372189, + 0.0034191622398793697, + 0.006904738023877144, + 0.003534728428348899, + 0.005605633370578289, + 0.007930065505206585, + 0.003113287501037121, + 0.006518213544040918, + 0.06616123020648956, + 0.017129210755228996, + 0.11068657040596008, + 0.027765898033976555, + 0.033793751150369644, + 0.05436115711927414, + 0.02706078253686428, + 0.04807938262820244, + 0.004487304948270321, + 0.002080193255096674, + 0.005093551240861416, + 0.0018186039524152875, + 0.004137067589908838, + 0.003231742652133107, + 0.002877228194847703, + 0.0031476602889597416, + 0.0030924933962523937, + 0.002492325147613883, + 0.0023804083466529846, + 0.002892107702791691, + 0.0027210479602217674, + 0.0037979790940880775, + 0.002930399961769581, + 0.0032357368618249893, + 0.007040214259177446, + 0.004270036239176989, + 0.009340489283204079, + 0.00356270931661129, + 0.003865137230604887, + 0.005536269396543503, + 0.005251822527498007, + 0.0055104936473071575, + 0.019151078537106514, + 0.008485677652060986, + 0.010650248266756535, + 0.008318960666656494, + 0.010967032052576542, + 0.020612109452486038, + 0.007603280246257782, + 0.013852478936314583, + 0.00359925115481019, + 0.0024135252460837364, + 0.009107035584747791, + 0.005376252345740795, + 0.003676936263218522, + 0.003660255577415228, + 0.0038715119007974863, + 0.004721948876976967, + 0.024607311934232712, + 0.013121568597853184, + 0.048471421003341675, + 0.021063528954982758, + 0.024351056665182114, + 0.015345594845712185, + 0.019924994558095932, + 0.02649804949760437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.002820087829604745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040924143977463245 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.0040924143977463245, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040924143977463245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0721793994307518, + 0.14594730734825134, + 0.18675221502780914, + 0.10016577690839767, + 0.038980867713689804, + 0.029614273458719254, + 0.06881166249513626, + 0.1362101435661316, + 0.19568881392478943, + 0.025649582967162132 + ], + "credit_map": [ + 0.013651457615196705, + 0.004404694773256779, + 0.012809377163648605, + 0.009719851426780224, + 0.005488511174917221, + 0.010619567707180977, + 0.005366058088839054, + 0.010119874961674213, + 0.036003340035676956, + 0.00452116783708334, + 0.011768442578613758, + 0.0045748320408165455, + 0.012627198360860348, + 0.042110320180654526, + 0.008709005080163479, + 0.025632986798882484, + 0.02357547916471958, + 0.007867654785513878, + 0.04499170556664467, + 0.011526621878147125, + 0.013873274438083172, + 0.04279564693570137, + 0.014357239007949829, + 0.027764583006501198, + 0.01788811758160591, + 0.004563671536743641, + 0.029659191146492958, + 0.007404190953820944, + 0.007858483120799065, + 0.013013479299843311, + 0.007063180208206177, + 0.012715461663901806, + 0.005637422204017639, + 0.0022948060650378466, + 0.010086319409310818, + 0.0037115346640348434, + 0.004173718858510256, + 0.005192822311073542, + 0.0027519180439412594, + 0.00513232359662652, + 0.004499501083046198, + 0.0023204521276056767, + 0.003039092058315873, + 0.0022414885461330414, + 0.0032264236360788345, + 0.006756315007805824, + 0.003264280268922448, + 0.0042667207308113575, + 0.01127572450786829, + 0.006011049263179302, + 0.014750509522855282, + 0.0053963554091751575, + 0.006656141951680183, + 0.008581713773310184, + 0.008048263378441334, + 0.008091894909739494, + 0.02737479843199253, + 0.009762424044311047, + 0.01275057252496481, + 0.011253515258431435, + 0.016267629340291023, + 0.02942441776394844, + 0.010294140316545963, + 0.01908264495432377, + 0.016720978543162346, + 0.017660068348050117, + 0.01913467049598694, + 0.03069227561354637, + 0.04153675585985184, + 0.02089597098529339, + 0.020775282755494118, + 0.02827281504869461, + 0.003563716309145093, + 0.002023888984695077, + 0.004168826621025801, + 0.002842144574970007, + 0.0034107808023691177, + 0.003890936728566885, + 0.00251650158315897, + 0.003232784103602171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.0040924143977463245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901198208332062 + ], + "opposite_logprob_tokens": [ + -2.4640119075775146 + ], + "logprob_reward": -0.08901198208332062, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901198208332062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03463936224579811, + 0.21576860547065735, + 0.12029394507408142, + 0.1504932940006256, + 0.1039288267493248, + 0.014267658814787865, + 0.011969934217631817, + 0.15976582467556, + 0.1619037240743637, + 0.026968752965331078 + ], + "credit_map": [ + 0.004228189587593079, + 0.0024860904086381197, + 0.00568727683275938, + 0.003321247175335884, + 0.005192779004573822, + 0.005547748412936926, + 0.002953646704554558, + 0.00522238202393055, + 0.048392292112112045, + 0.007335983216762543, + 0.016119379550218582, + 0.008757787756621838, + 0.020513910800218582, + 0.06297897547483444, + 0.015483680181205273, + 0.03618660569190979, + 0.015367437154054642, + 0.004623660817742348, + 0.031374212354421616, + 0.007574989926069975, + 0.00823329109698534, + 0.02578936330974102, + 0.00967006478458643, + 0.017660925164818764, + 0.025562478229403496, + 0.00667891651391983, + 0.042394768446683884, + 0.01083272136747837, + 0.013349544256925583, + 0.02176598645746708, + 0.010976199060678482, + 0.018932679668068886, + 0.014540006406605244, + 0.007810976356267929, + 0.024451592937111855, + 0.008408238179981709, + 0.015241844579577446, + 0.01207912527024746, + 0.008352632634341717, + 0.013044415041804314, + 0.001722149201668799, + 0.002144747180864215, + 0.0016329791396856308, + 0.0015584417851641774, + 0.002236247295513749, + 0.001510733738541603, + 0.0015209371922537684, + 0.0019414238631725311, + 0.0014913628110662103, + 0.0013454825384542346, + 0.0014386497205123305, + 0.0012254125904291868, + 0.0014848177088424563, + 0.0018923786701634526, + 0.0015801318222656846, + 0.0015116976574063301, + 0.03283079341053963, + 0.011116031557321548, + 0.015184818767011166, + 0.012606482952833176, + 0.019066115841269493, + 0.03328651934862137, + 0.013491899706423283, + 0.022183164954185486, + 0.013346781022846699, + 0.014725885353982449, + 0.015399081632494926, + 0.02551880106329918, + 0.03482351452112198, + 0.01754874177277088, + 0.017048465088009834, + 0.023492449894547462, + 0.0032212003134191036, + 0.004178155213594437, + 0.0028798943385481834, + 0.004038588609546423, + 0.00316802435554564, + 0.0035144200082868338, + 0.002612355398014188, + 0.003356112167239189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.08901198208332062 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion camera, filmed from a slightly upward angle, slowly approaches a massive white tiger, crouching in a languid yet powerful pose deep within an ancient, damp forest. Beneath it rests a massive, weathered rock, coated in a thick, moist, emerald green moss. In the background, layers of ferns and tangled vines loom in the light mist. The tiger's thick, snow-white fur gleams with a healthy sheen in the soft, diffused forest light, its inky black stripes flowing from head to tail outlining its muscular build. Its massive claws rest on one another, their sharp nails retracted. Its chest rises and falls gently with each deep, steady breath, its ears occasionally shifting alertly, picking up the subtlest sounds of the forest. The mottled sunlight penetrates the dense treetops, casting dynamic spots of light on it, and its sharp amber eyes are gazing majestically and quietly into the distance, as if in deep thought. The whole picture is filled with a tranquil and solemn atmosphere, showing a primitive and inviolable kingly demeanor.", + "seed": 10522, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_3/index_5685_seed_10522_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0589534685947001, + "vlm_logprob": -0.0589534685947001, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.0589534685947001, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09069908410310745 + ], + "opposite_logprob_tokens": [ + -2.4656991958618164 + ], + "logprob_reward": -0.09069908410310745, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09069908410310745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07336395978927612, + 0.04473545029759407, + 0.022852005437016487, + 0.024885255843400955, + 0.05239654704928398, + 0.5563967823982239, + 0.04852304607629776, + 0.03934482857584953, + 0.0247513260692358, + 0.11275090277194977 + ], + "credit_map": [ + 0.01004734169691801, + 0.008117898367345333, + 0.007925689220428467, + 0.018143290653824806, + 0.005867329891771078, + 0.006198185030370951, + 0.009570671245455742, + 0.007493541110306978, + 0.006737418472766876, + 0.004394959658384323, + 0.0035576221998780966, + 0.004486690741032362, + 0.005148359574377537, + 0.0105024678632617, + 0.004106226842850447, + 0.00580170564353466, + 0.0036808575969189405, + 0.003272596513852477, + 0.004631454590708017, + 0.0018974941922351718, + 0.0023550342302769423, + 0.002976781688630581, + 0.0018757223151624203, + 0.0021620625630021095, + 0.004249679856002331, + 0.002970556728541851, + 0.003116040024906397, + 0.0020928627345710993, + 0.0030816784128546715, + 0.0033322919625788927, + 0.002368277171626687, + 0.003673865459859371, + 0.009828828275203705, + 0.004950047004967928, + 0.013869776390492916, + 0.004733962006866932, + 0.004581536166369915, + 0.0040266369469463825, + 0.004400321282446384, + 0.006005432922393084, + 0.23273953795433044, + 0.011567490175366402, + 0.236890509724617, + 0.014495604671537876, + 0.008019725792109966, + 0.009231030009686947, + 0.011118649505078793, + 0.032334208488464355, + 0.01167446281760931, + 0.004496995825320482, + 0.0077714333310723305, + 0.00522707449272275, + 0.004905570298433304, + 0.004218284040689468, + 0.005307323299348354, + 0.004921896383166313, + 0.008690998889505863, + 0.004120509140193462, + 0.005863040219992399, + 0.00510946661233902, + 0.0052450490184128284, + 0.0029223437886685133, + 0.0034887203946709633, + 0.0039046998135745525, + 0.0025791225489228964, + 0.004043648485094309, + 0.003075310494750738, + 0.002043918939307332, + 0.004043073859065771, + 0.002655940130352974, + 0.0031650837045162916, + 0.0031452267430722713, + 0.013944411650300026, + 0.015649082139134407, + 0.01411369163542986, + 0.008896292187273502, + 0.016130078583955765, + 0.013126065954566002, + 0.016756350174546242, + 0.014134916476905346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.09069908410310745 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428377628326416 + ], + "opposite_logprob_tokens": [ + -2.0178377628326416 + ], + "logprob_reward": -0.1428377628326416, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428377628326416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21676701307296753, + 0.10073042660951614, + 0.17762763798236847, + 0.06112006679177284, + 0.08094071596860886, + 0.18831558525562286, + 0.0398922823369503, + 0.027895845472812653, + 0.04310694709420204, + 0.06360355019569397 + ], + "credit_map": [ + 0.03743952140212059, + 0.02232258766889572, + 0.028003467246890068, + 0.026567233726382256, + 0.02588658407330513, + 0.025258399546146393, + 0.02693173661828041, + 0.02435746043920517, + 0.015955615788698196, + 0.010672839358448982, + 0.009301695041358471, + 0.016429346054792404, + 0.013094278052449226, + 0.013663193210959435, + 0.011388108134269714, + 0.010225346311926842, + 0.02659308724105358, + 0.02430953085422516, + 0.02291366457939148, + 0.016903886571526527, + 0.021788114681839943, + 0.02395862154662609, + 0.01893102563917637, + 0.022229695692658424, + 0.009240330196917057, + 0.00819275714457035, + 0.00957062654197216, + 0.005773083306849003, + 0.006304240319877863, + 0.00662525137886405, + 0.005985741503536701, + 0.009428032673895359, + 0.01348781306296587, + 0.010961305350065231, + 0.016725024208426476, + 0.006445815786719322, + 0.008321132510900497, + 0.007462351582944393, + 0.00896331388503313, + 0.008573958650231361, + 0.061288099735975266, + 0.008979410864412785, + 0.06228635087609291, + 0.016562258824706078, + 0.006470496766269207, + 0.006208675913512707, + 0.009852920658886433, + 0.016667358577251434, + 0.006052969489246607, + 0.00543131772428751, + 0.0055846949107944965, + 0.004198711831122637, + 0.004621879663318396, + 0.0040906136855483055, + 0.004937656223773956, + 0.004974433686584234, + 0.005668782163411379, + 0.0035105233546346426, + 0.0041958908550441265, + 0.0030877813696861267, + 0.0032456805929541588, + 0.002936630742624402, + 0.002533097518607974, + 0.002717457013204694, + 0.01027540024369955, + 0.0055066063068807125, + 0.006934709846973419, + 0.0038258032873272896, + 0.0040269712917506695, + 0.003363854018971324, + 0.005195406265556812, + 0.003978190943598747, + 0.009714802727103233, + 0.008711103349924088, + 0.008502948097884655, + 0.004855386447161436, + 0.008699645288288593, + 0.0068902308121323586, + 0.008668123744428158, + 0.007561307866126299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.1428377628326416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00219818577170372 + ], + "opposite_logprob_tokens": [ + -6.127198219299316 + ], + "logprob_reward": -0.00219818577170372, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.00219818577170372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08739028871059418, + 0.06824807822704315, + 0.14024852216243744, + 0.14313039183616638, + 0.03835523873567581, + 0.3701387047767639, + 0.03380963206291199, + 0.06448087841272354, + 0.013653901405632496, + 0.040544357150793076 + ], + "credit_map": [ + 0.014468153938651085, + 0.010966470465064049, + 0.015372143127024174, + 0.01460824254900217, + 0.005374881438910961, + 0.007887491956353188, + 0.00857654120773077, + 0.010136367753148079, + 0.012437681667506695, + 0.007699443958699703, + 0.010610763914883137, + 0.010114962235093117, + 0.00518348952755332, + 0.006924434565007687, + 0.006708343047648668, + 0.008568959310650826, + 0.02294367179274559, + 0.019695471972227097, + 0.02355978637933731, + 0.026059933006763458, + 0.009781521745026112, + 0.010377654805779457, + 0.013706390745937824, + 0.01412409357726574, + 0.0192869920283556, + 0.026023754850029945, + 0.0238741934299469, + 0.01323406770825386, + 0.014281824231147766, + 0.014144550077617168, + 0.013524804264307022, + 0.01876020058989525, + 0.0057722171768546104, + 0.005025455728173256, + 0.0070483568124473095, + 0.005078652407974005, + 0.0031898412853479385, + 0.0036138466093689203, + 0.002944905310869217, + 0.005681963637471199, + 0.1519658863544464, + 0.009091882035136223, + 0.15496966242790222, + 0.011761766858398914, + 0.005761347711086273, + 0.006404366344213486, + 0.008133175782859325, + 0.022050589323043823, + 0.006534990388900042, + 0.0032495560590177774, + 0.005864426027983427, + 0.0028848459478467703, + 0.0030252484139055014, + 0.0027545157354325056, + 0.0031794412061572075, + 0.006316605489701033, + 0.02159883640706539, + 0.005053990054875612, + 0.011990123428404331, + 0.010446333326399326, + 0.004345855675637722, + 0.0031219387892633677, + 0.003798489226028323, + 0.00412531103938818, + 0.0038278386928141117, + 0.0015087348874658346, + 0.0023941758554428816, + 0.001667099422775209, + 0.0009094345150515437, + 0.0010941451182588935, + 0.001216456526890397, + 0.0010360152227804065, + 0.006911139935255051, + 0.004690965171903372, + 0.006111483555287123, + 0.00338886515237391, + 0.0056082941591739655, + 0.004062083084136248, + 0.005824434570968151, + 0.003947091288864613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.00219818577170372 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040932451374828815 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.0040932451374828815, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040932451374828815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2587021291255951, + 0.21466608345508575, + 0.04189879074692726, + 0.09226127713918686, + 0.050588082522153854, + 0.19775888323783875, + 0.03275178745388985, + 0.044428713619709015, + 0.03256404772400856, + 0.03438028320670128 + ], + "credit_map": [ + 0.039371002465486526, + 0.030453136190772057, + 0.028715549036860466, + 0.07012816518545151, + 0.017504557967185974, + 0.01701655611395836, + 0.036532770842313766, + 0.01898038759827614, + 0.034064631909132004, + 0.03410225361585617, + 0.018733033910393715, + 0.034017227590084076, + 0.01614423654973507, + 0.022466909140348434, + 0.024075137451291084, + 0.031062645837664604, + 0.00664573023095727, + 0.00670498376712203, + 0.006032307632267475, + 0.0038084376137703657, + 0.005348746664822102, + 0.005803489591926336, + 0.004427904728800058, + 0.003127193311229348, + 0.014404073357582092, + 0.014267024584114552, + 0.020397966727614403, + 0.010717635974287987, + 0.009448843076825142, + 0.008618968538939953, + 0.00711398059502244, + 0.007292776834219694, + 0.009398176334798336, + 0.005572137422859669, + 0.013832345604896545, + 0.004921631421893835, + 0.004239856731146574, + 0.004246245604008436, + 0.0034078259486705065, + 0.004969866015017033, + 0.07763192057609558, + 0.0072179497219622135, + 0.08014266192913055, + 0.007252446375787258, + 0.00430135615170002, + 0.004538252949714661, + 0.005048053804785013, + 0.01162625290453434, + 0.007572392467409372, + 0.004229502286761999, + 0.004932181444019079, + 0.0047956411726772785, + 0.0031431501265615225, + 0.002236134372651577, + 0.0025939352344721556, + 0.003248847322538495, + 0.012063752859830856, + 0.003971590660512447, + 0.006972231436520815, + 0.007237117737531662, + 0.002982216188684106, + 0.0027145554777234793, + 0.002881902502849698, + 0.005605349317193031, + 0.00678605493158102, + 0.004315559286624193, + 0.005198948550969362, + 0.005235730670392513, + 0.0025838464498519897, + 0.0024299549404531717, + 0.0032458356581628323, + 0.002768119564279914, + 0.006019238382577896, + 0.003567310981452465, + 0.005059858318418264, + 0.003726138034835458, + 0.004159558564424515, + 0.0030909229535609484, + 0.005130250938236713, + 0.0036270059645175934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0040932451374828815 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054939065128564835 + ], + "opposite_logprob_tokens": [ + -2.929939031600952 + ], + "logprob_reward": -0.054939065128564835, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054939065128564835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17937123775482178, + 0.16221584379673004, + 0.08681832253932953, + 0.05345599725842476, + 0.039220307022333145, + 0.2172725945711136, + 0.10366862267255783, + 0.0934688001871109, + 0.022769594565033913, + 0.04173871874809265 + ], + "credit_map": [ + 0.025833917781710625, + 0.015982547774910927, + 0.022931639105081558, + 0.031074145808815956, + 0.01977006159722805, + 0.022708989679813385, + 0.024803154170513153, + 0.016266772523522377, + 0.024591179564595222, + 0.01673070155084133, + 0.015077645890414715, + 0.015349765308201313, + 0.021069929003715515, + 0.03840073198080063, + 0.013866537250578403, + 0.01712934672832489, + 0.015140636824071407, + 0.012859039939939976, + 0.013920058496296406, + 0.009418264031410217, + 0.008005420677363873, + 0.010840755887329578, + 0.007652828004211187, + 0.008981305174529552, + 0.0076132286339998245, + 0.008405402302742004, + 0.007825500331819057, + 0.006091996096074581, + 0.00540105951949954, + 0.00574732618406415, + 0.005170412361621857, + 0.007201065309345722, + 0.0058541567996144295, + 0.0051150135695934296, + 0.006237993482500315, + 0.00434509851038456, + 0.0034682638943195343, + 0.0045130616053938866, + 0.004091970156878233, + 0.005594749003648758, + 0.08390825986862183, + 0.007632554043084383, + 0.08738100528717041, + 0.00793622899800539, + 0.005297182593494654, + 0.0051426151767373085, + 0.005739456973969936, + 0.014235268346965313, + 0.01725146360695362, + 0.014644371345639229, + 0.01432088389992714, + 0.01114470325410366, + 0.01327550783753395, + 0.008482175879180431, + 0.012398367747664452, + 0.012151142582297325, + 0.025297151878476143, + 0.009596332907676697, + 0.015697574242949486, + 0.013055216521024704, + 0.008734681643545628, + 0.006003835704177618, + 0.007203835062682629, + 0.007880165241658688, + 0.004166049417108297, + 0.003494057571515441, + 0.0031525157392024994, + 0.002267202129587531, + 0.0031072210986167192, + 0.0018298859940841794, + 0.002478980226442218, + 0.002273679245263338, + 0.005660119466483593, + 0.005293525289744139, + 0.0050428882241249084, + 0.0032557803206145763, + 0.005144702270627022, + 0.005040967371314764, + 0.004845831543207169, + 0.007454901002347469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.054939065128564835 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion camera, filmed from a slightly upward angle, slowly approaches a massive white tiger, crouching in a languid yet powerful pose deep within an ancient, damp forest. Beneath it rests a massive, weathered rock, coated in a thick, moist, emerald green moss. In the background, layers of ferns and tangled vines loom in the light mist. The tiger's thick, snow-white fur gleams with a healthy sheen in the soft, diffused forest light, its inky black stripes flowing from head to tail outlining its muscular build. Its massive claws rest on one another, their sharp nails retracted. Its chest rises and falls gently with each deep, steady breath, its ears occasionally shifting alertly, picking up the subtlest sounds of the forest. The mottled sunlight penetrates the dense treetops, casting dynamic spots of light on it, and its sharp amber eyes are gazing majestically and quietly into the distance, as if in deep thought. The whole picture is filled with a tranquil and solemn atmosphere, showing a primitive and inviolable kingly demeanor.", + "seed": 10523, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_3/index_5685_seed_10523_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04643463781103492, + "vlm_logprob": -0.04643463781103492, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.04643463781103492, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015718843787908554 + ], + "opposite_logprob_tokens": [ + -4.515718936920166 + ], + "logprob_reward": -0.015718843787908554, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.015718843787908554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11270468682050705, + 0.2050970047712326, + 0.10250809043645859, + 0.18985691666603088, + 0.0671435296535492, + 0.07077555358409882, + 0.14297538995742798, + 0.04709281399846077, + 0.0265098474919796, + 0.03533623367547989 + ], + "credit_map": [ + 0.016193510964512825, + 0.011767038144171238, + 0.013632598333060741, + 0.012545482255518436, + 0.022675257176160812, + 0.010357193648815155, + 0.011983317323029041, + 0.013550294563174248, + 0.04084545746445656, + 0.018368737772107124, + 0.025567423552274704, + 0.016003763303160667, + 0.03831132501363754, + 0.015473278239369392, + 0.02063717320561409, + 0.02988983504474163, + 0.018043069168925285, + 0.012183507904410362, + 0.019584491848945618, + 0.008834628388285637, + 0.014897682704031467, + 0.007202704902738333, + 0.009907836094498634, + 0.011854168958961964, + 0.04435177892446518, + 0.010883931070566177, + 0.024003589525818825, + 0.011831219308078289, + 0.01994413509964943, + 0.019284242764115334, + 0.015171773731708527, + 0.044386234134435654, + 0.010539158247411251, + 0.006424625404179096, + 0.010590652003884315, + 0.006320474203675985, + 0.012767496518790722, + 0.007407419849187136, + 0.007366921752691269, + 0.005726785399019718, + 0.01045258529484272, + 0.006644982844591141, + 0.008592034690082073, + 0.008650144562125206, + 0.014997191727161407, + 0.005964491982012987, + 0.007580050267279148, + 0.007894070819020271, + 0.01037986483424902, + 0.005729455966502428, + 0.04709877446293831, + 0.012587621808052063, + 0.007671247236430645, + 0.049697861075401306, + 0.004511621780693531, + 0.0052989511750638485, + 0.004015171434730291, + 0.0032750836107879877, + 0.010855750180780888, + 0.0025332155637443066, + 0.004390333313494921, + 0.013274258002638817, + 0.0032261833548545837, + 0.005522820167243481, + 0.003259536810219288, + 0.003580564633011818, + 0.002172244479879737, + 0.0022573345340788364, + 0.004284475930035114, + 0.0035694418475031853, + 0.003952640574425459, + 0.0034336084499955177, + 0.005045995116233826, + 0.002890736563131213, + 0.003031022846698761, + 0.0036585270427167416, + 0.004611889831721783, + 0.005478000734001398, + 0.0036948954220861197, + 0.006925166118890047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.015718843787908554 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270604431629181 + ], + "opposite_logprob_tokens": [ + -2.1270604133605957 + ], + "logprob_reward": -0.1270604431629181, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270604431629181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19317129254341125, + 0.05221985653042793, + 0.05089103430509567, + 0.09134380519390106, + 0.07494857907295227, + 0.09860485792160034, + 0.17492768168449402, + 0.03339093178510666, + 0.03551093116402626, + 0.19499097764492035 + ], + "credit_map": [ + 0.022627945989370346, + 0.010745581239461899, + 0.06316342949867249, + 0.015627702698111534, + 0.017608828842639923, + 0.009540501050651073, + 0.00910781230777502, + 0.04474949836730957, + 0.007068864069879055, + 0.0037230998277664185, + 0.0075486162677407265, + 0.004862764850258827, + 0.006821580696851015, + 0.00947867427021265, + 0.004318407271057367, + 0.008397852070629597, + 0.01073876116424799, + 0.003314543515443802, + 0.011521512642502785, + 0.005110460799187422, + 0.006773432716727257, + 0.0037330053746700287, + 0.0053487857803702354, + 0.004350538831204176, + 0.01947503350675106, + 0.007952013052999973, + 0.01042952574789524, + 0.006406905595213175, + 0.01178056187927723, + 0.007794688455760479, + 0.008571476675570011, + 0.018933609127998352, + 0.009971486404538155, + 0.006812178995460272, + 0.011652748100459576, + 0.007977860048413277, + 0.018036529421806335, + 0.00780624570325017, + 0.006730524357408285, + 0.005961012095212936, + 0.01673170179128647, + 0.009934400208294392, + 0.013179559260606766, + 0.013692330569028854, + 0.017760135233402252, + 0.009555840864777565, + 0.009502203203737736, + 0.008248692378401756, + 0.02111373282968998, + 0.007780017796903849, + 0.05424917861819267, + 0.012188704684376717, + 0.01011935155838728, + 0.05237874761223793, + 0.0063544209115207195, + 0.010743541643023491, + 0.003759195329621434, + 0.0030048759654164314, + 0.004202084615826607, + 0.004369343165308237, + 0.00455043651163578, + 0.004801676142960787, + 0.004980234429240227, + 0.003723091445863247, + 0.004567118361592293, + 0.0034787380136549473, + 0.003004479920491576, + 0.003076282562687993, + 0.005445695947855711, + 0.006287828553467989, + 0.004573040641844273, + 0.00507775042206049, + 0.022164247930049896, + 0.017134133726358414, + 0.015775665640830994, + 0.030845940113067627, + 0.03149520233273506, + 0.026150580495595932, + 0.022355031222105026, + 0.029070189222693443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.1270604431629181 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076242173090577126 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.0076242173090577126, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076242173090577126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12597890198230743, + 0.08594406396150589, + 0.046050067991018295, + 0.14071479439735413, + 0.01598595827817917, + 0.09846974909305573, + 0.38775190711021423, + 0.02906293235719204, + 0.045560337603092194, + 0.024481311440467834 + ], + "credit_map": [ + 0.0172237791121006, + 0.005453147925436497, + 0.03472019359469414, + 0.009825638495385647, + 0.009914188645780087, + 0.014748373068869114, + 0.006950895767658949, + 0.027142684906721115, + 0.01697695627808571, + 0.006009161937981844, + 0.014707095921039581, + 0.007816722616553307, + 0.009919106028974056, + 0.011914540082216263, + 0.005607177969068289, + 0.012993300333619118, + 0.006144202314317226, + 0.0050999680534005165, + 0.007864066399633884, + 0.004555635619908571, + 0.007710621226578951, + 0.0036876623053103685, + 0.004449042025953531, + 0.006538872141391039, + 0.03411395475268364, + 0.008715956471860409, + 0.01630234904587269, + 0.011329500935971737, + 0.014641644433140755, + 0.015149552375078201, + 0.008550729602575302, + 0.03191109746694565, + 0.0021450461354106665, + 0.0014140369603410363, + 0.001961382571607828, + 0.0017985926242545247, + 0.003721032990142703, + 0.001767967245541513, + 0.001505726482719183, + 0.0016721735009923577, + 0.01973765529692173, + 0.0094790980219841, + 0.012820183299481869, + 0.012338330037891865, + 0.017597397789359093, + 0.007882957346737385, + 0.00944161880761385, + 0.009172513149678707, + 0.03322218731045723, + 0.012897104024887085, + 0.1354285627603531, + 0.030910925939679146, + 0.015494372695684433, + 0.13847948610782623, + 0.0065817153081297874, + 0.014737551100552082, + 0.003664027899503708, + 0.0022925513330847025, + 0.003740152111276984, + 0.0024099983274936676, + 0.0038666047621518373, + 0.005576315801590681, + 0.003496066899970174, + 0.004017215222120285, + 0.004605667665600777, + 0.003962146118283272, + 0.01397786196321249, + 0.0021046195179224014, + 0.0021230182610452175, + 0.012534071691334248, + 0.0016231376212090254, + 0.004629812203347683, + 0.0032492594327777624, + 0.0024203055072575808, + 0.0018427643226459622, + 0.002256278647109866, + 0.0045094420202076435, + 0.0034080478362739086, + 0.002472885651513934, + 0.004322327673435211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.0076242173090577126 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028160461224615574 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.0028160461224615574, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028160461224615574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14096760749816895, + 0.08214588463306427, + 0.01662147417664528, + 0.029492352157831192, + 0.023274365812540054, + 0.07580264657735825, + 0.5664445757865906, + 0.017864491790533066, + 0.029710371047258377, + 0.017676232382655144 + ], + "credit_map": [ + 0.016829971224069595, + 0.006160257384181023, + 0.04652731865644455, + 0.008737259544432163, + 0.008414264768362045, + 0.008074401877820492, + 0.007280245888978243, + 0.038943879306316376, + 0.01872739940881729, + 0.005278321914374828, + 0.020469803363084793, + 0.0055444324389100075, + 0.008953381329774857, + 0.0069583929143846035, + 0.0064961654134094715, + 0.009717988781630993, + 0.0025826774071902037, + 0.001175155397504568, + 0.003007194958627224, + 0.0013542777160182595, + 0.0021964367479085922, + 0.0028395538683980703, + 0.001355723012238741, + 0.0021104575134813786, + 0.006791713181883097, + 0.0020904052071273327, + 0.0035388667602092028, + 0.0024839926045387983, + 0.0030022270511835814, + 0.002817728789523244, + 0.0026774739380925894, + 0.006089944392442703, + 0.004150683525949717, + 0.0020047577563673258, + 0.0032951601315289736, + 0.0019507550168782473, + 0.003496466437354684, + 0.003723566420376301, + 0.002157412702217698, + 0.002495564054697752, + 0.012566213496029377, + 0.009608075954020023, + 0.009651385247707367, + 0.010108543559908867, + 0.01203490886837244, + 0.006169470958411694, + 0.008349559269845486, + 0.007314490154385567, + 0.04933617636561394, + 0.01870930939912796, + 0.19726835191249847, + 0.044463951140642166, + 0.022611334919929504, + 0.2015952318906784, + 0.010080093517899513, + 0.02238013967871666, + 0.002263046335428953, + 0.0013723497977480292, + 0.003226673463359475, + 0.0012924150796607137, + 0.0017647076165303588, + 0.0038618165999650955, + 0.0019488149555400014, + 0.002134669339284301, + 0.0024370383471250534, + 0.002869619522243738, + 0.00806859228760004, + 0.0017317476449534297, + 0.0016806548228487372, + 0.007889851927757263, + 0.002558986423537135, + 0.002473879838362336, + 0.001494910684414208, + 0.001780626829713583, + 0.00127127964515239, + 0.0015704204561188817, + 0.0035421508364379406, + 0.002433702815324068, + 0.003679691581055522, + 0.0019034481374546885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0028160461224615574 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger with inky black stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger crouching in a pose with its massive claws resting on one another and sharp nails retracted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, weathered rock coated in thick emerald green moss visible beneath the tiger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the background feature layers of ferns and tangled vines in a damp, misty forest environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera filmed from a slightly upward angle and slowly approaching the tiger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895363867282867 + ], + "opposite_logprob_tokens": [ + -2.578953742980957 + ], + "logprob_reward": -0.07895363867282867, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895363867282867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07066335529088974, + 0.058550190180540085, + 0.04454899579286575, + 0.07186252623796463, + 0.041215166449546814, + 0.2993757128715515, + 0.24470971524715424, + 0.020223958417773247, + 0.0393504798412323, + 0.10949990153312683 + ], + "credit_map": [ + 0.007763813715428114, + 0.004904258530586958, + 0.015787065029144287, + 0.0046212696470320225, + 0.013558894395828247, + 0.004842214286327362, + 0.0058527435176074505, + 0.013333092443645, + 0.014245434664189816, + 0.0035779871977865696, + 0.014295709319412708, + 0.003655703505501151, + 0.006143603473901749, + 0.005230538547039032, + 0.0039297970943152905, + 0.0074714189395308495, + 0.008444622159004211, + 0.004421786870807409, + 0.009310776367783546, + 0.0033802352845668793, + 0.006510724779218435, + 0.003383834846317768, + 0.004362023901194334, + 0.004734990186989307, + 0.016617150977253914, + 0.00498540885746479, + 0.008707481436431408, + 0.005147513933479786, + 0.005273342598229647, + 0.007766928523778915, + 0.006394476629793644, + 0.01697022095322609, + 0.004759940318763256, + 0.003739861538633704, + 0.007584183476865292, + 0.0027949281502515078, + 0.00982831884175539, + 0.005110329948365688, + 0.004026195965707302, + 0.0033714082092046738, + 0.062151163816452026, + 0.029588943347334862, + 0.03846876323223114, + 0.03596701845526695, + 0.048501599580049515, + 0.026990221813321114, + 0.025567572563886642, + 0.032140422612428665, + 0.021594924852252007, + 0.0106512987986207, + 0.08082298934459686, + 0.019068416208028793, + 0.012982642278075218, + 0.08334100991487503, + 0.005790211260318756, + 0.0104582030326128, + 0.0016903738724067807, + 0.002315833466127515, + 0.0025041811168193817, + 0.0023351027630269527, + 0.003966948017477989, + 0.003568705404177308, + 0.00223565474152565, + 0.001607159385457635, + 0.0027451710775494576, + 0.004500207491219044, + 0.009759164415299892, + 0.0028648064471781254, + 0.004298923537135124, + 0.009871231392025948, + 0.0028630734886974096, + 0.002447904087603092, + 0.013331032358109951, + 0.01142998319119215, + 0.008018233813345432, + 0.012877220287919044, + 0.01735745742917061, + 0.015258245170116425, + 0.018793143332004547, + 0.012434592470526695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.07895363867282867 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_4/index_tensor([7169, 7169])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_4/index_tensor([7169, 7169])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a6c1273dbde84e3b38030d8ed04e231551f54a96 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_4/index_tensor([7169, 7169])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping long shot, the camera moving in smooth tracking motion, follows a hoverboard through the digital canyons of a futuristic city. Flanked by towering skyscrapers, their surfaces covered in massive holographic billboards and dynamic data waterfalls, countless aircraft of varying shapes weave through multi-tiered air traffic arteries. The sleek, dark carbon fiber composite hoverboard itself is a flowing light source, its surface and edges shimmering with neon blue and purple streams of binary code and complex geometric patterns. It silently accelerates, emitting a pale blue energy trail from its tail. Then, deftly, it tilts to the left, tracing a precise, smooth arc, effortlessly passing a bulkier, industrial-style hover truck next to it. The entire scene is meticulously rendered with the textures of high-fidelity CG, with physically accurate light and shadow reflecting off the cold metal and glass surfaces. The predominant color palette is a deep midnight blue, illuminated by the vibrant colors of the holographic advertisements, creating a dynamic atmosphere of high speed, awe, and a futuristic technological aesthetic.", + "seed": 30496, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_4/index_7169_seed_30496_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.020193961821496485, + "vlm_logprob": -0.020193961821496485, + "vlm_yes_no_margin": 3.9999999523162844, + "vlm_token_credit": -0.020193961821496485, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01989668607711792 + ], + "opposite_logprob_tokens": [ + -4.019896507263184 + ], + "logprob_reward": -0.01989668607711792, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01989668607711792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24181999266147614, + 0.05621204897761345, + 0.030070319771766663, + 0.045380450785160065, + 0.019785132259130478, + 0.04135182499885559, + 0.03224845603108406, + 0.4413204491138458, + 0.050631750375032425, + 0.0411795936524868 + ], + "credit_map": [ + 0.04245888069272041, + 0.02778938226401806, + 0.02815493568778038, + 0.02824144996702671, + 0.022716199979186058, + 0.028159838169813156, + 0.03787880018353462, + 0.026420515030622482, + 0.004047438036650419, + 0.004718869924545288, + 0.005342382937669754, + 0.005861587822437286, + 0.008382122963666916, + 0.011966748163104057, + 0.007702390663325787, + 0.008190508000552654, + 0.0029224823229014874, + 0.004046366550028324, + 0.0037781312130391598, + 0.002745091449469328, + 0.0025252695195376873, + 0.006298691034317017, + 0.0026585073210299015, + 0.0050957826897501945, + 0.006524034310132265, + 0.010706509463489056, + 0.004595656879246235, + 0.002949169836938381, + 0.009694898501038551, + 0.004825732670724392, + 0.0023946871515363455, + 0.003689764067530632, + 0.0018128880765289068, + 0.002942599356174469, + 0.0020713761914521456, + 0.002141506876796484, + 0.0025436817668378353, + 0.003325581783428788, + 0.0024942599702626467, + 0.0024532368406653404, + 0.005270063877105713, + 0.0073137530125677586, + 0.005121500696986914, + 0.003769720671698451, + 0.004607033915817738, + 0.007256328593939543, + 0.003161561908200383, + 0.004851860925555229, + 0.004425478633493185, + 0.005169272422790527, + 0.0024635237641632557, + 0.0024110060185194016, + 0.005400192923843861, + 0.005036641377955675, + 0.002282935194671154, + 0.005059405695647001, + 0.023880403488874435, + 0.03043234348297119, + 0.1777966469526291, + 0.027507362887263298, + 0.012188956141471863, + 0.013906278647482395, + 0.0097861522808671, + 0.14582230150699615, + 0.0035588156897574663, + 0.005819161422550678, + 0.006966074462980032, + 0.007237195502966642, + 0.006163957063108683, + 0.004024159163236618, + 0.008105832152068615, + 0.008756553754210472, + 0.007134883664548397, + 0.005827388726174831, + 0.004778996575623751, + 0.003292944049462676, + 0.0056214905343949795, + 0.008058695122599602, + 0.0040730563923716545, + 0.0023921371903270483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.01989668607711792 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369118645787239 + ], + "opposite_logprob_tokens": [ + -3.408691167831421 + ], + "logprob_reward": -0.03369118645787239, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369118645787239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27772650122642517, + 0.03774169832468033, + 0.04071320220828056, + 0.14200085401535034, + 0.13242077827453613, + 0.0847950354218483, + 0.05557749792933464, + 0.11225198209285736, + 0.058023255318403244, + 0.05874915421009064 + ], + "credit_map": [ + 0.046903010457754135, + 0.025395188480615616, + 0.0387728251516819, + 0.03191975876688957, + 0.02436833269894123, + 0.03466752916574478, + 0.04325956851243973, + 0.032440297305583954, + 0.0029903894755989313, + 0.0038345151115208864, + 0.004005791619420052, + 0.00410336721688509, + 0.005042353179305792, + 0.006951788440346718, + 0.006034351419657469, + 0.004779141396284103, + 0.005873026791960001, + 0.004772781394422054, + 0.005581957753747702, + 0.004621017258614302, + 0.004890525247901678, + 0.0054407184943556786, + 0.0050631179474294186, + 0.004470056388527155, + 0.019454965367913246, + 0.03355612978339195, + 0.012509962543845177, + 0.010363906621932983, + 0.02646614797413349, + 0.02221042476594448, + 0.008159411139786243, + 0.00927990023046732, + 0.01591171696782112, + 0.02081959694623947, + 0.014797958545386791, + 0.023419957607984543, + 0.015123959630727768, + 0.017470678314566612, + 0.013942547142505646, + 0.010934355668723583, + 0.010620874352753162, + 0.012043602764606476, + 0.009706137701869011, + 0.01056302897632122, + 0.009364808909595013, + 0.011583131738007069, + 0.010855864733457565, + 0.010057583451271057, + 0.007483852095901966, + 0.010445513762533665, + 0.005052622873336077, + 0.005228719208389521, + 0.007639528252184391, + 0.007442875765264034, + 0.0038503578398376703, + 0.008434026502072811, + 0.007808350957930088, + 0.011031911708414555, + 0.030697915703058243, + 0.010895984247326851, + 0.007642529904842377, + 0.008312389254570007, + 0.0068309796042740345, + 0.029031917452812195, + 0.0041905734688043594, + 0.007227207534015179, + 0.009326448664069176, + 0.008333900012075901, + 0.007715565152466297, + 0.005080689210444689, + 0.00914781168103218, + 0.007001059129834175, + 0.011767333373427391, + 0.005118536297231913, + 0.0055707297287881374, + 0.0061230286955833435, + 0.004295156802982092, + 0.012527995742857456, + 0.00783949438482523, + 0.005506876856088638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.03369118645787239 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018175598233938217 + ], + "opposite_logprob_tokens": [ + -4.0181756019592285 + ], + "logprob_reward": -0.018175598233938217, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018175598233938217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15087890625, + 0.11680188030004501, + 0.14545468986034393, + 0.11590644717216492, + 0.0429188497364521, + 0.11817615479230881, + 0.03723835572600365, + 0.10731436312198639, + 0.09146016091108322, + 0.07385028898715973 + ], + "credit_map": [ + 0.024907201528549194, + 0.01586708426475525, + 0.01611945778131485, + 0.014623185619711876, + 0.015345973894000053, + 0.017635691910982132, + 0.027117151767015457, + 0.019263146445155144, + 0.01006214041262865, + 0.010264878161251545, + 0.010950376279652119, + 0.008188929408788681, + 0.015256705693900585, + 0.03376283496618271, + 0.009198945946991444, + 0.019117062911391258, + 0.010143505409359932, + 0.019200898706912994, + 0.014138337224721909, + 0.008624122478067875, + 0.017010392621159554, + 0.045437246561050415, + 0.014217098243534565, + 0.016683073714375496, + 0.017633091658353806, + 0.02308737486600876, + 0.010954110883176327, + 0.006873009726405144, + 0.02400464378297329, + 0.015823468565940857, + 0.0073088896460831165, + 0.010221846401691437, + 0.00454058637842536, + 0.006862899754196405, + 0.0035044534597545862, + 0.004518799018114805, + 0.0049860356375575066, + 0.009337827563285828, + 0.004108829889446497, + 0.005059412680566311, + 0.015582064166665077, + 0.016321366652846336, + 0.01525490079075098, + 0.011299979873001575, + 0.015987155959010124, + 0.020569050684571266, + 0.011596128344535828, + 0.011565492488443851, + 0.00649136770516634, + 0.005512755364179611, + 0.004037566017359495, + 0.0032396342139691114, + 0.005104315932840109, + 0.006223008967936039, + 0.002148305531591177, + 0.004481395706534386, + 0.007419895380735397, + 0.01225428469479084, + 0.03454853594303131, + 0.009854411706328392, + 0.005435771308839321, + 0.004984251689165831, + 0.004778975620865822, + 0.028038227930665016, + 0.010511315427720547, + 0.009515584446489811, + 0.009374281391501427, + 0.01627757027745247, + 0.009343624114990234, + 0.008410634472966194, + 0.010984615422785282, + 0.017042528837919235, + 0.013104988262057304, + 0.009726895950734615, + 0.007551416289061308, + 0.0060174185782670975, + 0.007683972362428904, + 0.012586278840899467, + 0.010057906620204449, + 0.0071214004419744015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.018175598233938217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577752962708473 + ], + "opposite_logprob_tokens": [ + -3.8955776691436768 + ], + "logprob_reward": -0.020577752962708473, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577752962708473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1056787297129631, + 0.06033293157815933, + 0.07346341013908386, + 0.22910264134407043, + 0.03444376587867737, + 0.09088864922523499, + 0.06270316988229752, + 0.07306288182735443, + 0.04718833789229393, + 0.22313553094863892 + ], + "credit_map": [ + 0.01656721718609333, + 0.010194395668804646, + 0.014777163043618202, + 0.010838175192475319, + 0.009496049955487251, + 0.014330794103443623, + 0.017145169898867607, + 0.012329760007560253, + 0.004535256884992123, + 0.004449981264770031, + 0.00794942770153284, + 0.006067103706300259, + 0.00899647269397974, + 0.012362521141767502, + 0.007904108613729477, + 0.008068063296377659, + 0.005727785639464855, + 0.011116266250610352, + 0.009950317442417145, + 0.006959504447877407, + 0.006828757468611002, + 0.016208436340093613, + 0.008506843820214272, + 0.008165501989424229, + 0.040805842727422714, + 0.01868365705013275, + 0.021541010588407516, + 0.01750091277062893, + 0.05032472312450409, + 0.045359913259744644, + 0.012497680261731148, + 0.02238890528678894, + 0.003390474943444133, + 0.004528135526925325, + 0.004782599397003651, + 0.007290856447070837, + 0.003191329538822174, + 0.003281351877376437, + 0.005053676199167967, + 0.002925345441326499, + 0.012994536198675632, + 0.014877610839903355, + 0.013648527674376965, + 0.006760335061699152, + 0.011545672081410885, + 0.01563263311982155, + 0.006311546079814434, + 0.009117781184613705, + 0.009190298616886139, + 0.009031151421368122, + 0.004679450299590826, + 0.0038179983384907246, + 0.010240224190056324, + 0.011043121106922626, + 0.004357037600129843, + 0.010343887843191624, + 0.007167447358369827, + 0.008388897404074669, + 0.02237633801996708, + 0.004849928896874189, + 0.004758493509143591, + 0.004381109029054642, + 0.0030169300734996796, + 0.01812373474240303, + 0.00569543382152915, + 0.005666972603648901, + 0.006405035965144634, + 0.007087259087711573, + 0.005396902095526457, + 0.0049414485692977905, + 0.006464971229434013, + 0.005530316382646561, + 0.06516634672880173, + 0.019155947491526604, + 0.012755931355059147, + 0.011029538698494434, + 0.013076912611722946, + 0.07726962864398956, + 0.013055732473731041, + 0.011625489220023155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.020577752962708473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008628585375845432 + ], + "opposite_logprob_tokens": [ + -4.7586283683776855 + ], + "logprob_reward": -0.008628585375845432, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008628585375845432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26229092478752136, + 0.07166283577680588, + 0.0551198311150074, + 0.07809088379144669, + 0.03961982950568199, + 0.09863465279340744, + 0.05720505490899086, + 0.08105332404375076, + 0.139091357588768, + 0.11723125725984573 + ], + "credit_map": [ + 0.04148107394576073, + 0.02770218439400196, + 0.038329314440488815, + 0.027036121115088463, + 0.027351848781108856, + 0.033070482313632965, + 0.03704220429062843, + 0.030277715995907784, + 0.006220721174031496, + 0.005228531546890736, + 0.006867214106023312, + 0.008348885923624039, + 0.012896702624857426, + 0.010548163205385208, + 0.01194069441407919, + 0.009611918590962887, + 0.006145870313048363, + 0.008128859102725983, + 0.010652253404259682, + 0.005115555599331856, + 0.005417900625616312, + 0.005888390354812145, + 0.006192481145262718, + 0.007578516844660044, + 0.015355700626969337, + 0.010791115462779999, + 0.009004498831927776, + 0.003775697434321046, + 0.01844833418726921, + 0.00943782925605774, + 0.00419789319857955, + 0.007079813629388809, + 0.0029340211767703295, + 0.007539774291217327, + 0.0037239789962768555, + 0.00632616737857461, + 0.004004678688943386, + 0.008101516403257847, + 0.0037084019277244806, + 0.003281290177255869, + 0.01181795634329319, + 0.0193281639367342, + 0.011283871717751026, + 0.007760791108012199, + 0.014935724437236786, + 0.016608459874987602, + 0.007069461513310671, + 0.009830229915678501, + 0.008998328819870949, + 0.008968205191195011, + 0.004867753479629755, + 0.005097256973385811, + 0.008448132313787937, + 0.009385869838297367, + 0.0035997244995087385, + 0.007839782163500786, + 0.00851353071630001, + 0.013371829874813557, + 0.02027115970849991, + 0.008066277951002121, + 0.005611999426037073, + 0.004920464474707842, + 0.004257868975400925, + 0.016040194779634476, + 0.013414513319730759, + 0.01644117757678032, + 0.013908331282436848, + 0.022257432341575623, + 0.012173956260085106, + 0.009859487414360046, + 0.017413081601262093, + 0.03362337872385979, + 0.02478809654712677, + 0.009834232740104198, + 0.00797550193965435, + 0.00903447438031435, + 0.007270938251167536, + 0.037361811846494675, + 0.00971270352602005, + 0.011253505945205688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.008628585375845432 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping long shot, the camera moving in smooth tracking motion, follows a hoverboard through the digital canyons of a futuristic city. Flanked by towering skyscrapers, their surfaces covered in massive holographic billboards and dynamic data waterfalls, countless aircraft of varying shapes weave through multi-tiered air traffic arteries. The sleek, dark carbon fiber composite hoverboard itself is a flowing light source, its surface and edges shimmering with neon blue and purple streams of binary code and complex geometric patterns. It silently accelerates, emitting a pale blue energy trail from its tail. Then, deftly, it tilts to the left, tracing a precise, smooth arc, effortlessly passing a bulkier, industrial-style hover truck next to it. The entire scene is meticulously rendered with the textures of high-fidelity CG, with physically accurate light and shadow reflecting off the cold metal and glass surfaces. The predominant color palette is a deep midnight blue, illuminated by the vibrant colors of the holographic advertisements, creating a dynamic atmosphere of high speed, awe, and a futuristic technological aesthetic.", + "seed": 30497, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_4/index_7169_seed_30497_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.021193619817495346, + "vlm_logprob": -0.021193619817495346, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.021193619817495346, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031979840248823166 + ], + "opposite_logprob_tokens": [ + -3.53197979927063 + ], + "logprob_reward": -0.031979840248823166, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.031979840248823166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4410468339920044, + 0.04937934875488281, + 0.032173771411180496, + 0.020268907770514488, + 0.15586955845355988, + 0.060147061944007874, + 0.07938288897275925, + 0.0339617021381855, + 0.0777885913848877, + 0.04998141527175903 + ], + "credit_map": [ + 0.017920663580298424, + 0.012636294588446617, + 0.019765332341194153, + 0.018029727041721344, + 0.019602341577410698, + 0.02318740077316761, + 0.16142790019512177, + 0.1684771329164505, + 0.0074074119329452515, + 0.006234521511942148, + 0.00552753172814846, + 0.0047313142567873, + 0.006773932371288538, + 0.006246830802410841, + 0.006219691596925259, + 0.006238110829144716, + 0.0033484462182968855, + 0.002914266660809517, + 0.0036404032725840807, + 0.0032948218286037445, + 0.0050559984520077705, + 0.004126117564737797, + 0.005507512949407101, + 0.004286203067749739, + 0.002917446196079254, + 0.002019862411543727, + 0.0026192045770585537, + 0.0022438503801822662, + 0.0022987392731010914, + 0.0026935474015772343, + 0.0025390407536178827, + 0.002937217243015766, + 0.030045945197343826, + 0.02229963429272175, + 0.010746341198682785, + 0.012466385029256344, + 0.029341360554099083, + 0.019988805055618286, + 0.008001995272934437, + 0.022979069501161575, + 0.0063821496441960335, + 0.011617620475590229, + 0.006829061545431614, + 0.007825829088687897, + 0.0053961980156600475, + 0.005857208743691444, + 0.007423991337418556, + 0.008814996108412743, + 0.014331472106277943, + 0.015374653972685337, + 0.010373011231422424, + 0.011172785423696041, + 0.008789054118096828, + 0.005349493585526943, + 0.008919871412217617, + 0.005072546657174826, + 0.0036714880261570215, + 0.008249929174780846, + 0.004432243760675192, + 0.0036886371672153473, + 0.0027029479388147593, + 0.0020708520896732807, + 0.004695724695920944, + 0.004449875093996525, + 0.00899643823504448, + 0.012027465738356113, + 0.013229075819253922, + 0.017632408067584038, + 0.006869497708976269, + 0.00508052296936512, + 0.007233318407088518, + 0.006719859782606363, + 0.008106281980872154, + 0.005205465480685234, + 0.0077739437110722065, + 0.009500824846327305, + 0.005235607735812664, + 0.0036110193468630314, + 0.004127566237002611, + 0.006420703139156103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.031979840248823166 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343757286667824 + ], + "opposite_logprob_tokens": [ + -3.651343822479248 + ], + "logprob_reward": -0.026343757286667824, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343757286667824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4527154564857483, + 0.06537828594446182, + 0.04059203341603279, + 0.028455402702093124, + 0.0855017751455307, + 0.03416109085083008, + 0.05908050015568733, + 0.03808737173676491, + 0.085418201982975, + 0.11060990393161774 + ], + "credit_map": [ + 0.022294415161013603, + 0.014899041503667831, + 0.02122051827609539, + 0.01779904030263424, + 0.02254861779510975, + 0.02559574693441391, + 0.16248346865177155, + 0.16587458550930023, + 0.006882932502776384, + 0.010003208182752132, + 0.004436877556145191, + 0.004186437465250492, + 0.009837058372795582, + 0.007522038649767637, + 0.011626406572759151, + 0.010883326642215252, + 0.005825960077345371, + 0.004391975235193968, + 0.003414417849853635, + 0.004134311806410551, + 0.00599775742739439, + 0.005272052716463804, + 0.005516348406672478, + 0.006039209198206663, + 0.004160535987466574, + 0.003331708023324609, + 0.003100891597568989, + 0.0032775497529655695, + 0.0031234475318342447, + 0.0028540901839733124, + 0.004286493174731731, + 0.004320686683058739, + 0.009882490150630474, + 0.00898413173854351, + 0.006783317308872938, + 0.008439469151198864, + 0.015760203823447227, + 0.013415026478469372, + 0.00999982189387083, + 0.012237313203513622, + 0.0036658740136772394, + 0.005131904501467943, + 0.005444624926894903, + 0.005687153898179531, + 0.0027939684223383665, + 0.0029213018715381622, + 0.004919691476970911, + 0.003596570109948516, + 0.008310288190841675, + 0.008212846703827381, + 0.007078262511640787, + 0.009493553079664707, + 0.006799086928367615, + 0.005473003257066011, + 0.007153663784265518, + 0.006559793837368488, + 0.004205459728837013, + 0.00748129328712821, + 0.00618356978520751, + 0.004852063022553921, + 0.0036189397796988487, + 0.002700126962736249, + 0.006359845399856567, + 0.002686070278286934, + 0.008846932090818882, + 0.009536189958453178, + 0.009700375609099865, + 0.01587134227156639, + 0.009453527629375458, + 0.009129254147410393, + 0.012589839287102222, + 0.01029073353856802, + 0.011295846663415432, + 0.009810950607061386, + 0.012345334514975548, + 0.011492141522467136, + 0.018959470093250275, + 0.015574264340102673, + 0.012428473681211472, + 0.018703429028391838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.026343757286667824 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016059452667832375 + ], + "opposite_logprob_tokens": [ + -4.141059398651123 + ], + "logprob_reward": -0.016059452667832375, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016059452667832375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3539694547653198, + 0.08155906200408936, + 0.03804522007703781, + 0.039888445287942886, + 0.1494537740945816, + 0.14156341552734375, + 0.03547309711575508, + 0.07120738178491592, + 0.04010731726884842, + 0.04873281344771385 + ], + "credit_map": [ + 0.03585168346762657, + 0.02261742576956749, + 0.027358513325452805, + 0.03904883563518524, + 0.024586467072367668, + 0.026312610134482384, + 0.08538395911455154, + 0.09280993789434433, + 0.01074565015733242, + 0.012757780030369759, + 0.007403309922665358, + 0.00858577061444521, + 0.00936594046652317, + 0.01262578833848238, + 0.008731949143111706, + 0.011342877522110939, + 0.0036419357638806105, + 0.0030253701843321323, + 0.004115283489227295, + 0.005846226587891579, + 0.005067487247288227, + 0.004577231593430042, + 0.006500087212771177, + 0.005271595437079668, + 0.004586317110806704, + 0.00398576445877552, + 0.0033843396231532097, + 0.0036504282616078854, + 0.006744327023625374, + 0.005794146563857794, + 0.005310805048793554, + 0.006432315334677696, + 0.03245360776782036, + 0.018630124628543854, + 0.010272281244397163, + 0.011621832847595215, + 0.0235156137496233, + 0.01731683872640133, + 0.008314288221299648, + 0.02732919529080391, + 0.016980575397610664, + 0.021891964599490166, + 0.01584591343998909, + 0.018081778660416603, + 0.015178169123828411, + 0.012233424000442028, + 0.018042132258415222, + 0.023309458047151566, + 0.005148599855601788, + 0.004565390292555094, + 0.004621964879333973, + 0.0051391227170825005, + 0.004140547942370176, + 0.00399080291390419, + 0.004984558559954166, + 0.002882113214582205, + 0.00729084387421608, + 0.01981852389872074, + 0.00545476982370019, + 0.006112770177423954, + 0.007703522685915232, + 0.004644988104701042, + 0.012990021146833897, + 0.007191936019808054, + 0.0033806150313466787, + 0.004659062717109919, + 0.005159391555935144, + 0.01226147823035717, + 0.0037570479325950146, + 0.003643855219706893, + 0.0039912024512887, + 0.003254665993154049, + 0.005552442744374275, + 0.006578767206519842, + 0.006018530111759901, + 0.007359516806900501, + 0.007889905013144016, + 0.004595980513840914, + 0.004810497630387545, + 0.005927171558141708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.016059452667832375 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340622454881668 + ], + "opposite_logprob_tokens": [ + -3.6513407230377197 + ], + "logprob_reward": -0.026340622454881668, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340622454881668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35039621591567993, + 0.07246016710996628, + 0.07250726222991943, + 0.07024511694908142, + 0.06437020748853683, + 0.035447269678115845, + 0.0369938500225544, + 0.07223933935165405, + 0.15884579718112946, + 0.06649474799633026 + ], + "credit_map": [ + 0.03268738463521004, + 0.02375996671617031, + 0.033648937940597534, + 0.043521586805582047, + 0.02562696300446987, + 0.027145177125930786, + 0.07905454933643341, + 0.08495163172483444, + 0.009400081820786, + 0.009042629040777683, + 0.007915934547781944, + 0.008640662766993046, + 0.007792389020323753, + 0.008242560550570488, + 0.00949812214821577, + 0.01192778255790472, + 0.01365711260586977, + 0.006471909582614899, + 0.006706863641738892, + 0.007884539663791656, + 0.008352035656571388, + 0.010968666523694992, + 0.009010487236082554, + 0.009455650113523006, + 0.007409885060042143, + 0.006293291691690683, + 0.0056249531917274, + 0.014909139834344387, + 0.009072612039744854, + 0.00925907026976347, + 0.00805637612938881, + 0.009619785472750664, + 0.008984181098639965, + 0.007589534856379032, + 0.005767450667917728, + 0.009242710657417774, + 0.007352485787123442, + 0.006262634880840778, + 0.0087353540584445, + 0.010435855947434902, + 0.004560514818876982, + 0.005089741665869951, + 0.004348651506006718, + 0.004569605924189091, + 0.00479406351223588, + 0.0040063755586743355, + 0.0037519927136600018, + 0.004326323512941599, + 0.00483894906938076, + 0.0054490710608661175, + 0.0042677451856434345, + 0.006305434741079807, + 0.004427633713930845, + 0.0026676226407289505, + 0.005666468758136034, + 0.003370922990143299, + 0.006896005943417549, + 0.007209449540823698, + 0.012583017349243164, + 0.008475696668028831, + 0.0075568147003650665, + 0.00390663230791688, + 0.01702389121055603, + 0.008587834425270557, + 0.021937860175967216, + 0.023016871884465218, + 0.025158260017633438, + 0.026853879913687706, + 0.016702275723218918, + 0.013255264610052109, + 0.017262760549783707, + 0.014658637344837189, + 0.006521269213408232, + 0.007786715868860483, + 0.006915612146258354, + 0.007242381572723389, + 0.012118586339056492, + 0.007892938330769539, + 0.007362179458141327, + 0.010655066929757595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.026340622454881668 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005244426429271698 + ], + "opposite_logprob_tokens": [ + -5.255244255065918 + ], + "logprob_reward": -0.005244426429271698, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005244426429271698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3126574158668518, + 0.04634527862071991, + 0.043557558208703995, + 0.05533226579427719, + 0.06451920419931412, + 0.04790318384766579, + 0.08236212283372879, + 0.1757650077342987, + 0.09739481657743454, + 0.07416316866874695 + ], + "credit_map": [ + 0.05428125709295273, + 0.034336235374212265, + 0.04173519089818001, + 0.038255318999290466, + 0.025596201419830322, + 0.04010691121220589, + 0.03373683989048004, + 0.04460948705673218, + 0.004858597181737423, + 0.006979623809456825, + 0.00401208596304059, + 0.003733427496626973, + 0.005766424350440502, + 0.007138686720281839, + 0.007278041448444128, + 0.006578394211828709, + 0.004738690797239542, + 0.0035686444025486708, + 0.0035828673280775547, + 0.004609034396708012, + 0.007808694615960121, + 0.005583987571299076, + 0.008136066608130932, + 0.0055295745842158794, + 0.005029681138694286, + 0.005835028365254402, + 0.003549141576513648, + 0.008384373039007187, + 0.005929626990109682, + 0.008824115619063377, + 0.009188275784254074, + 0.008592020720243454, + 0.006770542357116938, + 0.006282861810177565, + 0.004525805823504925, + 0.009781471453607082, + 0.011045815423130989, + 0.009952457621693611, + 0.007659256923943758, + 0.008500990457832813, + 0.00559498043730855, + 0.007704340852797031, + 0.005328305996954441, + 0.008439941331744194, + 0.006152000278234482, + 0.003239906160160899, + 0.006015433464199305, + 0.005428273230791092, + 0.01650058664381504, + 0.016420844942331314, + 0.009932858869433403, + 0.010746382176876068, + 0.009939323179423809, + 0.005700357723981142, + 0.00829912256449461, + 0.004822643473744392, + 0.01922469586133957, + 0.016406670212745667, + 0.043421342968940735, + 0.018117666244506836, + 0.013092459179461002, + 0.0077827321365475655, + 0.03610249236226082, + 0.02161693386733532, + 0.011321390978991985, + 0.014513831585645676, + 0.017511621117591858, + 0.01828240603208542, + 0.009600200690329075, + 0.005968855693936348, + 0.01073502842336893, + 0.0094614801928401, + 0.007674746215343475, + 0.008538063615560532, + 0.010054205544292927, + 0.010645592585206032, + 0.01153680682182312, + 0.007431773468852043, + 0.00917928945273161, + 0.009102688170969486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.005244426429271698 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping long shot, the camera moving in smooth tracking motion, follows a hoverboard through the digital canyons of a futuristic city. Flanked by towering skyscrapers, their surfaces covered in massive holographic billboards and dynamic data waterfalls, countless aircraft of varying shapes weave through multi-tiered air traffic arteries. The sleek, dark carbon fiber composite hoverboard itself is a flowing light source, its surface and edges shimmering with neon blue and purple streams of binary code and complex geometric patterns. It silently accelerates, emitting a pale blue energy trail from its tail. Then, deftly, it tilts to the left, tracing a precise, smooth arc, effortlessly passing a bulkier, industrial-style hover truck next to it. The entire scene is meticulously rendered with the textures of high-fidelity CG, with physically accurate light and shadow reflecting off the cold metal and glass surfaces. The predominant color palette is a deep midnight blue, illuminated by the vibrant colors of the holographic advertisements, creating a dynamic atmosphere of high speed, awe, and a futuristic technological aesthetic.", + "seed": 30498, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_4/index_7169_seed_30498_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02357504181563854, + "vlm_logprob": -0.02357504181563854, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.02357504181563854, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011976477690041065 + ], + "opposite_logprob_tokens": [ + -4.63697624206543 + ], + "logprob_reward": -0.011976477690041065, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011976477690041065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2936469614505768, + 0.10972903668880463, + 0.11567211896181107, + 0.1212468221783638, + 0.054491255432367325, + 0.06310155987739563, + 0.07396723330020905, + 0.042637452483177185, + 0.06069089472293854, + 0.06481671333312988 + ], + "credit_map": [ + 0.03091866336762905, + 0.041378263384103775, + 0.047666389495134354, + 0.04246607795357704, + 0.03606361150741577, + 0.02430080994963646, + 0.03887176141142845, + 0.03198134899139404, + 0.011526011861860752, + 0.014237625524401665, + 0.014622495509684086, + 0.019499246031045914, + 0.013135124929249287, + 0.016057470813393593, + 0.01178748719394207, + 0.008863560855388641, + 0.008772327564656734, + 0.01244593970477581, + 0.01682281866669655, + 0.012634660117328167, + 0.013649790547788143, + 0.026124557480216026, + 0.013303727842867374, + 0.011918285861611366, + 0.01141444593667984, + 0.016674960032105446, + 0.018748372793197632, + 0.017108425498008728, + 0.018423890694975853, + 0.01400275994092226, + 0.013389323838055134, + 0.011484636925160885, + 0.005229478236287832, + 0.0068429457023739815, + 0.008219520561397076, + 0.008125812746584415, + 0.008593159727752209, + 0.004834399092942476, + 0.007025618106126785, + 0.005620316136628389, + 0.015108702704310417, + 0.005366260651499033, + 0.005999596323817968, + 0.005071362946182489, + 0.004597743973135948, + 0.012079891748726368, + 0.007316021714359522, + 0.007561979815363884, + 0.007803961168974638, + 0.011735957115888596, + 0.006338207982480526, + 0.013782105408608913, + 0.011771542020142078, + 0.006696837488561869, + 0.008571148850023746, + 0.007267468608915806, + 0.0050367494113743305, + 0.005269486457109451, + 0.005071809981018305, + 0.006621318403631449, + 0.005837469361722469, + 0.007946675643324852, + 0.003692899364978075, + 0.0031610403675585985, + 0.008028157986700535, + 0.008249467238783836, + 0.007434001658111811, + 0.005805888678878546, + 0.015419445000588894, + 0.006322260946035385, + 0.004159499891102314, + 0.005272164940834045, + 0.006658921949565411, + 0.006162532605230808, + 0.004934364464133978, + 0.005752764642238617, + 0.010685157030820847, + 0.010698231868445873, + 0.013093425892293453, + 0.006831309292465448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.011976477690041065 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232852790504694 + ], + "opposite_logprob_tokens": [ + -3.773285388946533 + ], + "logprob_reward": -0.0232852790504694, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232852790504694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1200370118021965, + 0.09625588357448578, + 0.054743438959121704, + 0.08851610869169235, + 0.07659859955310822, + 0.14146171510219574, + 0.22197450697422028, + 0.06165003776550293, + 0.05386446416378021, + 0.08489815145730972 + ], + "credit_map": [ + 0.011446856893599033, + 0.00942661240696907, + 0.016540652140975, + 0.022365473210811615, + 0.013331098482012749, + 0.01101146824657917, + 0.022906266152858734, + 0.013008601032197475, + 0.00968968030065298, + 0.007405078504234552, + 0.013015889562666416, + 0.011574623174965382, + 0.00892750546336174, + 0.020341781899333, + 0.012673638761043549, + 0.012627699412405491, + 0.008264531381428242, + 0.007682922761887312, + 0.007060663774609566, + 0.005551931913942099, + 0.005753562320023775, + 0.008061559870839119, + 0.00573126133531332, + 0.006637013051658869, + 0.00890713557600975, + 0.010608972050249577, + 0.016165466979146004, + 0.014689921401441097, + 0.008639439009130001, + 0.007439497392624617, + 0.011908088810741901, + 0.010157598182559013, + 0.007027884945273399, + 0.012521692551672459, + 0.009367043152451515, + 0.008555393666028976, + 0.0110155725851655, + 0.01220189779996872, + 0.008873874321579933, + 0.0070352498441934586, + 0.024100275710225105, + 0.013301359489560127, + 0.010174460709095001, + 0.018585193902254105, + 0.018243003636598587, + 0.015599971637129784, + 0.022359246388077736, + 0.01909821666777134, + 0.018806038424372673, + 0.028390949591994286, + 0.01814374513924122, + 0.051363952457904816, + 0.03940541669726372, + 0.01593119464814663, + 0.028892235830426216, + 0.021041003987193108, + 0.005084152799099684, + 0.006619812920689583, + 0.007830665446817875, + 0.014239275828003883, + 0.008928625844419003, + 0.007759233005344868, + 0.00643494725227356, + 0.004753326065838337, + 0.007708503864705563, + 0.007381506264209747, + 0.005328190512955189, + 0.006280592177063227, + 0.013536914251744747, + 0.005449852906167507, + 0.00405894173309207, + 0.004119968973100185, + 0.00763423228636384, + 0.00995621271431446, + 0.009238970465958118, + 0.008876271545886993, + 0.01225938368588686, + 0.012803727760910988, + 0.014001498930156231, + 0.01012786477804184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0232852790504694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978767640888691 + ], + "opposite_logprob_tokens": [ + -3.52978777885437 + ], + "logprob_reward": -0.02978767640888691, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978767640888691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.318152517080307, + 0.06134364753961563, + 0.09636715054512024, + 0.10731679201126099, + 0.03701882064342499, + 0.07727035880088806, + 0.15598812699317932, + 0.06462395191192627, + 0.025418583303689957, + 0.05650000646710396 + ], + "credit_map": [ + 0.02616407535970211, + 0.032192714512348175, + 0.0329887717962265, + 0.0698128268122673, + 0.031010981649160385, + 0.026359139010310173, + 0.07659934461116791, + 0.023024728521704674, + 0.007527146488428116, + 0.006019530817866325, + 0.009683972224593163, + 0.014039631932973862, + 0.004278514068573713, + 0.005513490177690983, + 0.007198432460427284, + 0.0070829326286911964, + 0.010000452399253845, + 0.008765097707509995, + 0.012907624244689941, + 0.01388910785317421, + 0.008162273094058037, + 0.010673479177057743, + 0.019743112847208977, + 0.012226015329360962, + 0.00960650946944952, + 0.012169912457466125, + 0.01430723536759615, + 0.023168860003352165, + 0.012255319394171238, + 0.009303835220634937, + 0.013712014071643353, + 0.01279311254620552, + 0.0034677647054195404, + 0.0058983489871025085, + 0.004461569711565971, + 0.004826280288398266, + 0.0046164533123373985, + 0.0054063741117715836, + 0.004600035026669502, + 0.0037419977597892284, + 0.01398675236850977, + 0.007752734702080488, + 0.005282623693346977, + 0.010419461876153946, + 0.010043146088719368, + 0.006839575711637735, + 0.012268219143152237, + 0.010677851736545563, + 0.016961149871349335, + 0.023358497768640518, + 0.008853600360453129, + 0.02904113568365574, + 0.032051555812358856, + 0.014627738855779171, + 0.01980755664408207, + 0.011286906898021698, + 0.007084992248564959, + 0.00869505014270544, + 0.0069670178927481174, + 0.008351637981832027, + 0.01138390228152275, + 0.0066460780799388885, + 0.009726272895932198, + 0.00576900877058506, + 0.004450421314686537, + 0.003869841806590557, + 0.0031786663457751274, + 0.0030943253077566624, + 0.003136704908683896, + 0.0032960905227810144, + 0.0016954350285232067, + 0.0026971001643687487, + 0.007857773452997208, + 0.0076636201702058315, + 0.00632592523470521, + 0.005385014694184065, + 0.008977132849395275, + 0.007012858986854553, + 0.007220106199383736, + 0.0060575841926038265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.02978767640888691 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305369406938553 + ], + "opposite_logprob_tokens": [ + -3.16805362701416 + ], + "logprob_reward": -0.04305369406938553, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305369406938553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20112740993499756, + 0.05592799931764603, + 0.06578546017408371, + 0.07675078511238098, + 0.07087421417236328, + 0.0615437850356102, + 0.29025232791900635, + 0.06256847083568573, + 0.032958563417196274, + 0.08221093565225601 + ], + "credit_map": [ + 0.02426733449101448, + 0.020132169127464294, + 0.02273506298661232, + 0.03784160315990448, + 0.020694980397820473, + 0.016878796741366386, + 0.038265328854322433, + 0.020312149077653885, + 0.006359852850437164, + 0.005979006644338369, + 0.008364175446331501, + 0.010685106739401817, + 0.005352389998733997, + 0.008243036456406116, + 0.006158053409308195, + 0.004786388482898474, + 0.006648705806583166, + 0.006564110517501831, + 0.011814508587121964, + 0.007374319713562727, + 0.006329822354018688, + 0.00864826887845993, + 0.011384543962776661, + 0.007021186873316765, + 0.007733214180916548, + 0.0076817236840724945, + 0.013210796751081944, + 0.012917292304337025, + 0.009796014055609703, + 0.0067709037102758884, + 0.009587924927473068, + 0.009052921086549759, + 0.00756543455645442, + 0.009639314375817776, + 0.00988541729748249, + 0.009178074076771736, + 0.009305432438850403, + 0.007223167922347784, + 0.00962860882282257, + 0.008448775857686996, + 0.008989821188151836, + 0.008024154230952263, + 0.003757180878892541, + 0.008958812803030014, + 0.006735616363584995, + 0.006554174702614546, + 0.01061205007135868, + 0.007911978289484978, + 0.03006688877940178, + 0.04266961291432381, + 0.019533179700374603, + 0.0573100782930851, + 0.05087866261601448, + 0.02795909158885479, + 0.0353313572704792, + 0.026503482833504677, + 0.007235435768961906, + 0.00656826002523303, + 0.008951907977461815, + 0.010765188373625278, + 0.008067893795669079, + 0.008787604980170727, + 0.007336015347391367, + 0.004856165964156389, + 0.006591170560568571, + 0.005004874896258116, + 0.0028415462002158165, + 0.005406532436609268, + 0.003864375641569495, + 0.0036261335480958223, + 0.002440344775095582, + 0.003183589316904545, + 0.009470346383750439, + 0.0071254572831094265, + 0.008042017929255962, + 0.009078827686607838, + 0.010956257581710815, + 0.010595680214464664, + 0.01822785660624504, + 0.008714496158063412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.04305369406938553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009772081859409809 + ], + "opposite_logprob_tokens": [ + -4.634772300720215 + ], + "logprob_reward": -0.009772081859409809, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009772081859409809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1440931111574173, + 0.07327258586883545, + 0.21605245769023895, + 0.052823297679424286, + 0.0481434166431427, + 0.11675172299146652, + 0.10292860120534897, + 0.11942382156848907, + 0.01919555477797985, + 0.10731552541255951 + ], + "credit_map": [ + 0.014972417615354061, + 0.013540973886847496, + 0.012772904708981514, + 0.03147366642951965, + 0.014251724816858768, + 0.013136298395693302, + 0.028818905353546143, + 0.015126215294003487, + 0.007369098253548145, + 0.01027614250779152, + 0.010906122624874115, + 0.011861209757626057, + 0.007410550955682993, + 0.008780255913734436, + 0.008460133336484432, + 0.008209063671529293, + 0.020180171355605125, + 0.017696978524327278, + 0.03365708887577057, + 0.025023018941283226, + 0.021884437650442123, + 0.02492368407547474, + 0.045922424644231796, + 0.026764625683426857, + 0.004382798448204994, + 0.005747857969254255, + 0.008054854348301888, + 0.008527313359081745, + 0.006011095363646746, + 0.006141727324575186, + 0.008785848505795002, + 0.005171798635274172, + 0.004911847412586212, + 0.008305828087031841, + 0.006046094466000795, + 0.004406878724694252, + 0.006474328693002462, + 0.007691600825637579, + 0.004698559641838074, + 0.005608272273093462, + 0.0255221426486969, + 0.0096204224973917, + 0.010001926682889462, + 0.01149565540254116, + 0.009710201993584633, + 0.009195315651595592, + 0.014541404321789742, + 0.026664625853300095, + 0.012096067890524864, + 0.013377680443227291, + 0.007847685366868973, + 0.018581600859761238, + 0.014628485776484013, + 0.01205262541770935, + 0.015523815527558327, + 0.008820627816021442, + 0.016306661069393158, + 0.014324686489999294, + 0.01238871831446886, + 0.01767784357070923, + 0.013435657136142254, + 0.022250693291425705, + 0.013530331663787365, + 0.009509210474789143, + 0.0027735603507608175, + 0.00235479767434299, + 0.0021180908661335707, + 0.0033555254340171814, + 0.001773034455254674, + 0.002546532079577446, + 0.0017251934623345733, + 0.0025488194078207016, + 0.013858314603567123, + 0.009806358255445957, + 0.009885724633932114, + 0.015705596655607224, + 0.01244497299194336, + 0.011360601522028446, + 0.021933268755674362, + 0.012320682406425476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.009772081859409809 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping long shot, the camera moving in smooth tracking motion, follows a hoverboard through the digital canyons of a futuristic city. Flanked by towering skyscrapers, their surfaces covered in massive holographic billboards and dynamic data waterfalls, countless aircraft of varying shapes weave through multi-tiered air traffic arteries. The sleek, dark carbon fiber composite hoverboard itself is a flowing light source, its surface and edges shimmering with neon blue and purple streams of binary code and complex geometric patterns. It silently accelerates, emitting a pale blue energy trail from its tail. Then, deftly, it tilts to the left, tracing a precise, smooth arc, effortlessly passing a bulkier, industrial-style hover truck next to it. The entire scene is meticulously rendered with the textures of high-fidelity CG, with physically accurate light and shadow reflecting off the cold metal and glass surfaces. The predominant color palette is a deep midnight blue, illuminated by the vibrant colors of the holographic advertisements, creating a dynamic atmosphere of high speed, awe, and a futuristic technological aesthetic.", + "seed": 30499, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_4/index_7169_seed_30499_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020562912337481976, + "vlm_logprob": -0.020562912337481976, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.020562912337481976, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006347969174385071 + ], + "opposite_logprob_tokens": [ + -5.381348133087158 + ], + "logprob_reward": -0.006347969174385071, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.006347969174385071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12140926718711853, + 0.08061549812555313, + 0.22689686715602875, + 0.19752371311187744, + 0.11714434623718262, + 0.03458341211080551, + 0.03749322146177292, + 0.06270994246006012, + 0.03831587731838226, + 0.08330786973237991 + ], + "credit_map": [ + 0.009776519611477852, + 0.011928847059607506, + 0.009323694743216038, + 0.011821706779301167, + 0.01364098396152258, + 0.0254437867552042, + 0.019770625978708267, + 0.019703097641468048, + 0.00929358508437872, + 0.005848522763699293, + 0.007760519161820412, + 0.007234249264001846, + 0.0106198163703084, + 0.017238354310393333, + 0.011089157313108444, + 0.011531288735568523, + 0.019651640206575394, + 0.02175939828157425, + 0.02864081785082817, + 0.03236548975110054, + 0.0181140024214983, + 0.027675103396177292, + 0.03841853141784668, + 0.04027187079191208, + 0.021921522915363312, + 0.020499324426054955, + 0.02702050656080246, + 0.0282338447868824, + 0.018459798768162727, + 0.026015078648924828, + 0.03202065825462341, + 0.023352961987257004, + 0.013413713313639164, + 0.019515644758939743, + 0.020716248080134392, + 0.011537655256688595, + 0.016014620661735535, + 0.020207736641168594, + 0.006902184337377548, + 0.008836543187499046, + 0.004308526404201984, + 0.004458303097635508, + 0.005054938141256571, + 0.0018847411265596747, + 0.005804014857858419, + 0.00851111114025116, + 0.0021931652445346117, + 0.0023686124477535486, + 0.005353105254471302, + 0.006260056979954243, + 0.0035571777261793613, + 0.0035440523643046618, + 0.005094996653497219, + 0.007673180662095547, + 0.0031438858713954687, + 0.0028667685110121965, + 0.0066564991138875484, + 0.006616794504225254, + 0.006991222966462374, + 0.004111460410058498, + 0.014842133969068527, + 0.0065468428656458855, + 0.0033851484768092632, + 0.01355984527617693, + 0.005485546309500933, + 0.005881006829440594, + 0.004697102587670088, + 0.004087221808731556, + 0.005850167479366064, + 0.006389375776052475, + 0.0031430660746991634, + 0.002782392082735896, + 0.010093226097524166, + 0.009212377481162548, + 0.012799144722521305, + 0.009557797573506832, + 0.0088850948959589, + 0.015500197187066078, + 0.008529065176844597, + 0.008730960078537464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.006347969174385071 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808938339352608 + ], + "opposite_logprob_tokens": [ + -3.2880892753601074 + ], + "logprob_reward": -0.03808938339352608, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808938339352608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11396841704845428, + 0.15611247718334198, + 0.18480947613716125, + 0.04315909370779991, + 0.07403716444969177, + 0.06359294056892395, + 0.12094562500715256, + 0.07215363532304764, + 0.0541251040995121, + 0.11709608137607574 + ], + "credit_map": [ + 0.008135039359331131, + 0.011393574997782707, + 0.013775632716715336, + 0.010394283570349216, + 0.01337909884750843, + 0.022277984768152237, + 0.016537919640541077, + 0.018074873834848404, + 0.012712113559246063, + 0.0114979837089777, + 0.012899674475193024, + 0.018112557008862495, + 0.022179635241627693, + 0.03370596468448639, + 0.020017677918076515, + 0.024986855685710907, + 0.020131969824433327, + 0.024268146604299545, + 0.025450026616454124, + 0.03227457404136658, + 0.01933942176401615, + 0.02599513530731201, + 0.02250155247747898, + 0.014848628081381321, + 0.00521861482411623, + 0.004116071853786707, + 0.005449904594570398, + 0.006415808107703924, + 0.004484167788177729, + 0.0065002781338989735, + 0.006169854197651148, + 0.004804390016943216, + 0.00693622138351202, + 0.010686671361327171, + 0.011850181967020035, + 0.007460231427103281, + 0.010125914588570595, + 0.009895735420286655, + 0.007974025793373585, + 0.009108168072998524, + 0.005855415482074022, + 0.010506821796298027, + 0.01167716458439827, + 0.004388557281345129, + 0.009565792977809906, + 0.011325999163091183, + 0.004901609383523464, + 0.005371572449803352, + 0.010836141183972359, + 0.016102924942970276, + 0.013582801446318626, + 0.01549915224313736, + 0.01676202565431595, + 0.017400434240698814, + 0.017611360177397728, + 0.013150773011147976, + 0.0073101636953651905, + 0.006058104336261749, + 0.006099571008235216, + 0.0052766455337405205, + 0.016766032204031944, + 0.011294751428067684, + 0.00478073675185442, + 0.014567622914910316, + 0.00360841304063797, + 0.0044899932108819485, + 0.006959329824894667, + 0.006615549325942993, + 0.006828943733125925, + 0.009086251258850098, + 0.010902347043156624, + 0.005634270142763853, + 0.010865838266909122, + 0.011517460457980633, + 0.01852479577064514, + 0.015481404028832912, + 0.012789005413651466, + 0.018443306908011436, + 0.015605032444000244, + 0.01386923249810934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.03808938339352608 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016057223081588745 + ], + "opposite_logprob_tokens": [ + -4.141057014465332 + ], + "logprob_reward": -0.016057223081588745, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016057223081588745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13596951961517334, + 0.09491787850856781, + 0.3083544671535492, + 0.034212760627269745, + 0.03790624067187309, + 0.07983211427927017, + 0.05491826683282852, + 0.10053404420614243, + 0.0792040303349495, + 0.07415065914392471 + ], + "credit_map": [ + 0.010949417017400265, + 0.011193269863724709, + 0.014632456004619598, + 0.013894107192754745, + 0.014220290817320347, + 0.02581358142197132, + 0.02321365848183632, + 0.02205275371670723, + 0.007740089204162359, + 0.008107303641736507, + 0.013363166712224483, + 0.011974851600825787, + 0.017063219100236893, + 0.014020479284226894, + 0.011064921505749226, + 0.011583855375647545, + 0.032872144132852554, + 0.03423909470438957, + 0.03949376940727234, + 0.043966639786958694, + 0.031447820365428925, + 0.037138182669878006, + 0.04880605265498161, + 0.04039078950881958, + 0.004940928425639868, + 0.004003013949841261, + 0.004730336368083954, + 0.0049544754438102245, + 0.002929635113105178, + 0.004247749224305153, + 0.004342854022979736, + 0.0040637715719640255, + 0.003323717974126339, + 0.0041939509101212025, + 0.00540680019184947, + 0.004995414521545172, + 0.005933440290391445, + 0.004258322063833475, + 0.004248378332704306, + 0.005546220112591982, + 0.010176314041018486, + 0.010979513637721539, + 0.012082545086741447, + 0.006452643778175116, + 0.01165690179914236, + 0.01626153476536274, + 0.00518132233992219, + 0.0070413448847830296, + 0.0055517759174108505, + 0.01052782405167818, + 0.004961244761943817, + 0.005484007298946381, + 0.007392529863864183, + 0.010499134659767151, + 0.005694796331226826, + 0.0048069641925394535, + 0.01215786300599575, + 0.010651933960616589, + 0.009573654271662235, + 0.005492828786373138, + 0.024960588663816452, + 0.013040851801633835, + 0.004810203332453966, + 0.01984613575041294, + 0.00922033004462719, + 0.012794900685548782, + 0.011026386171579361, + 0.007005774881690741, + 0.012287138029932976, + 0.015005978755652905, + 0.007283943705260754, + 0.00457958597689867, + 0.009943355806171894, + 0.009189016185700893, + 0.010508575476706028, + 0.009518636390566826, + 0.008363929577171803, + 0.009372479282319546, + 0.008987760171294212, + 0.008266913704574108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.016057223081588745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369199112057686 + ], + "opposite_logprob_tokens": [ + -3.408691883087158 + ], + "logprob_reward": -0.03369199112057686, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369199112057686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07173197716474533, + 0.1012563556432724, + 0.28786617517471313, + 0.06892308592796326, + 0.0500970222055912, + 0.11903072893619537, + 0.09103842824697495, + 0.10450389981269836, + 0.06480171531438828, + 0.04075060039758682 + ], + "credit_map": [ + 0.00659015029668808, + 0.0066656144335865974, + 0.01007213443517685, + 0.0071259718388319016, + 0.00854659453034401, + 0.009994289837777615, + 0.011271974071860313, + 0.011465253308415413, + 0.009426780045032501, + 0.011372080072760582, + 0.01106040459126234, + 0.015532818622887135, + 0.015572204254567623, + 0.014305808581411839, + 0.014238080941140652, + 0.0097481869161129, + 0.027157654985785484, + 0.027447117492556572, + 0.03553326055407524, + 0.04767778515815735, + 0.02288993075489998, + 0.033340562134981155, + 0.0490211620926857, + 0.044798724353313446, + 0.008490180596709251, + 0.007614922244101763, + 0.008217738009989262, + 0.009807869791984558, + 0.00708469282835722, + 0.008801475167274475, + 0.01084877923130989, + 0.008057435974478722, + 0.005112229380756617, + 0.00763497082516551, + 0.007671034429222345, + 0.005426964722573757, + 0.005769188050180674, + 0.005199857987463474, + 0.006689473055303097, + 0.006593306548893452, + 0.014390027150511742, + 0.01965503953397274, + 0.020130440592765808, + 0.005835529416799545, + 0.017619390040636063, + 0.025521086528897285, + 0.00692771328613162, + 0.008951506577432156, + 0.01690412126481533, + 0.020520752295851707, + 0.011054006405174732, + 0.0056897737085819244, + 0.012143990024924278, + 0.013959440402686596, + 0.004875369369983673, + 0.005890980828553438, + 0.016254374757409096, + 0.012279094196856022, + 0.018531955778598785, + 0.009529692120850086, + 0.01650756411254406, + 0.009750714525580406, + 0.008389195427298546, + 0.013261313550174236, + 0.011664316989481449, + 0.00875862967222929, + 0.009546884335577488, + 0.005576620809733868, + 0.007608854211866856, + 0.01271939929574728, + 0.005306541919708252, + 0.0036204762291163206, + 0.0060797398909926414, + 0.004992474336177111, + 0.006162106990814209, + 0.004820897709578276, + 0.003989495802670717, + 0.006136350333690643, + 0.004910495597869158, + 0.0036590376403182745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.03369199112057686 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a futuristic city with towering skyscrapers featuring holographic billboards and data waterfalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a sleek, dark carbon fiber composite hoverboard emitting a pale blue energy trail from its tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hoverboard accelerate, tilt to the left, and trace a smooth arc while passing a bulkier industrial-style hover truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hoverboard's surface shimmering with neon blue and purple streams of binary code and complex geometric patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a smooth tracking motion following the hoverboard in a sweeping long shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008627994917333126 + ], + "opposite_logprob_tokens": [ + -4.758627891540527 + ], + "logprob_reward": -0.008627994917333126, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008627994917333126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12137839943170547, + 0.0981767550110817, + 0.3514504134654999, + 0.03951791301369667, + 0.047621116042137146, + 0.05183957889676094, + 0.045460496097803116, + 0.06527416408061981, + 0.04305754974484444, + 0.13622356951236725 + ], + "credit_map": [ + 0.009517303667962551, + 0.007330873981118202, + 0.012910121120512486, + 0.011791874654591084, + 0.014095119200646877, + 0.01604411005973816, + 0.028887053951621056, + 0.020801948383450508, + 0.00817402545362711, + 0.008222595788538456, + 0.00737333158031106, + 0.007653241511434317, + 0.013342607766389847, + 0.021641835570335388, + 0.011852476745843887, + 0.019916651770472527, + 0.034505464136600494, + 0.03294900432229042, + 0.04248540475964546, + 0.05915923789143562, + 0.029666351154446602, + 0.04154190793633461, + 0.0568644218146801, + 0.0542786605656147, + 0.005815558135509491, + 0.0034693796187639236, + 0.008243177086114883, + 0.006589769851416349, + 0.0039438121020793915, + 0.002929945942014456, + 0.004502086900174618, + 0.004024187568575144, + 0.005375250708311796, + 0.007519290782511234, + 0.007842648774385452, + 0.005278532858937979, + 0.006109245587140322, + 0.005872614681720734, + 0.0038280722219496965, + 0.005795465316623449, + 0.0059844329953193665, + 0.005558063741773367, + 0.007735128980129957, + 0.004353611264377832, + 0.00856638140976429, + 0.01153567060828209, + 0.004200993571430445, + 0.003905300749465823, + 0.004537161439657211, + 0.008425040170550346, + 0.005142385605722666, + 0.003908355254679918, + 0.0054793404415249825, + 0.0062159341759979725, + 0.007574534043669701, + 0.004177750553935766, + 0.007443114183843136, + 0.008308825083076954, + 0.005769580602645874, + 0.00453610485419631, + 0.01385316252708435, + 0.009652258828282356, + 0.0046849907375872135, + 0.011026138439774513, + 0.007680321112275124, + 0.006672514136880636, + 0.005241480190306902, + 0.003048292361199856, + 0.0054682716727256775, + 0.008821453899145126, + 0.003413188038393855, + 0.0027120334561914206, + 0.025891050696372986, + 0.016691794618964195, + 0.022088440135121346, + 0.010976758785545826, + 0.016919411718845367, + 0.017441727221012115, + 0.017406003549695015, + 0.008808386512100697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.008627994917333126 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_5/index_tensor([3219, 3219])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_5/index_tensor([3219, 3219])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0436b6a04dd8c45d01dd075a29e7f346c650fda1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_5/index_tensor([3219, 3219])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera follows a young man from a low angle, parallel to him, panning smoothly to the right to capture him as he briskly navigates the streets on his bicycle. The young man, with his short, slightly tousled dark brown hair fluttering gently in the breeze, wears a relaxed, cheerful smile. He wears a simple white cotton T-shirt over an unbuttoned olive green work shirt, slim-fit washed blue jeans, and classic white canvas shoes. He rides a retro-style city bike with a bright yellow frame, a brown leather seat and grips, and a beige canvas messenger bag hanging from the handlebars. He leans slightly forward, his hands resting effortlessly on the handlebars, his legs pedaling with a steady, rhythmic rhythm, propelling the bike smoothly forward over the uneven cobblestone road. They were walking through a narrow, winding alleyway in the ancient town, flanked by traditional Chinese houses with mottled whitewashed walls and green-tiled roofs. Verdant vines clung to the walls, and red lanterns occasionally hung from the eaves. The warm afternoon sun filtered through the gaps between the buildings above, casting dappled, playful shadows on the ground and on him. The dominant colors of the scene were the gray of the slate and the white of the walls, accented by the bright yellow of the bicycle, the green of the vines, and the colors of his clothing, creating a tranquil, leisurely, and nostalgic atmosphere full of the flavor of life.", + "seed": 17218, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_5/index_3219_seed_17218_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0696592541411519, + "vlm_logprob": -0.0696592541411519, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.0696592541411519, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03744402527809143 + ], + "opposite_logprob_tokens": [ + -4.287444114685059 + ], + "logprob_reward": -0.03744402527809143, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.03744402527809143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6338849663734436, + 0.036055244505405426, + 0.03595820814371109, + 0.05759013071656227, + 0.06202252209186554, + 0.03945567458868027, + 0.036652155220508575, + 0.033645160496234894, + 0.022867165505886078, + 0.041868723928928375 + ], + "credit_map": [ + 0.053823065012693405, + 0.027529720216989517, + 0.20553037524223328, + 0.031884144991636276, + 0.02135610394179821, + 0.07099808752536774, + 0.04287726804614067, + 0.17988628149032593, + 0.003646568162366748, + 0.006252991501241922, + 0.003293185029178858, + 0.004737687297165394, + 0.004333284217864275, + 0.0031966338865458965, + 0.007706734351813793, + 0.002888163784518838, + 0.003062507836148143, + 0.005664001684635878, + 0.006142575293779373, + 0.005095252301543951, + 0.004178905393928289, + 0.004247363191097975, + 0.004066980443894863, + 0.0035006250254809856, + 0.004808751400560141, + 0.008336808532476425, + 0.006341641303151846, + 0.008034657686948776, + 0.010473264381289482, + 0.005411844700574875, + 0.00993064884096384, + 0.004252519924193621, + 0.004306409507989883, + 0.005277311895042658, + 0.015701133757829666, + 0.003647798905149102, + 0.006915129255503416, + 0.018036166206002235, + 0.004249921068549156, + 0.003888659877702594, + 0.004727332387119532, + 0.004086473491042852, + 0.005484194960445166, + 0.008360435254871845, + 0.004710166249424219, + 0.0038857138715684414, + 0.004646891728043556, + 0.0035544715356081724, + 0.006368331611156464, + 0.00627495301887393, + 0.0027840929105877876, + 0.005182276945561171, + 0.0035112188197672367, + 0.005134051665663719, + 0.002748670522123575, + 0.004648564383387566, + 0.0034754008520394564, + 0.003192853182554245, + 0.0020421259105205536, + 0.003587419632822275, + 0.007741875480860472, + 0.004772411659359932, + 0.003928157966583967, + 0.004904921166598797, + 0.001891297404654324, + 0.0024605803191661835, + 0.0022259552497416735, + 0.0048791212029755116, + 0.003128949785605073, + 0.002886816393584013, + 0.003050147322937846, + 0.0023442984092980623, + 0.006064770743250847, + 0.008642674423754215, + 0.0069175586104393005, + 0.002938107820227742, + 0.00572860985994339, + 0.004714416339993477, + 0.003190091345459223, + 0.0036724957171827555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.03744402527809143 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054962534457445145 + ], + "opposite_logprob_tokens": [ + -2.929962635040283 + ], + "logprob_reward": -0.054962534457445145, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054962534457445145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5544931292533875, + 0.033482179045677185, + 0.09091920405626297, + 0.059010379016399384, + 0.0693860873579979, + 0.04062303528189659, + 0.0299989040941, + 0.023267192766070366, + 0.05872117355465889, + 0.04009876027703285 + ], + "credit_map": [ + 0.04651618003845215, + 0.023621702566742897, + 0.18177488446235657, + 0.029740510508418083, + 0.01998445950448513, + 0.06478079408407211, + 0.03801169991493225, + 0.1500629037618637, + 0.0034855795092880726, + 0.007258898578584194, + 0.004257113207131624, + 0.005000603850930929, + 0.002418017480522394, + 0.00358016905374825, + 0.0036357587669044733, + 0.0038460376672446728, + 0.008004622533917427, + 0.014656070619821548, + 0.010725406929850578, + 0.019540343433618546, + 0.011811681091785431, + 0.010186458006501198, + 0.009833148680627346, + 0.0061614722944796085, + 0.005302791483700275, + 0.00679956516250968, + 0.0066405790857970715, + 0.008547544479370117, + 0.00797696877270937, + 0.005757632199674845, + 0.011439455673098564, + 0.006545841693878174, + 0.005312931723892689, + 0.0077212657779455185, + 0.013350402005016804, + 0.01059695240110159, + 0.006452248431742191, + 0.011082638055086136, + 0.009610469453036785, + 0.005259180907160044, + 0.003504229476675391, + 0.003294197143986821, + 0.005839769262820482, + 0.0065499115735292435, + 0.005339812021702528, + 0.006334759294986725, + 0.004965642001479864, + 0.0047947135753929615, + 0.003062174655497074, + 0.003613237524405122, + 0.004554034676402807, + 0.006842546630650759, + 0.002591485157608986, + 0.0038540286477655172, + 0.0020596515387296677, + 0.0034217475913465023, + 0.003400113433599472, + 0.002254555467516184, + 0.003954503685235977, + 0.0027249460108578205, + 0.0030801836401224136, + 0.0023314093705266714, + 0.0026055341586470604, + 0.0029159486293792725, + 0.005638955160975456, + 0.006125913932919502, + 0.008657353930175304, + 0.012432100251317024, + 0.007444211281836033, + 0.0065549686551094055, + 0.006240656599402428, + 0.005627016071230173, + 0.004598389845341444, + 0.005354644730687141, + 0.008542945608496666, + 0.004451099317520857, + 0.00390935642644763, + 0.005709989462047815, + 0.0036135513801127672, + 0.003918784204870462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.054962534457445145 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781171567738056 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.009781171567738056, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781171567738056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41073086857795715, + 0.030784402042627335, + 0.06546523422002792, + 0.21826036274433136, + 0.07803761214017868, + 0.047393135726451874, + 0.034374698996543884, + 0.033984195441007614, + 0.04458538070321083, + 0.03638409823179245 + ], + "credit_map": [ + 0.03526343032717705, + 0.01860824227333069, + 0.13231642544269562, + 0.022460386157035828, + 0.01453815121203661, + 0.04518818482756615, + 0.02829364687204361, + 0.11406239122152328, + 0.0038056648336350918, + 0.0030311380978673697, + 0.003597662318497896, + 0.004646644927561283, + 0.0034395884722471237, + 0.0031397053971886635, + 0.006214280147105455, + 0.002909719245508313, + 0.003947637043893337, + 0.006302507594227791, + 0.008406688459217548, + 0.013199464417994022, + 0.008655603043735027, + 0.0077918898314237595, + 0.009331139735877514, + 0.007830305024981499, + 0.03716391697525978, + 0.021941667422652245, + 0.029075834900140762, + 0.022291703149676323, + 0.014978374354541302, + 0.01854550838470459, + 0.04930147901177406, + 0.024961883202195168, + 0.003922439180314541, + 0.005560371559113264, + 0.016517946496605873, + 0.009365396574139595, + 0.009829113259911537, + 0.016730403527617455, + 0.009601357392966747, + 0.006510588806122541, + 0.003786621615290642, + 0.004462862852960825, + 0.004984621424227953, + 0.00765796285122633, + 0.006483144126832485, + 0.007904238067567348, + 0.0072109270840883255, + 0.004902760032564402, + 0.0030831971671432257, + 0.004295746795833111, + 0.00398337934166193, + 0.006523061078041792, + 0.0025458589661866426, + 0.0047307671047747135, + 0.0033452766947448254, + 0.005867414176464081, + 0.004092135466635227, + 0.0031370511278510094, + 0.003613680135458708, + 0.004622845910489559, + 0.004309493582695723, + 0.0039909309707582, + 0.003504400607198477, + 0.006713656708598137, + 0.004549398552626371, + 0.004420670215040445, + 0.0067579722963273525, + 0.00937721598893404, + 0.005487598013132811, + 0.004475958179682493, + 0.004887340124696493, + 0.004629222676157951, + 0.005395232234150171, + 0.007862789556384087, + 0.004857406485825777, + 0.0037923003546893597, + 0.0038643707521259785, + 0.004716414958238602, + 0.0026944384444504976, + 0.003201147308573127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.009781171567738056 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057354897260666 + ], + "opposite_logprob_tokens": [ + -3.895573616027832 + ], + "logprob_reward": -0.02057354897260666, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057354897260666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7991141676902771, + 0.03180829435586929, + 0.023729292675852776, + 0.014631514437496662, + 0.025308851152658463, + 0.012862890027463436, + 0.016034526750445366, + 0.024431094527244568, + 0.01326857227832079, + 0.038810715079307556 + ], + "credit_map": [ + 0.06780461966991425, + 0.03349408134818077, + 0.26469171047210693, + 0.042402494698762894, + 0.02555989660322666, + 0.09558438509702682, + 0.05232134461402893, + 0.21725577116012573, + 0.003142446978017688, + 0.0040943981148302555, + 0.003142887493595481, + 0.004030134528875351, + 0.004517989698797464, + 0.003033592365682125, + 0.005923005286604166, + 0.0039238473400473595, + 0.0023824195377528667, + 0.004612621385604143, + 0.003646714612841606, + 0.0023830747231841087, + 0.0029348537791520357, + 0.0030009467154741287, + 0.0025475118309259415, + 0.0022211528848856688, + 0.002405479084700346, + 0.002305146772414446, + 0.002006791066378355, + 0.0016217613592743874, + 0.0017332007410004735, + 0.001351002836599946, + 0.0014746901579201221, + 0.0017334448639303446, + 0.003508699359372258, + 0.004594879224896431, + 0.003308625193312764, + 0.002511693863198161, + 0.0022981760557740927, + 0.003920330200344324, + 0.0020538214594125748, + 0.003112632082775235, + 0.0016556611517444253, + 0.0016017499146983027, + 0.001776783843524754, + 0.001953801838681102, + 0.0018298033392056823, + 0.001237240619957447, + 0.00154910609126091, + 0.001258745207451284, + 0.001512340852059424, + 0.002497553825378418, + 0.001811190857551992, + 0.002601283136755228, + 0.0023054094053804874, + 0.0020245504565536976, + 0.0012650784337893128, + 0.002017123159021139, + 0.005372822284698486, + 0.002401074394583702, + 0.0032231116201728582, + 0.003022644901648164, + 0.0034731649793684483, + 0.003397194901481271, + 0.001857787137851119, + 0.0016832962865009904, + 0.0012609312543645501, + 0.0022132601588964462, + 0.0015160789480432868, + 0.001719602383673191, + 0.00341561296954751, + 0.001059147180058062, + 0.0010827064979821444, + 0.001001235214062035, + 0.004843942355364561, + 0.005487622693181038, + 0.008329399861395359, + 0.005049341823905706, + 0.004813177511096001, + 0.0027618734166026115, + 0.004442776553332806, + 0.0030825866851955652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.02057354897260666 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22553499042987823 + ], + "opposite_logprob_tokens": [ + -1.6005350351333618 + ], + "logprob_reward": -0.22553499042987823, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22553499042987823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5839791297912598, + 0.09380494058132172, + 0.08220987766981125, + 0.02779342420399189, + 0.07195071876049042, + 0.01866540126502514, + 0.014665167778730392, + 0.016606954857707024, + 0.028997249901294708, + 0.06132715567946434 + ], + "credit_map": [ + 0.04971284046769142, + 0.024461599066853523, + 0.19283929467201233, + 0.03005223348736763, + 0.01940970867872238, + 0.07025366276502609, + 0.03941437602043152, + 0.1578354388475418, + 0.01094078179448843, + 0.009401082992553711, + 0.010027259588241577, + 0.008910995908081532, + 0.014323823153972626, + 0.009513928554952145, + 0.018548494204878807, + 0.012138569727540016, + 0.008230714127421379, + 0.013959730975329876, + 0.011753744445741177, + 0.014859282411634922, + 0.009458954446017742, + 0.009622769430279732, + 0.008243872784078121, + 0.0060808053240180016, + 0.004465007223188877, + 0.0025946609675884247, + 0.003519620280712843, + 0.0032969373278319836, + 0.0028353568632155657, + 0.002419309690594673, + 0.00546020595356822, + 0.0032023251987993717, + 0.007084968034178019, + 0.006949230562895536, + 0.018515408039093018, + 0.004516506101936102, + 0.004713002592325211, + 0.021604139357805252, + 0.004251491278409958, + 0.004315970931202173, + 0.0015472444938495755, + 0.002045569708570838, + 0.002372047631070018, + 0.0029527300503104925, + 0.0033143297769129276, + 0.002419795375317335, + 0.0024383936543017626, + 0.0015752894105389714, + 0.0016227388987317681, + 0.0023351332638412714, + 0.0017713940469548106, + 0.0019184850389137864, + 0.002130133332684636, + 0.001624134136363864, + 0.001197870122268796, + 0.0020652783568948507, + 0.002184659242630005, + 0.0019909567199647427, + 0.002364659681916237, + 0.0019168101716786623, + 0.0025334074161946774, + 0.0025756917893886566, + 0.001511898241005838, + 0.001528872293420136, + 0.002520054578781128, + 0.0051548476330935955, + 0.0027499564457684755, + 0.004669437650591135, + 0.008974334225058556, + 0.0020619055721908808, + 0.001597104943357408, + 0.00126961141359061, + 0.008517374284565449, + 0.01311295572668314, + 0.009483146481215954, + 0.007478684652596712, + 0.006314877420663834, + 0.006744615267962217, + 0.005876325070858002, + 0.003799178870394826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.22553499042987823 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera follows a young man from a low angle, parallel to him, panning smoothly to the right to capture him as he briskly navigates the streets on his bicycle. The young man, with his short, slightly tousled dark brown hair fluttering gently in the breeze, wears a relaxed, cheerful smile. He wears a simple white cotton T-shirt over an unbuttoned olive green work shirt, slim-fit washed blue jeans, and classic white canvas shoes. He rides a retro-style city bike with a bright yellow frame, a brown leather seat and grips, and a beige canvas messenger bag hanging from the handlebars. He leans slightly forward, his hands resting effortlessly on the handlebars, his legs pedaling with a steady, rhythmic rhythm, propelling the bike smoothly forward over the uneven cobblestone road. They were walking through a narrow, winding alleyway in the ancient town, flanked by traditional Chinese houses with mottled whitewashed walls and green-tiled roofs. Verdant vines clung to the walls, and red lanterns occasionally hung from the eaves. The warm afternoon sun filtered through the gaps between the buildings above, casting dappled, playful shadows on the ground and on him. The dominant colors of the scene were the gray of the slate and the white of the walls, accented by the bright yellow of the bicycle, the green of the vines, and the colors of his clothing, creating a tranquil, leisurely, and nostalgic atmosphere full of the flavor of life.", + "seed": 17219, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_5/index_3219_seed_17219_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05477635767310858, + "vlm_logprob": -0.05477635767310858, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.05477635767310858, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0423666313290596 + ], + "opposite_logprob_tokens": [ + -3.6673667430877686 + ], + "logprob_reward": -0.0423666313290596, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0423666313290596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3758560121059418, + 0.0776209682226181, + 0.20636595785617828, + 0.04352492839097977, + 0.03653835505247116, + 0.02663496695458889, + 0.05731885880231857, + 0.09673995524644852, + 0.055252932012081146, + 0.02414710633456707 + ], + "credit_map": [ + 0.024463120847940445, + 0.027849432080984116, + 0.03267386928200722, + 0.030311236158013344, + 0.11809297651052475, + 0.04026633873581886, + 0.024427123367786407, + 0.07777192443609238, + 0.009003964252769947, + 0.005534389056265354, + 0.01031467691063881, + 0.009386508725583553, + 0.008017608895897865, + 0.013961132615804672, + 0.011434456333518028, + 0.009968229569494724, + 0.01103934459388256, + 0.007704201620072126, + 0.04546380043029785, + 0.009937473572790623, + 0.010567559860646725, + 0.08761433511972427, + 0.02205505594611168, + 0.011984183453023434, + 0.004592636600136757, + 0.004389544483274221, + 0.003552696667611599, + 0.004870461765676737, + 0.005363318137824535, + 0.004569920711219311, + 0.007854785770177841, + 0.008331559598445892, + 0.003530546324327588, + 0.002638624981045723, + 0.005653936415910721, + 0.004355387296527624, + 0.005869888700544834, + 0.005428365431725979, + 0.0035050739534199238, + 0.005556534044444561, + 0.0024248482659459114, + 0.0016120978398248553, + 0.003114280290901661, + 0.0029910090379416943, + 0.0041452450677752495, + 0.003632217412814498, + 0.005134141072630882, + 0.003581127617508173, + 0.0048418911173939705, + 0.007456548977643251, + 0.006195436231791973, + 0.007241037208586931, + 0.006175660528242588, + 0.00702289491891861, + 0.009875018149614334, + 0.008510369807481766, + 0.010569727048277855, + 0.01031075231730938, + 0.014644099399447441, + 0.012759697623550892, + 0.011617700569331646, + 0.010908432304859161, + 0.013939481228590012, + 0.011990062892436981, + 0.005397114437073469, + 0.006657759193331003, + 0.0070015438832342625, + 0.00787914264947176, + 0.008978638797998428, + 0.008070330135524273, + 0.004902961663901806, + 0.0063654412515461445, + 0.002908439375460148, + 0.004009932745248079, + 0.00264293747022748, + 0.002389302011579275, + 0.0032166638411581516, + 0.0029388947878032923, + 0.0033503773156553507, + 0.0026905564591288567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0423666313290596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033714696764945984 + ], + "opposite_logprob_tokens": [ + -3.408714771270752 + ], + "logprob_reward": -0.033714696764945984, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033714696764945984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24757546186447144, + 0.08149784803390503, + 0.11776388436555862, + 0.15012525022029877, + 0.07140068709850311, + 0.06768771260976791, + 0.054220713675022125, + 0.05251815915107727, + 0.11919517070055008, + 0.03801503777503967 + ], + "credit_map": [ + 0.012564809992909431, + 0.013036479242146015, + 0.015129378996789455, + 0.013976308517158031, + 0.08336224406957626, + 0.022454405203461647, + 0.01387760229408741, + 0.07317423820495605, + 0.009060811251401901, + 0.01009584404528141, + 0.007658911403268576, + 0.008152542635798454, + 0.007789044175297022, + 0.012197824195027351, + 0.011206467635929585, + 0.015336411073803902, + 0.0061732386238873005, + 0.004123971331864595, + 0.027695313096046448, + 0.006971126422286034, + 0.007205563597381115, + 0.046316687017679214, + 0.012366031296551228, + 0.0069119613617658615, + 0.016954990103840828, + 0.01870599575340748, + 0.018525514751672745, + 0.029107412323355675, + 0.015546109527349472, + 0.015354545786976814, + 0.015970705077052116, + 0.01995999366044998, + 0.00660077016800642, + 0.00837201438844204, + 0.00881139375269413, + 0.013479454442858696, + 0.009520788677036762, + 0.00887337327003479, + 0.007569810841232538, + 0.008173075504601002, + 0.005234740674495697, + 0.008495046757161617, + 0.007326886989176273, + 0.011323739774525166, + 0.008618350140750408, + 0.00929159764200449, + 0.011282330378890038, + 0.006115019787102938, + 0.005147741641849279, + 0.009093032218515873, + 0.004961131606251001, + 0.0061834873631596565, + 0.006930483039468527, + 0.00804393831640482, + 0.006935696583241224, + 0.006925204303115606, + 0.005761097650974989, + 0.0059339627623558044, + 0.009259374812245369, + 0.006651636678725481, + 0.006757459137588739, + 0.005982238799333572, + 0.006495969370007515, + 0.005676422733813524, + 0.013322023674845695, + 0.023570764809846878, + 0.013611345551908016, + 0.01115355920046568, + 0.020249055698513985, + 0.013436696492135525, + 0.009922615252435207, + 0.013929111883044243, + 0.005149041768163443, + 0.007206350099295378, + 0.004562806338071823, + 0.0036777034401893616, + 0.004539784509688616, + 0.00393823953345418, + 0.004102921113371849, + 0.004838190041482449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.033714696764945984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283764719963074 + ], + "opposite_logprob_tokens": [ + -3.7732837200164795 + ], + "logprob_reward": -0.023283764719963074, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283764719963074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.151264950633049, + 0.07920616865158081, + 0.19340088963508606, + 0.2975879907608032, + 0.08366785198450089, + 0.053738489747047424, + 0.04794609919190407, + 0.040846724063158035, + 0.03422437235713005, + 0.018116420134902 + ], + "credit_map": [ + 0.013597476296126842, + 0.014869551174342632, + 0.015252872370183468, + 0.03229545056819916, + 0.018545465543866158, + 0.023007063195109367, + 0.013233261182904243, + 0.02046382799744606, + 0.007936207577586174, + 0.010207172483205795, + 0.008638939820230007, + 0.0072625321336090565, + 0.008883318863809109, + 0.011511048302054405, + 0.01046829391270876, + 0.014298664405941963, + 0.014063124544918537, + 0.008767317049205303, + 0.04349227994680405, + 0.01414911076426506, + 0.01395360566675663, + 0.06714356690645218, + 0.019466795027256012, + 0.012365113943815231, + 0.03641631454229355, + 0.03165256232023239, + 0.04107305780053139, + 0.05344317480921745, + 0.03740758076310158, + 0.03178093954920769, + 0.03483232855796814, + 0.030982043594121933, + 0.008096727542579174, + 0.010455403476953506, + 0.008296928368508816, + 0.015918364748358727, + 0.010259468108415604, + 0.010235990397632122, + 0.011184342205524445, + 0.009220637381076813, + 0.003858864540234208, + 0.004770446103066206, + 0.004454421810805798, + 0.008233669213950634, + 0.006805688142776489, + 0.007335505448281765, + 0.013979186303913593, + 0.004300713539123535, + 0.005350676830857992, + 0.007216088939458132, + 0.005118178203701973, + 0.003959571942687035, + 0.004039255436509848, + 0.005807612091302872, + 0.010589562356472015, + 0.005865154322236776, + 0.0051445504650473595, + 0.004451545421034098, + 0.0062195616774261, + 0.00445467047393322, + 0.005376629997044802, + 0.004832340870052576, + 0.006336122285574675, + 0.004031307063996792, + 0.0034253995399922132, + 0.004607244394719601, + 0.004239846020936966, + 0.004715846385806799, + 0.004390169866383076, + 0.005359178874641657, + 0.0032833730801939964, + 0.004203315358608961, + 0.002480193739756942, + 0.003672546474263072, + 0.0017674424452707171, + 0.0018852411303669214, + 0.0020706295035779476, + 0.0027269890997558832, + 0.0016098774503916502, + 0.0019035015720874071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.023283764719963074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418930571526289 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.01418930571526289, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418930571526289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35620084404945374, + 0.13220442831516266, + 0.2284470796585083, + 0.03293287009000778, + 0.09567088633775711, + 0.039495326578617096, + 0.05087858438491821, + 0.025816716253757477, + 0.020722869783639908, + 0.017630409449338913 + ], + "credit_map": [ + 0.020794741809368134, + 0.02336641028523445, + 0.0300286915153265, + 0.03187631070613861, + 0.08463823050260544, + 0.1005806028842926, + 0.029030611738562584, + 0.03588520362973213, + 0.00971995946019888, + 0.009581495076417923, + 0.024613559246063232, + 0.019556116312742233, + 0.008512749336659908, + 0.03310105949640274, + 0.0153266079723835, + 0.01179287675768137, + 0.010127821937203407, + 0.006623660214245319, + 0.04870617017149925, + 0.008453235030174255, + 0.015330648981034756, + 0.09718989580869675, + 0.02824368327856064, + 0.013771945610642433, + 0.003079333808273077, + 0.004597690887749195, + 0.0032541395630687475, + 0.005702151916921139, + 0.005111285950988531, + 0.0030447603203356266, + 0.004375769756734371, + 0.003767730901017785, + 0.009256022982299328, + 0.009681579656898975, + 0.008402466773986816, + 0.01165765430778265, + 0.02040526270866394, + 0.013737400993704796, + 0.008778193965554237, + 0.013752294704318047, + 0.003948685713112354, + 0.0029499726369976997, + 0.004356579855084419, + 0.004605397116392851, + 0.006271626800298691, + 0.0055466992780566216, + 0.004859830252826214, + 0.006956529337912798, + 0.005645188502967358, + 0.005818900652229786, + 0.0047511206939816475, + 0.00640325341373682, + 0.007318864576518536, + 0.006079375743865967, + 0.007793890777975321, + 0.00706798629835248, + 0.004777964204549789, + 0.0021727944258600473, + 0.004990784451365471, + 0.0031933323480188847, + 0.002374539617449045, + 0.0029576742090284824, + 0.0025719311088323593, + 0.002777694258838892, + 0.0022337750997394323, + 0.0029264476615935564, + 0.002597537124529481, + 0.0022044258657842875, + 0.0037993991281837225, + 0.002536872634664178, + 0.0022367401979863644, + 0.0021876704413443804, + 0.0023642031010240316, + 0.0038787717930972576, + 0.0012718315701931715, + 0.0013109608553349972, + 0.0020449531730264425, + 0.002243664348497987, + 0.0019601969979703426, + 0.0025558252818882465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.01418930571526289 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032738983631134 + ], + "opposite_logprob_tokens": [ + -1.910327434539795 + ], + "logprob_reward": -0.16032738983631134, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032738983631134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26205962896347046, + 0.2523408532142639, + 0.24830743670463562, + 0.03917215019464493, + 0.04401196539402008, + 0.05189060792326927, + 0.02048591524362564, + 0.030986212193965912, + 0.03337801992893219, + 0.017367228865623474 + ], + "credit_map": [ + 0.012623845599591732, + 0.015785787254571915, + 0.015380213037133217, + 0.035129714757204056, + 0.08214307576417923, + 0.03240090608596802, + 0.016750195994973183, + 0.05184587091207504, + 0.02523854561150074, + 0.028109071776270866, + 0.025097085162997246, + 0.054201267659664154, + 0.0193649809807539, + 0.037611402571201324, + 0.0328892357647419, + 0.029829254373908043, + 0.023017410188913345, + 0.03000102937221527, + 0.03361024335026741, + 0.03066415525972843, + 0.029184948652982712, + 0.05603129789233208, + 0.028551151975989342, + 0.017247166484594345, + 0.0032900681253522635, + 0.004987888503819704, + 0.004734336398541927, + 0.005793599411845207, + 0.005418385379016399, + 0.003731603967025876, + 0.006549995392560959, + 0.004666266962885857, + 0.00543275335803628, + 0.0025749702472239733, + 0.005434192251414061, + 0.0041542500257492065, + 0.009224156849086285, + 0.0064565339125692844, + 0.0057158032432198524, + 0.005019303411245346, + 0.0034394976682960987, + 0.006227795034646988, + 0.00441498588770628, + 0.007816586643457413, + 0.007705307565629482, + 0.006921074353158474, + 0.009423235431313515, + 0.005942119285464287, + 0.0014342668000608683, + 0.002065191511064768, + 0.0019411962712183595, + 0.002953864634037018, + 0.0028599731158465147, + 0.0024456477258354425, + 0.004362335428595543, + 0.0024234377779066563, + 0.00419216463342309, + 0.0029662211891263723, + 0.005095093045383692, + 0.004814215935766697, + 0.002866356400772929, + 0.0035124223213642836, + 0.003914139233529568, + 0.0036255924496799707, + 0.0029503621626645327, + 0.0037914300337433815, + 0.00390090374276042, + 0.006722953170537949, + 0.004541925620287657, + 0.00464699836447835, + 0.003257650649175048, + 0.0035657898988574743, + 0.002171361120417714, + 0.0029205994214862585, + 0.0016871581319719553, + 0.0018574533751234412, + 0.0020767387468367815, + 0.0024719296488910913, + 0.0018145213834941387, + 0.002367465989664197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.16032738983631134 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera follows a young man from a low angle, parallel to him, panning smoothly to the right to capture him as he briskly navigates the streets on his bicycle. The young man, with his short, slightly tousled dark brown hair fluttering gently in the breeze, wears a relaxed, cheerful smile. He wears a simple white cotton T-shirt over an unbuttoned olive green work shirt, slim-fit washed blue jeans, and classic white canvas shoes. He rides a retro-style city bike with a bright yellow frame, a brown leather seat and grips, and a beige canvas messenger bag hanging from the handlebars. He leans slightly forward, his hands resting effortlessly on the handlebars, his legs pedaling with a steady, rhythmic rhythm, propelling the bike smoothly forward over the uneven cobblestone road. They were walking through a narrow, winding alleyway in the ancient town, flanked by traditional Chinese houses with mottled whitewashed walls and green-tiled roofs. Verdant vines clung to the walls, and red lanterns occasionally hung from the eaves. The warm afternoon sun filtered through the gaps between the buildings above, casting dappled, playful shadows on the ground and on him. The dominant colors of the scene were the gray of the slate and the white of the walls, accented by the bright yellow of the bicycle, the green of the vines, and the colors of his clothing, creating a tranquil, leisurely, and nostalgic atmosphere full of the flavor of life.", + "seed": 17220, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_5/index_3219_seed_17220_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05111979860812425, + "vlm_logprob": -0.05111979860812425, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.05111979860812425, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037442419677972794 + ], + "opposite_logprob_tokens": [ + -3.787442445755005 + ], + "logprob_reward": -0.037442419677972794, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.037442419677972794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12039569020271301, + 0.07147695124149323, + 0.06714431941509247, + 0.07011257857084274, + 0.055136363953351974, + 0.08934573084115982, + 0.15425704419612885, + 0.2449919879436493, + 0.05152539908885956, + 0.07561398297548294 + ], + "credit_map": [ + 0.008116306737065315, + 0.007659868337213993, + 0.012629708275198936, + 0.01812160387635231, + 0.011518248356878757, + 0.016373232007026672, + 0.017535027116537094, + 0.028441693633794785, + 0.0043348404578864574, + 0.007639870513230562, + 0.006909607443958521, + 0.013205258175730705, + 0.010506951250135899, + 0.008555642329156399, + 0.011913587339222431, + 0.008411199785768986, + 0.005281743593513966, + 0.008321577683091164, + 0.008128322660923004, + 0.01053531002253294, + 0.010049534030258656, + 0.008477882482111454, + 0.01080225594341755, + 0.005547688342630863, + 0.006182257551699877, + 0.006349789910018444, + 0.009640037082135677, + 0.013809313997626305, + 0.009433564729988575, + 0.007110466714948416, + 0.009674116969108582, + 0.00791302602738142, + 0.007394444663077593, + 0.005855216644704342, + 0.007390571758151054, + 0.006221845280379057, + 0.008503950200974941, + 0.006503306794911623, + 0.005852089263498783, + 0.007414940278977156, + 0.004933683201670647, + 0.007060424890369177, + 0.01080517377704382, + 0.013005072250962257, + 0.008476777002215385, + 0.021471550688147545, + 0.010858775116503239, + 0.012734276242554188, + 0.007791025098413229, + 0.0042428127489984035, + 0.011787123046815395, + 0.005578831769526005, + 0.006624729372560978, + 0.061464715749025345, + 0.004628016613423824, + 0.05213978514075279, + 0.014936095103621483, + 0.03405311703681946, + 0.024519452825188637, + 0.039917752146720886, + 0.028746096417307854, + 0.020844124257564545, + 0.05531599745154381, + 0.026659352704882622, + 0.00612733606249094, + 0.01046791672706604, + 0.005306067876517773, + 0.005270384252071381, + 0.009214671328663826, + 0.005749750416725874, + 0.0038020997308194637, + 0.005587175954133272, + 0.0063642170280218124, + 0.009233874268829823, + 0.011113609187304974, + 0.008535901084542274, + 0.014305188320577145, + 0.00786536280065775, + 0.008562793023884296, + 0.00963302981108427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.037442419677972794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866386204957962 + ], + "opposite_logprob_tokens": [ + -3.048663854598999 + ], + "logprob_reward": -0.04866386204957962, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866386204957962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23793374001979828, + 0.08238933235406876, + 0.07344509661197662, + 0.04911946877837181, + 0.06423182040452957, + 0.06822893023490906, + 0.08898266404867172, + 0.19560708105564117, + 0.039084676653146744, + 0.10097716748714447 + ], + "credit_map": [ + 0.005355506669729948, + 0.011470806784927845, + 0.01622829958796501, + 0.019466830417513847, + 0.02516695111989975, + 0.017078209668397903, + 0.07760410755872726, + 0.06556302309036255, + 0.007438156753778458, + 0.006746935658156872, + 0.010134442709386349, + 0.013660488650202751, + 0.009871359914541245, + 0.011763143353164196, + 0.011391938664019108, + 0.011382865719497204, + 0.004280788358300924, + 0.010845181532204151, + 0.007903244346380234, + 0.016352973878383636, + 0.010727828368544579, + 0.006033643614500761, + 0.010939723812043667, + 0.006361716892570257, + 0.0039753178134560585, + 0.0053341928869485855, + 0.005570263136178255, + 0.012314860709011555, + 0.005907602142542601, + 0.005011708941310644, + 0.005687437951564789, + 0.005318082869052887, + 0.004597028251737356, + 0.006273467093706131, + 0.007196452002972364, + 0.007099600043147802, + 0.013964639976620674, + 0.008603928610682487, + 0.006594902370125055, + 0.009901799261569977, + 0.003420848399400711, + 0.006981879472732544, + 0.00901535339653492, + 0.00978563167154789, + 0.008593963459134102, + 0.012872765772044659, + 0.007497726008296013, + 0.010060764849185944, + 0.010038523934781551, + 0.013060607016086578, + 0.01000980380922556, + 0.008187063969671726, + 0.0073635815642774105, + 0.01863410696387291, + 0.005728594027459621, + 0.01596038229763508, + 0.011078793555498123, + 0.040744468569755554, + 0.01942470110952854, + 0.02191375195980072, + 0.03033585287630558, + 0.01983252353966236, + 0.02846626751124859, + 0.023810740560293198, + 0.0030582176987081766, + 0.011996092274785042, + 0.0034084999933838844, + 0.004526403266936541, + 0.0047034635208547115, + 0.003972145728766918, + 0.0038879727944731712, + 0.003531880909577012, + 0.006559383124113083, + 0.011105581186711788, + 0.013268370181322098, + 0.011477813124656677, + 0.021123889833688736, + 0.011603649705648422, + 0.015592173673212528, + 0.010246303863823414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.04866386204957962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192244037985802 + ], + "opposite_logprob_tokens": [ + -4.2641921043396 + ], + "logprob_reward": -0.014192244037985802, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192244037985802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10277828574180603, + 0.06699302047491074, + 0.23600345849990845, + 0.10946016013622284, + 0.08094494044780731, + 0.1288367211818695, + 0.06022860109806061, + 0.1191718801856041, + 0.04294804483652115, + 0.052634984254837036 + ], + "credit_map": [ + 0.004439693875610828, + 0.00976426899433136, + 0.010896407999098301, + 0.01893317699432373, + 0.011502436362206936, + 0.010879568755626678, + 0.02120061405003071, + 0.015162115916609764, + 0.005353918764740229, + 0.0051335422322154045, + 0.010928476229310036, + 0.014666170813143253, + 0.008484537713229656, + 0.007451674900949001, + 0.008662208914756775, + 0.006312488112598658, + 0.011561866849660873, + 0.0332166962325573, + 0.02978767827153206, + 0.06112150475382805, + 0.03182148560881615, + 0.01731140725314617, + 0.031368695199489594, + 0.019814111292362213, + 0.006091875024139881, + 0.009345076978206635, + 0.01662410795688629, + 0.036231160163879395, + 0.008992452174425125, + 0.010165992192924023, + 0.014053655788302422, + 0.007955842651426792, + 0.006277443841099739, + 0.01013458427041769, + 0.007396681699901819, + 0.010128209367394447, + 0.01571413315832615, + 0.010038222186267376, + 0.01108886394649744, + 0.010166806168854237, + 0.007070349063724279, + 0.008220406249165535, + 0.01623794622719288, + 0.01569269597530365, + 0.007057229056954384, + 0.03833691030740738, + 0.013004070147871971, + 0.023217106238007545, + 0.004473404958844185, + 0.004839498549699783, + 0.00467882351949811, + 0.004983526188880205, + 0.004757930524647236, + 0.016521254554390907, + 0.004181742202490568, + 0.015792420133948326, + 0.0072863101959228516, + 0.01981443352997303, + 0.012461846694350243, + 0.02082098089158535, + 0.014892006292939186, + 0.009637105278670788, + 0.023833423852920532, + 0.01042576041072607, + 0.0032440160866826773, + 0.0056387013755738735, + 0.004083074163645506, + 0.00505165895447135, + 0.0100852120667696, + 0.004876904655247927, + 0.005949396640062332, + 0.0040190815925598145, + 0.004520211834460497, + 0.00693177105858922, + 0.009454903192818165, + 0.007831646129488945, + 0.006996459793299437, + 0.005628907587379217, + 0.006639257073402405, + 0.00463182432577014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.014192244037985802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531115673482418 + ], + "opposite_logprob_tokens": [ + -4.387531280517578 + ], + "logprob_reward": -0.012531115673482418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531115673482418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2867654860019684, + 0.1059148907661438, + 0.060301586985588074, + 0.15124239027500153, + 0.022444769740104675, + 0.043825663626194, + 0.05372738465666771, + 0.09336761385202408, + 0.11296629160642624, + 0.06944392621517181 + ], + "credit_map": [ + 0.02040007710456848, + 0.020970765501260757, + 0.027238119393587112, + 0.032096944749355316, + 0.0570690855383873, + 0.041278108954429626, + 0.04577510058879852, + 0.04193727299571037, + 0.006881785113364458, + 0.006843348033726215, + 0.01307312399148941, + 0.013836851343512535, + 0.013802765868604183, + 0.014283780939877033, + 0.018840458244085312, + 0.018352780491113663, + 0.004255898762494326, + 0.006261724513024092, + 0.005621249787509441, + 0.007651702035218477, + 0.013585505075752735, + 0.0076584480702877045, + 0.008123225532472134, + 0.007143831346184015, + 0.015151824802160263, + 0.015723098069429398, + 0.016645830124616623, + 0.017013777047395706, + 0.023739362135529518, + 0.018352119252085686, + 0.016672415658831596, + 0.027943963184952736, + 0.001984470523893833, + 0.004345226567238569, + 0.0030807014554739, + 0.003610053798183799, + 0.0028522503562271595, + 0.0022359180729836226, + 0.0024247446563094854, + 0.0019114037277176976, + 0.003011768450960517, + 0.004476009868085384, + 0.007071787491440773, + 0.0066872574388980865, + 0.0038928543217480183, + 0.009422038681805134, + 0.0038839743938297033, + 0.005379972048103809, + 0.007614941336214542, + 0.0053838277235627174, + 0.00699967285618186, + 0.0038497678469866514, + 0.004070448223501444, + 0.012426280416548252, + 0.003795342054218054, + 0.009587102569639683, + 0.008077381178736687, + 0.023353124037384987, + 0.01051407027989626, + 0.010495240800082684, + 0.014913225546479225, + 0.009318282827734947, + 0.009029914624989033, + 0.007666370831429958, + 0.013258656486868858, + 0.02416445128619671, + 0.019723283126950264, + 0.011501049622893333, + 0.010373784229159355, + 0.01294703409075737, + 0.008644075132906437, + 0.012353967875242233, + 0.008374769240617752, + 0.009301981888711452, + 0.010802322067320347, + 0.00842045247554779, + 0.0075425258837640285, + 0.008337925188243389, + 0.006881088949739933, + 0.00978285912424326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.012531115673482418 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14276935160160065 + ], + "opposite_logprob_tokens": [ + -2.0177693367004395 + ], + "logprob_reward": -0.14276935160160065, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14276935160160065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3051471412181854, + 0.040803421288728714, + 0.04092218354344368, + 0.04399982467293739, + 0.0462443083524704, + 0.02121679298579693, + 0.15179769694805145, + 0.05878068506717682, + 0.22094011306762695, + 0.0701478123664856 + ], + "credit_map": [ + 0.007169699762016535, + 0.012522682547569275, + 0.01528262346982956, + 0.029179612174630165, + 0.02704145945608616, + 0.035059355199337006, + 0.08691415935754776, + 0.09197753667831421, + 0.002177781192585826, + 0.004253284074366093, + 0.004245372489094734, + 0.00464395759627223, + 0.00939780380576849, + 0.005004390608519316, + 0.005439148284494877, + 0.00564168393611908, + 0.00294628762640059, + 0.0036485393065959215, + 0.004988415166735649, + 0.0060945069417357445, + 0.007836021482944489, + 0.0048540555872023106, + 0.006543547380715609, + 0.004010808188468218, + 0.002983675803989172, + 0.003505266737192869, + 0.005704022943973541, + 0.006340615451335907, + 0.008649243041872978, + 0.0047218273393809795, + 0.005305363796651363, + 0.006789807695895433, + 0.0037067960947752, + 0.004329124465584755, + 0.005867500323802233, + 0.005576157011091709, + 0.0066282604821026325, + 0.006370848510414362, + 0.0048647732473909855, + 0.008900847285985947, + 0.001633013249374926, + 0.001995319500565529, + 0.0031047204975038767, + 0.003222673898562789, + 0.0032355180010199547, + 0.0029281931929290295, + 0.003203429514542222, + 0.0018939239671453834, + 0.007310147862881422, + 0.0047206333838403225, + 0.01085555087774992, + 0.007262262050062418, + 0.005239780060946941, + 0.05929335579276085, + 0.003749702125787735, + 0.053366273641586304, + 0.0057752118445932865, + 0.015790633857250214, + 0.005099398549646139, + 0.005867138039320707, + 0.008693505078554153, + 0.006245431490242481, + 0.0062389071099460125, + 0.0050704600289464, + 0.018802788108587265, + 0.0691947415471077, + 0.02556692622601986, + 0.02488594502210617, + 0.02841302566230297, + 0.01920551247894764, + 0.01779644377529621, + 0.0170747060328722, + 0.007336258422583342, + 0.009116333909332752, + 0.012606380507349968, + 0.010937770828604698, + 0.007842945866286755, + 0.007370220497250557, + 0.008869383484125137, + 0.006068523041903973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.14276935160160065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera follows a young man from a low angle, parallel to him, panning smoothly to the right to capture him as he briskly navigates the streets on his bicycle. The young man, with his short, slightly tousled dark brown hair fluttering gently in the breeze, wears a relaxed, cheerful smile. He wears a simple white cotton T-shirt over an unbuttoned olive green work shirt, slim-fit washed blue jeans, and classic white canvas shoes. He rides a retro-style city bike with a bright yellow frame, a brown leather seat and grips, and a beige canvas messenger bag hanging from the handlebars. He leans slightly forward, his hands resting effortlessly on the handlebars, his legs pedaling with a steady, rhythmic rhythm, propelling the bike smoothly forward over the uneven cobblestone road. They were walking through a narrow, winding alleyway in the ancient town, flanked by traditional Chinese houses with mottled whitewashed walls and green-tiled roofs. Verdant vines clung to the walls, and red lanterns occasionally hung from the eaves. The warm afternoon sun filtered through the gaps between the buildings above, casting dappled, playful shadows on the ground and on him. The dominant colors of the scene were the gray of the slate and the white of the walls, accented by the bright yellow of the bicycle, the green of the vines, and the colors of his clothing, creating a tranquil, leisurely, and nostalgic atmosphere full of the flavor of life.", + "seed": 17221, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_5/index_3219_seed_17221_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05546652693301439, + "vlm_logprob": -0.05546652693301439, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.05546652693301439, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03923511877655983 + ], + "opposite_logprob_tokens": [ + -4.1642351150512695 + ], + "logprob_reward": -0.03923511877655983, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03923511877655983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10220743715763092, + 0.04610472172498703, + 0.06577108800411224, + 0.02288958989083767, + 0.03608012571930885, + 0.11892378330230713, + 0.05312875285744667, + 0.11550980806350708, + 0.34396615624427795, + 0.0954185202717781 + ], + "credit_map": [ + 0.009926105849444866, + 0.009710230864584446, + 0.022813035175204277, + 0.01080353744328022, + 0.013206038624048233, + 0.013612608425319195, + 0.011294154450297356, + 0.01084172260016203, + 0.004937296267598867, + 0.004198883660137653, + 0.011355557478964329, + 0.004214162938296795, + 0.004884680733084679, + 0.005022518336772919, + 0.0054046716541051865, + 0.0060869487933814526, + 0.0061080786399543285, + 0.007779378909617662, + 0.009224859066307545, + 0.009192283265292645, + 0.010917924344539642, + 0.008008544333279133, + 0.006871570833027363, + 0.007668449077755213, + 0.002193449530750513, + 0.0032350383698940277, + 0.0025704947765916586, + 0.002868077950552106, + 0.0024388188030570745, + 0.002745498437434435, + 0.003992541693150997, + 0.002845670096576214, + 0.003916094079613686, + 0.005389315076172352, + 0.008939411491155624, + 0.003891345113515854, + 0.0031883581541478634, + 0.004219173453748226, + 0.0030106864869594574, + 0.003525740932673216, + 0.011666719801723957, + 0.02742326818406582, + 0.03653382882475853, + 0.00759090157225728, + 0.01156496349722147, + 0.009507880546152592, + 0.00593327172100544, + 0.00870294775813818, + 0.005838165059685707, + 0.008170978166162968, + 0.006275868508964777, + 0.007165994960814714, + 0.006574228405952454, + 0.008668387308716774, + 0.005135385785251856, + 0.005299742799252272, + 0.016460955142974854, + 0.01020937878638506, + 0.017722943797707558, + 0.010727960616350174, + 0.013312478549778461, + 0.03172748163342476, + 0.006809459533542395, + 0.00853915885090828, + 0.03809018433094025, + 0.023428095504641533, + 0.05049111694097519, + 0.051306091248989105, + 0.02915208600461483, + 0.04431530833244324, + 0.027582908049225807, + 0.07960037887096405, + 0.011337527073919773, + 0.009033901616930962, + 0.010502399876713753, + 0.014409159310162067, + 0.009292992763221264, + 0.010098105296492577, + 0.01809169352054596, + 0.012652735225856304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.03923511877655983 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810223564505577 + ], + "opposite_logprob_tokens": [ + -3.288102149963379 + ], + "logprob_reward": -0.03810223564505577, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810223564505577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20650403201580048, + 0.08334287256002426, + 0.07644446194171906, + 0.11514803767204285, + 0.05417404696345329, + 0.11073299497365952, + 0.08380301296710968, + 0.11416757851839066, + 0.11339862644672394, + 0.04228430241346359 + ], + "credit_map": [ + 0.013952625915408134, + 0.016094200313091278, + 0.03321760892868042, + 0.023994309827685356, + 0.023848019540309906, + 0.020366134122014046, + 0.03117302991449833, + 0.043858107179403305, + 0.007071756757795811, + 0.008911827579140663, + 0.013636669144034386, + 0.008987612091004848, + 0.010329828597605228, + 0.012013789266347885, + 0.013896575197577477, + 0.008494810201227665, + 0.007493935991078615, + 0.008429020643234253, + 0.012120763771235943, + 0.01237022876739502, + 0.010806403122842312, + 0.009307111613452435, + 0.008502534590661526, + 0.007414462044835091, + 0.009831439703702927, + 0.012495269998908043, + 0.018351761624217033, + 0.009469544515013695, + 0.014617991633713245, + 0.01484256237745285, + 0.02108367346227169, + 0.014455792494118214, + 0.005096414592117071, + 0.007002615835517645, + 0.009905653074383736, + 0.0052521368488669395, + 0.006961286999285221, + 0.006834762170910835, + 0.005512441508471966, + 0.007608737330883741, + 0.01812830939888954, + 0.018443910405039787, + 0.02216624654829502, + 0.007516978308558464, + 0.009253125637769699, + 0.01234806515276432, + 0.008395510725677013, + 0.014480849727988243, + 0.012354237958788872, + 0.016961699351668358, + 0.009648429229855537, + 0.009248259477317333, + 0.008039383217692375, + 0.010255485773086548, + 0.009882878512144089, + 0.007412636652588844, + 0.023108942434191704, + 0.01174463052302599, + 0.03347198665142059, + 0.008239026181399822, + 0.011405054479837418, + 0.011274161748588085, + 0.007843749597668648, + 0.007080026436597109, + 0.015296977013349533, + 0.008404918946325779, + 0.014457282610237598, + 0.015368525870144367, + 0.00990355759859085, + 0.012718893587589264, + 0.010667353868484497, + 0.026581110432744026, + 0.005269329063594341, + 0.005872667767107487, + 0.005654932465404272, + 0.004858001600950956, + 0.0059589785523712635, + 0.004584393464028835, + 0.0057253725826740265, + 0.0043606264516711235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.03810223564505577 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092651419341564 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092651419341564, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092651419341564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20423126220703125, + 0.08980301767587662, + 0.04994380846619606, + 0.07683999091386795, + 0.037374723702669144, + 0.12300441414117813, + 0.13690325617790222, + 0.10760168731212616, + 0.13118012249469757, + 0.043117690831422806 + ], + "credit_map": [ + 0.014416016638278961, + 0.02177044376730919, + 0.03697306290268898, + 0.029425395652651787, + 0.024616599082946777, + 0.017501549795269966, + 0.0338979996740818, + 0.02563020959496498, + 0.008967509493231773, + 0.007685784716159105, + 0.011664657853543758, + 0.008270114660263062, + 0.011678879149258137, + 0.009855947457253933, + 0.02298365719616413, + 0.008696465753018856, + 0.005620571784675121, + 0.0063365790992975235, + 0.006423394661396742, + 0.008461247198283672, + 0.006524631287902594, + 0.005868089385330677, + 0.005606785416603088, + 0.005102509632706642, + 0.008034526370465755, + 0.008637591265141964, + 0.0118997972458601, + 0.007675216533243656, + 0.008305537514388561, + 0.00894884578883648, + 0.013191796839237213, + 0.010146675631403923, + 0.005275677423924208, + 0.005192198324948549, + 0.005177427548915148, + 0.003758551785722375, + 0.0042497701942920685, + 0.0035693119280040264, + 0.005258827470242977, + 0.00489296019077301, + 0.017821379005908966, + 0.027431972324848175, + 0.030414029955863953, + 0.008571866899728775, + 0.012065347284078598, + 0.008638867177069187, + 0.009345032274723053, + 0.008715919218957424, + 0.013812031596899033, + 0.019776947796344757, + 0.015815744176506996, + 0.013087471015751362, + 0.014896505512297153, + 0.010095649398863316, + 0.03085276298224926, + 0.01856614090502262, + 0.01828920654952526, + 0.012320688925683498, + 0.021143600344657898, + 0.00914771482348442, + 0.015525322407484055, + 0.014746423810720444, + 0.008855226449668407, + 0.007573505397886038, + 0.015681346878409386, + 0.009190586395561695, + 0.022183634340763092, + 0.01981055736541748, + 0.010961451567709446, + 0.015575557947158813, + 0.01138933002948761, + 0.02638765051960945, + 0.006589867174625397, + 0.003703912952914834, + 0.005796045530587435, + 0.005278491880744696, + 0.004738706164062023, + 0.005628291051834822, + 0.005359689239412546, + 0.0060226875357329845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.004092651419341564 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605827920138836 + ], + "opposite_logprob_tokens": [ + -4.141058444976807 + ], + "logprob_reward": -0.01605827920138836, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605827920138836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.268564373254776, + 0.09045897424221039, + 0.051778580993413925, + 0.026689792051911354, + 0.024964086711406708, + 0.11281498521566391, + 0.029632139950990677, + 0.0664648562669754, + 0.19852274656295776, + 0.1301094889640808 + ], + "credit_map": [ + 0.01984049566090107, + 0.026718059554696083, + 0.042749639600515366, + 0.04033790901303291, + 0.034566305577754974, + 0.02955697849392891, + 0.03493744879961014, + 0.039857540279626846, + 0.010354017838835716, + 0.010183590464293957, + 0.023375991731882095, + 0.009269839152693748, + 0.011960712261497974, + 0.008089185692369938, + 0.009344123303890228, + 0.007881511002779007, + 0.005626526661217213, + 0.005792444571852684, + 0.007066451944410801, + 0.007485457696020603, + 0.006059123203158379, + 0.0059588076546788216, + 0.007411439437419176, + 0.0063783288933336735, + 0.0029943028930574656, + 0.002869731979444623, + 0.002302170032635331, + 0.002798727946355939, + 0.0038596028462052345, + 0.0032058937940746546, + 0.005476250313222408, + 0.003183112246915698, + 0.0035804680082947016, + 0.0031534263398498297, + 0.002860552631318569, + 0.002438134979456663, + 0.0037674603518098593, + 0.003867093473672867, + 0.0026748718228191137, + 0.002622079337015748, + 0.009732173755764961, + 0.028091834858059883, + 0.03947855532169342, + 0.007071611471474171, + 0.008465876802802086, + 0.008658673614263535, + 0.004309362731873989, + 0.007006898056715727, + 0.003140877466648817, + 0.003970182500779629, + 0.005159701220691204, + 0.003782257903367281, + 0.0029007040429860353, + 0.004198150709271431, + 0.0034939772449433804, + 0.0029862907249480486, + 0.017261704429984093, + 0.004105899482965469, + 0.021282268688082695, + 0.0037529149558395147, + 0.004649065434932709, + 0.00873094704002142, + 0.002831833204254508, + 0.0038502272218465805, + 0.02235385961830616, + 0.014263064600527287, + 0.034656230360269547, + 0.028975272551178932, + 0.01807466335594654, + 0.025672325864434242, + 0.017124062404036522, + 0.037403278052806854, + 0.020781733095645905, + 0.014022043906152248, + 0.01576048880815506, + 0.020710468292236328, + 0.011498942971229553, + 0.017085861414670944, + 0.01357698068022728, + 0.01667296327650547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.01605827920138836 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and panning smoothly to the right while following a young man on a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the young man wearing a white cotton T-shirt over an unbuttoned olive green work shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a retro-style city bike with a bright yellow frame and a beige canvas messenger bag on the handlebars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a narrow, winding alleyway in an ancient town with traditional Chinese houses featuring whitewashed walls and green-tiled roofs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red lanterns visible hanging from the eaves of the buildings in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17984434962272644 + ], + "opposite_logprob_tokens": [ + -1.8048443794250488 + ], + "logprob_reward": -0.17984434962272644, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17984434962272644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2527196407318115, + 0.18366988003253937, + 0.03665449470281601, + 0.03767487779259682, + 0.05271066352725029, + 0.12612739205360413, + 0.023357829079031944, + 0.0536421500146389, + 0.15217585861682892, + 0.08126720041036606 + ], + "credit_map": [ + 0.018860390409827232, + 0.019591644406318665, + 0.05374445393681526, + 0.023086804896593094, + 0.02906649000942707, + 0.02968365140259266, + 0.03231150656938553, + 0.04637472704052925, + 0.019218605011701584, + 0.029606293886899948, + 0.031191803514957428, + 0.01916743628680706, + 0.020568370819091797, + 0.018341924995183945, + 0.029045086354017258, + 0.016530372202396393, + 0.0034756609238684177, + 0.005131619982421398, + 0.005282711237668991, + 0.003560486249625683, + 0.006969166453927755, + 0.004246055148541927, + 0.004333273507654667, + 0.0036555228289216757, + 0.005744236521422863, + 0.004384090192615986, + 0.004856976680457592, + 0.0033316630870103836, + 0.0048869880847632885, + 0.0046117729507386684, + 0.005070116836577654, + 0.004789035767316818, + 0.005300929769873619, + 0.006757382769137621, + 0.009510359726846218, + 0.003874181304126978, + 0.007004197686910629, + 0.0067947437055408955, + 0.006744279991835356, + 0.00672459090128541, + 0.013285189867019653, + 0.0283826757222414, + 0.036736030131578445, + 0.007080230861902237, + 0.009800621308386326, + 0.00863056629896164, + 0.007569046691060066, + 0.014643036760389805, + 0.0032288343645632267, + 0.0029073399491608143, + 0.005485530011355877, + 0.002404755214229226, + 0.0016727170441299677, + 0.0034864649642258883, + 0.002308529568836093, + 0.0018636594759300351, + 0.013056594878435135, + 0.005071430467069149, + 0.007647543679922819, + 0.0037368398625403643, + 0.004965541884303093, + 0.013128766790032387, + 0.0025117737241089344, + 0.003523661522194743, + 0.018070803955197334, + 0.01184054184705019, + 0.01864716038107872, + 0.025353651493787766, + 0.015801990404725075, + 0.018188245594501495, + 0.01335070189088583, + 0.030922790989279747, + 0.014023268595337868, + 0.00981076993048191, + 0.012299594469368458, + 0.010875795967876911, + 0.008515730500221252, + 0.009328344836831093, + 0.0066563040018081665, + 0.009757393039762974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.17984434962272644 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_6/index_tensor([8167, 8167])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_6/index_tensor([8167, 8167])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ffe6a01297ccb7d790aa7a9cb5b5a82c5f4b8ad8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_6/index_tensor([8167, 8167])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static shot captures a towering, floor-to-ceiling bookshelf constructed of warm, honey-toned pine, lined with books in a sun-drenched log cabin. A silent, magical river of literature begins to cascade down from the upper shelves; hardcovers with frayed fabric spines and glossy paperbacks in a variety of colors tumble and slide against each other in not a chaotic collision but a fluid, flowing movement. The camera slowly tilts downward, carefully following the books as they fall onto the polished wooden floor, forming a winding stream that meanders through the room. Bright, golden afternoon sunlight streams in through a large window, illuminating the scene and capturing countless motes of dust dancing in the air. As the books drift across the floor, the pages of some open volumes flutter gently and deliberately, the thin paper catching the light, creating a serene and enchanting atmosphere of quiet, impossible magic within the peaceful, sun-drenched sanctuary.", + "seed": 9187, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_6/index_8167_seed_9187_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03960212394595146, + "vlm_logprob": -0.03960212394595146, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.03960212394595146, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07287811487913132 + ], + "opposite_logprob_tokens": [ + -2.822878122329712 + ], + "logprob_reward": -0.07287811487913132, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07287811487913132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19195398688316345, + 0.0823431983590126, + 0.05612294748425484, + 0.14252540469169617, + 0.039978958666324615, + 0.060238976031541824, + 0.07633432000875473, + 0.03081500343978405, + 0.2865332365036011, + 0.03315403684973717 + ], + "credit_map": [ + 0.03181037679314613, + 0.0215576384216547, + 0.023519203066825867, + 0.02434936724603176, + 0.019292859360575676, + 0.02067233994603157, + 0.023307209834456444, + 0.027444977313280106, + 0.008851381950080395, + 0.0072217779234051704, + 0.00765924621373415, + 0.008836215361952782, + 0.015460140071809292, + 0.01298756618052721, + 0.011040310375392437, + 0.010286548174917698, + 0.005430853925645351, + 0.0069972192868590355, + 0.00532541936263442, + 0.0057915206998586655, + 0.009966589510440826, + 0.010032366961240768, + 0.006599354092031717, + 0.005979619454592466, + 0.015353413298726082, + 0.021545397117733955, + 0.024886377155780792, + 0.011757421307265759, + 0.016615895554423332, + 0.016177291050553322, + 0.015509738586843014, + 0.020679844543337822, + 0.0036816394422203302, + 0.0028284923173487186, + 0.00760896410793066, + 0.004423077683895826, + 0.004125707782804966, + 0.005957497749477625, + 0.005789936054497957, + 0.005563637241721153, + 0.007402964401990175, + 0.0035336159635335207, + 0.0076525406911969185, + 0.012628651224076748, + 0.004254053346812725, + 0.012906319461762905, + 0.005856805481016636, + 0.006004018243402243, + 0.009497626684606075, + 0.008138800039887428, + 0.011475678533315659, + 0.011976185254752636, + 0.006500516086816788, + 0.009152604267001152, + 0.012999186292290688, + 0.006593711208552122, + 0.0035606150049716234, + 0.0029038090724498034, + 0.00520040700212121, + 0.0038332862313836813, + 0.0017833213787525892, + 0.0038394189905375242, + 0.004781992174685001, + 0.004912149161100388, + 0.014611568301916122, + 0.02069527469575405, + 0.02362510934472084, + 0.08959253132343292, + 0.015897361561655998, + 0.012648165225982666, + 0.02500983327627182, + 0.08445335924625397, + 0.004198180977255106, + 0.003554453607648611, + 0.006146587897092104, + 0.003819498233497143, + 0.0034067779779434204, + 0.0026176737155765295, + 0.004063062835484743, + 0.005347798112779856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.07287811487913132 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381559394299984 + ], + "opposite_logprob_tokens": [ + -3.288156032562256 + ], + "logprob_reward": -0.0381559394299984, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381559394299984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1491396725177765, + 0.10562724620103836, + 0.04607260599732399, + 0.09222385287284851, + 0.04834461212158203, + 0.09084863215684891, + 0.04502907767891884, + 0.09763459861278534, + 0.24297964572906494, + 0.08210005611181259 + ], + "credit_map": [ + 0.018464282155036926, + 0.01411391980946064, + 0.01611139439046383, + 0.015773797407746315, + 0.016077201813459396, + 0.027135582640767097, + 0.01819014549255371, + 0.02327335812151432, + 0.01585233397781849, + 0.008534309454262257, + 0.01792330853641033, + 0.008182565681636333, + 0.019845619797706604, + 0.013877615332603455, + 0.010832914151251316, + 0.010578585788607597, + 0.003994911443442106, + 0.004340336658060551, + 0.003803822211921215, + 0.00373928458429873, + 0.01186070591211319, + 0.006129457149654627, + 0.005025890190154314, + 0.007178195286542177, + 0.010659495368599892, + 0.010722021572291851, + 0.013512895442545414, + 0.008398880250751972, + 0.010897382162511349, + 0.013549042865633965, + 0.010215947404503822, + 0.014268186874687672, + 0.005013711284846067, + 0.004561977926641703, + 0.008364989422261715, + 0.004638585727661848, + 0.004582180175930262, + 0.0075636752881109715, + 0.006096805911511183, + 0.007522685453295708, + 0.01281858328729868, + 0.0063344999216496944, + 0.01382144819945097, + 0.020540988072752953, + 0.0058111948892474174, + 0.015275425277650356, + 0.008814042434096336, + 0.0074324537999928, + 0.007680450100451708, + 0.0051494138315320015, + 0.004927029833197594, + 0.006790141575038433, + 0.00411165039986372, + 0.007116673514246941, + 0.004730318207293749, + 0.00452340068295598, + 0.007140832953155041, + 0.010406922549009323, + 0.013929814100265503, + 0.014365773648023605, + 0.0052097090519964695, + 0.016107844188809395, + 0.010830898769199848, + 0.019642801955342293, + 0.01969737932085991, + 0.017977967858314514, + 0.034204449504613876, + 0.0631103590130806, + 0.01904142089188099, + 0.014486102387309074, + 0.02352011762559414, + 0.050941839814186096, + 0.008554833941161633, + 0.012616951018571854, + 0.010592068545520306, + 0.011309142224490643, + 0.00844207126647234, + 0.008281323127448559, + 0.011888600885868073, + 0.010415065102279186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0381559394299984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043081097304821014 + ], + "opposite_logprob_tokens": [ + -3.168081045150757 + ], + "logprob_reward": -0.043081097304821014, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043081097304821014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1900644153356552, + 0.056410402059555054, + 0.05153711885213852, + 0.054168447852134705, + 0.052837975323200226, + 0.06891177594661713, + 0.06520155817270279, + 0.1859574317932129, + 0.18335233628749847, + 0.0915585532784462 + ], + "credit_map": [ + 0.03358246758580208, + 0.01827363669872284, + 0.023212993517518044, + 0.021739425137639046, + 0.017932284623384476, + 0.022055761888623238, + 0.02588008902966976, + 0.027387745678424835, + 0.007400193717330694, + 0.004841092508286238, + 0.006110748276114464, + 0.006465874146670103, + 0.007440551184117794, + 0.008054090663790703, + 0.009429987519979477, + 0.006667868234217167, + 0.003919727634638548, + 0.0072076451033353806, + 0.005289417691528797, + 0.005979551468044519, + 0.008164149709045887, + 0.0077939629554748535, + 0.005032572895288467, + 0.008150092326104641, + 0.006367192603647709, + 0.009330649860203266, + 0.006649541202932596, + 0.004419874865561724, + 0.005502010695636272, + 0.008483177982270718, + 0.006663701497018337, + 0.006752300076186657, + 0.006733031943440437, + 0.004891139455139637, + 0.01051340438425541, + 0.006432961206883192, + 0.0029699201695621014, + 0.009241696447134018, + 0.004037069156765938, + 0.008018751628696918, + 0.010460587218403816, + 0.0055891890078783035, + 0.007664297707378864, + 0.006737080868333578, + 0.005899770651012659, + 0.017459983006119728, + 0.007277274504303932, + 0.00782359205186367, + 0.008741292171180248, + 0.006205819547176361, + 0.009419862180948257, + 0.011595329269766808, + 0.004510290455073118, + 0.006336209364235401, + 0.008999438025057316, + 0.009393316693603992, + 0.011620640754699707, + 0.01713525876402855, + 0.026525259017944336, + 0.028203945606946945, + 0.010416238568723202, + 0.031730860471725464, + 0.0192401185631752, + 0.04108509421348572, + 0.021848566830158234, + 0.01866999641060829, + 0.035366933792829514, + 0.032787103205919266, + 0.016862424090504646, + 0.018974455073475838, + 0.01801462471485138, + 0.02082821913063526, + 0.009430757723748684, + 0.011728924699127674, + 0.01530288252979517, + 0.010645410977303982, + 0.006775086745619774, + 0.007029079832136631, + 0.017485346645116806, + 0.0131610669195652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.043081097304821014 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020599357783794403 + ], + "opposite_logprob_tokens": [ + -3.895599365234375 + ], + "logprob_reward": -0.020599357783794403, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020599357783794403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05567893385887146, + 0.09807945042848587, + 0.09929980337619781, + 0.08084625005722046, + 0.03431384265422821, + 0.1388498842716217, + 0.08603668957948685, + 0.09236529469490051, + 0.2766379415988922, + 0.03789195418357849 + ], + "credit_map": [ + 0.008734305389225483, + 0.006750110536813736, + 0.005921903997659683, + 0.006691078655421734, + 0.005695454776287079, + 0.007173576392233372, + 0.0059012058191001415, + 0.00881129503250122, + 0.016960343345999718, + 0.006845502648502588, + 0.02070813998579979, + 0.010095814242959023, + 0.0086193373426795, + 0.010214024223387241, + 0.012437071651220322, + 0.012199205346405506, + 0.007707170210778713, + 0.01196583267301321, + 0.010885612107813358, + 0.008972224779427052, + 0.024305863305926323, + 0.01423147227615118, + 0.009891706518828869, + 0.011339915916323662, + 0.0067397779785096645, + 0.025278856977820396, + 0.010615277104079723, + 0.006743364967405796, + 0.009206466376781464, + 0.008301561698317528, + 0.006665690336376429, + 0.007295252755284309, + 0.0034488386008888483, + 0.0032581561245024204, + 0.005757798906415701, + 0.00447714002802968, + 0.002888195216655731, + 0.005596998147666454, + 0.004271415527909994, + 0.004615296144038439, + 0.01683024875819683, + 0.007334867492318153, + 0.025138208642601967, + 0.01718626730144024, + 0.0085899131372571, + 0.0356348492205143, + 0.010217737406492233, + 0.017917778342962265, + 0.011530080810189247, + 0.009396897628903389, + 0.010656311176717281, + 0.01417098380625248, + 0.005569281056523323, + 0.010749628767371178, + 0.015982158482074738, + 0.007981334812939167, + 0.005619762931019068, + 0.007157288026064634, + 0.012519216164946556, + 0.013652852736413479, + 0.006415226962417364, + 0.013652144931256771, + 0.015280219726264477, + 0.018068572506308556, + 0.017697103321552277, + 0.024496398866176605, + 0.024959886446595192, + 0.07556598633527756, + 0.01590537093579769, + 0.013088572770357132, + 0.024230046197772026, + 0.08069454878568649, + 0.004510826896876097, + 0.0038229210767894983, + 0.005559536628425121, + 0.006456015631556511, + 0.0031525427475571632, + 0.0036559959407895803, + 0.00572325149551034, + 0.005010863300412893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.020599357783794403 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023296110332012177 + ], + "opposite_logprob_tokens": [ + -3.7732961177825928 + ], + "logprob_reward": -0.023296110332012177, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023296110332012177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10752563923597336, + 0.07581904530525208, + 0.018292728811502457, + 0.05106031894683838, + 0.03563637658953667, + 0.037868235260248184, + 0.0513911210000515, + 0.11060450226068497, + 0.46567201614379883, + 0.046129997819662094 + ], + "credit_map": [ + 0.016062337905168533, + 0.011967523023486137, + 0.013859730213880539, + 0.012611191719770432, + 0.010587192140519619, + 0.014794008806347847, + 0.013185277581214905, + 0.014458381570875645, + 0.012295177206397057, + 0.007410747464746237, + 0.008815715089440346, + 0.006905163172632456, + 0.012434146367013454, + 0.008289248682558537, + 0.011063922196626663, + 0.008604923263192177, + 0.001979239983484149, + 0.0020403617527335882, + 0.0017248551594093442, + 0.0018944633193314075, + 0.0027302983216941357, + 0.002653432311490178, + 0.0025034977588802576, + 0.002766580320894718, + 0.0058198790065944195, + 0.014946148730814457, + 0.010109823197126389, + 0.0033240735065191984, + 0.005431501194834709, + 0.005615041125565767, + 0.0022467360831797123, + 0.003567120525985956, + 0.0042096818797290325, + 0.004603755194693804, + 0.004649484064429998, + 0.0035261367447674274, + 0.003129989840090275, + 0.008211597800254822, + 0.0029866325203329325, + 0.004319098312407732, + 0.006261887960135937, + 0.002314404584467411, + 0.00442416500300169, + 0.008025387302041054, + 0.0026020784862339497, + 0.005837905220687389, + 0.0041832770220935345, + 0.0042191301472485065, + 0.006880741100758314, + 0.0050946930423378944, + 0.0073679969646036625, + 0.009424515999853611, + 0.0034686685539782047, + 0.004977007862180471, + 0.005971682723611593, + 0.00820581242442131, + 0.009264860302209854, + 0.009509380906820297, + 0.020732859149575233, + 0.015959568321704865, + 0.006391193717718124, + 0.013373314402997494, + 0.01596957817673683, + 0.01940375566482544, + 0.023274438455700874, + 0.03382885083556175, + 0.055792249739170074, + 0.15311899781227112, + 0.029439762234687805, + 0.0259451512247324, + 0.0400603711605072, + 0.10421217978000641, + 0.006728396750986576, + 0.0061121415346860886, + 0.009061271324753761, + 0.005214260891079903, + 0.004136171191930771, + 0.0031929670367389917, + 0.005285811610519886, + 0.006398973520845175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.023296110332012177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot captures a towering, floor-to-ceiling bookshelf constructed of warm, honey-toned pine, lined with books in a sun-drenched log cabin. A silent, magical river of literature begins to cascade down from the upper shelves; hardcovers with frayed fabric spines and glossy paperbacks in a variety of colors tumble and slide against each other in not a chaotic collision but a fluid, flowing movement. The camera slowly tilts downward, carefully following the books as they fall onto the polished wooden floor, forming a winding stream that meanders through the room. Bright, golden afternoon sunlight streams in through a large window, illuminating the scene and capturing countless motes of dust dancing in the air. As the books drift across the floor, the pages of some open volumes flutter gently and deliberately, the thin paper catching the light, creating a serene and enchanting atmosphere of quiet, impossible magic within the peaceful, sun-drenched sanctuary.", + "seed": 9188, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_6/index_8167_seed_9188_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022446887753903865, + "vlm_logprob": -0.022446887753903865, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.022446887753903865, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024902723729610443 + ], + "opposite_logprob_tokens": [ + -4.024902820587158 + ], + "logprob_reward": -0.024902723729610443, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.024902723729610443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.182792067527771, + 0.09007933735847473, + 0.03927340731024742, + 0.05854778736829758, + 0.03938090056180954, + 0.14588136970996857, + 0.10325299948453903, + 0.07794161140918732, + 0.18746879696846008, + 0.07538164407014847 + ], + "credit_map": [ + 0.043665047734975815, + 0.016708673909306526, + 0.033258918672800064, + 0.013016512617468834, + 0.01802912913262844, + 0.019862113520503044, + 0.014226236380636692, + 0.024025442078709602, + 0.01339517068117857, + 0.006602129433304071, + 0.01512348186224699, + 0.012847482226788998, + 0.008376172743737698, + 0.013684181496500969, + 0.009628424420952797, + 0.010422303341329098, + 0.003950111567974091, + 0.0029245216865092516, + 0.006643669679760933, + 0.0025379329454153776, + 0.0037125847302377224, + 0.0068949200212955475, + 0.004043500404804945, + 0.008566172793507576, + 0.0071745337918400764, + 0.0034081267658621073, + 0.011002200655639172, + 0.004234505817294121, + 0.0071202428080141544, + 0.009216918610036373, + 0.008073548786342144, + 0.008317714557051659, + 0.004363524727523327, + 0.003862199140712619, + 0.006627811584621668, + 0.0036068595945835114, + 0.005524509120732546, + 0.006883128080517054, + 0.004279086831957102, + 0.00423378124833107, + 0.012267802841961384, + 0.009746668860316277, + 0.02202465571463108, + 0.010146986693143845, + 0.02473130077123642, + 0.02342604659497738, + 0.024027012288570404, + 0.019510909914970398, + 0.011843744665384293, + 0.008017381653189659, + 0.02307828888297081, + 0.007203059736639261, + 0.01381602417677641, + 0.007283979095518589, + 0.010947631672024727, + 0.02106289751827717, + 0.01527527067810297, + 0.010041067376732826, + 0.009415177628397942, + 0.0081791328266263, + 0.010442997328937054, + 0.008578669279813766, + 0.006791798397898674, + 0.009217496030032635, + 0.03516441211104393, + 0.014764709398150444, + 0.02040986716747284, + 0.017602985724806786, + 0.023860221728682518, + 0.01798149198293686, + 0.01293451339006424, + 0.044750601053237915, + 0.010983318090438843, + 0.006363352295011282, + 0.011119192466139793, + 0.006124429404735565, + 0.014635068364441395, + 0.010672468692064285, + 0.00667569600045681, + 0.008808123879134655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.024902723729610443 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374754637479782 + ], + "opposite_logprob_tokens": [ + -3.408747434616089 + ], + "logprob_reward": -0.03374754637479782, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374754637479782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14685499668121338, + 0.09505774080753326, + 0.049917854368686676, + 0.07394582778215408, + 0.037572238594293594, + 0.03716180473566055, + 0.24563078582286835, + 0.058342739939689636, + 0.10294895619153976, + 0.152567058801651 + ], + "credit_map": [ + 0.022992709651589394, + 0.014663034118711948, + 0.019673962146043777, + 0.013795166276395321, + 0.019955480471253395, + 0.01757223717868328, + 0.01458575576543808, + 0.023616664111614227, + 0.021448131650686264, + 0.006767576094716787, + 0.023935062810778618, + 0.006908383686095476, + 0.008899044245481491, + 0.010373161174356937, + 0.007903791964054108, + 0.008822591975331306, + 0.004758094437420368, + 0.0038249485660344362, + 0.006131882779300213, + 0.0037070694379508495, + 0.0062180450186133385, + 0.008310838602483273, + 0.005000207107514143, + 0.01196676678955555, + 0.007282203063368797, + 0.009035316295921803, + 0.007753871846944094, + 0.007797806058079004, + 0.01456222403794527, + 0.009447451680898666, + 0.009823642671108246, + 0.008243313990533352, + 0.00430096173658967, + 0.003085613949224353, + 0.00841639656573534, + 0.003976008854806423, + 0.004532671999186277, + 0.00563467713072896, + 0.0036312108859419823, + 0.003994698636233807, + 0.0029131723567843437, + 0.003704026574268937, + 0.005807932466268539, + 0.005210641771554947, + 0.0052807386964559555, + 0.004192420281469822, + 0.004921847488731146, + 0.00513102300465107, + 0.029126355424523354, + 0.01469914335757494, + 0.06849586218595505, + 0.021131278946995735, + 0.024812709540128708, + 0.017153466120362282, + 0.02508242055773735, + 0.04512955620884895, + 0.01430533267557621, + 0.004242219962179661, + 0.0075502898544073105, + 0.0057427347637712955, + 0.007180366199463606, + 0.006519571412354708, + 0.0048217796720564365, + 0.007980446331202984, + 0.017763566225767136, + 0.008257493376731873, + 0.010487192310392857, + 0.007680912502110004, + 0.013436829671263695, + 0.015266353264451027, + 0.007872862741351128, + 0.022183751687407494, + 0.019571207463741302, + 0.01909920945763588, + 0.02340785041451454, + 0.015773581340909004, + 0.023630905896425247, + 0.01754435896873474, + 0.016829947009682655, + 0.016710001975297928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.03374754637479782 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023299604654312134 + ], + "opposite_logprob_tokens": [ + -3.7732996940612793 + ], + "logprob_reward": -0.023299604654312134, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023299604654312134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1602362096309662, + 0.12376787513494492, + 0.03653310239315033, + 0.08756450563669205, + 0.03541119769215584, + 0.08780042082071304, + 0.10545165091753006, + 0.1418161392211914, + 0.11493410170078278, + 0.10648476332426071 + ], + "credit_map": [ + 0.025687186047434807, + 0.008146310225129128, + 0.028158819302916527, + 0.012487105093896389, + 0.02211111970245838, + 0.022339224815368652, + 0.01694743148982525, + 0.024359045550227165, + 0.022569453343749046, + 0.010778140276670456, + 0.015728143975138664, + 0.019941985607147217, + 0.010862895287573338, + 0.017294878140091896, + 0.01327427290380001, + 0.013318111188709736, + 0.0043046134524047375, + 0.0029470704030245543, + 0.004788302816450596, + 0.00279831001535058, + 0.006148809567093849, + 0.005285760387778282, + 0.004653565119951963, + 0.005606673192232847, + 0.008947777561843395, + 0.007615615148097277, + 0.022274194285273552, + 0.010851483792066574, + 0.010348907671868801, + 0.011940699070692062, + 0.008602645248174667, + 0.006983189843595028, + 0.0031234044581651688, + 0.0027567255310714245, + 0.005539632868021727, + 0.0030740343499928713, + 0.006202869117259979, + 0.006141364108771086, + 0.0044800033792853355, + 0.004093172028660774, + 0.007196118589490652, + 0.008382884785532951, + 0.011740152724087238, + 0.010447606444358826, + 0.012652288191020489, + 0.01089635118842125, + 0.013169816695153713, + 0.013315216638147831, + 0.01148831844329834, + 0.008071149699389935, + 0.02024579979479313, + 0.01097630150616169, + 0.016013193875551224, + 0.01013235468417406, + 0.008775928989052773, + 0.019748609513044357, + 0.02088235504925251, + 0.01374140940606594, + 0.015908822417259216, + 0.020777134224772453, + 0.02740705944597721, + 0.017920158803462982, + 0.01261377613991499, + 0.012565432116389275, + 0.019150031730532646, + 0.01596192456781864, + 0.00970066525042057, + 0.010748220607638359, + 0.01324677374213934, + 0.016480350866913795, + 0.008170252665877342, + 0.021475890651345253, + 0.01205856166779995, + 0.017648136243224144, + 0.014722657389938831, + 0.010689008049666882, + 0.011450967751443386, + 0.01501796767115593, + 0.01259232684969902, + 0.012305151671171188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.023299604654312134 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081271693110466 + ], + "opposite_logprob_tokens": [ + -4.1410813331604 + ], + "logprob_reward": -0.016081271693110466, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081271693110466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08491270989179611, + 0.0738106444478035, + 0.04283139854669571, + 0.08519294112920761, + 0.07075300067663193, + 0.11379349231719971, + 0.20114187896251678, + 0.12593093514442444, + 0.16169629991054535, + 0.039936650544404984 + ], + "credit_map": [ + 0.01656683348119259, + 0.0055557070299983025, + 0.015757905319333076, + 0.005203955806791782, + 0.012056231498718262, + 0.012720529921352863, + 0.007253843359649181, + 0.009797700680792332, + 0.007452338468283415, + 0.0074425945058465, + 0.007165240123867989, + 0.013253556564450264, + 0.007655698340386152, + 0.011006679385900497, + 0.008396457880735397, + 0.011438075453042984, + 0.005642290227115154, + 0.0031155485194176435, + 0.008800751529633999, + 0.0025099331978708506, + 0.004760097246617079, + 0.006321124266833067, + 0.004232896491885185, + 0.007448755670338869, + 0.00997072085738182, + 0.009105007164180279, + 0.01616744138300419, + 0.015338505618274212, + 0.005669326987117529, + 0.010317305102944374, + 0.01067479606717825, + 0.007949840277433395, + 0.0058144559152424335, + 0.005877618677914143, + 0.011792694218456745, + 0.00677682738751173, + 0.012208385393023491, + 0.01148675475269556, + 0.007531088311225176, + 0.009265175089240074, + 0.011211305856704712, + 0.00836162082850933, + 0.018259674310684204, + 0.012125748209655285, + 0.016739187762141228, + 0.017517641186714172, + 0.011827651411294937, + 0.017750663682818413, + 0.017500562593340874, + 0.012791968882083893, + 0.03614276275038719, + 0.027211608365178108, + 0.01606805995106697, + 0.029353972524404526, + 0.027595166116952896, + 0.03447776287794113, + 0.025049209594726562, + 0.01125205960124731, + 0.014985746704041958, + 0.02271432615816593, + 0.013455572538077831, + 0.00979837030172348, + 0.012643721885979176, + 0.016031930223107338, + 0.02612071856856346, + 0.014645522460341454, + 0.01906026527285576, + 0.013659742660820484, + 0.02302984520792961, + 0.025388209149241447, + 0.014484805054962635, + 0.025307193398475647, + 0.004964351654052734, + 0.004840787034481764, + 0.004246759694069624, + 0.004891310352832079, + 0.007223485969007015, + 0.005741138942539692, + 0.0038678557612001896, + 0.004160958342254162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.016081271693110466 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203292317688465 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014203292317688465, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203292317688465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09407717734575272, + 0.0560629740357399, + 0.02980133891105652, + 0.043945055454969406, + 0.05300920829176903, + 0.0783638283610344, + 0.09621194750070572, + 0.3040064871311188, + 0.06237668916583061, + 0.18214526772499084 + ], + "credit_map": [ + 0.014047556556761265, + 0.009739390574395657, + 0.011798177845776081, + 0.009605880826711655, + 0.011511928401887417, + 0.011998284608125687, + 0.009089669212698936, + 0.01628628745675087, + 0.009761102497577667, + 0.006190719548612833, + 0.009698249399662018, + 0.005728616379201412, + 0.004844037815928459, + 0.006299538072198629, + 0.005039731971919537, + 0.008500977419316769, + 0.003265428589656949, + 0.0023241008166223764, + 0.004185468889772892, + 0.002559977350756526, + 0.005517013370990753, + 0.003692039754241705, + 0.0029372316785156727, + 0.0053200796246528625, + 0.00534100504592061, + 0.0027948494534939528, + 0.0064263902604579926, + 0.005740791093558073, + 0.005152806174010038, + 0.007062437012791634, + 0.005936515051871538, + 0.005490259733051062, + 0.003301959950476885, + 0.003976058214902878, + 0.0074175940826535225, + 0.004752594977617264, + 0.010993722826242447, + 0.010316958650946617, + 0.005942031741142273, + 0.00630828645080328, + 0.004249176010489464, + 0.006673469208180904, + 0.010580932721495628, + 0.007112558931112289, + 0.0133390873670578, + 0.010823700577020645, + 0.014346031472086906, + 0.011238872073590755, + 0.01121118851006031, + 0.007154843769967556, + 0.020526546984910965, + 0.006460141856223345, + 0.012449484318494797, + 0.006254309788346291, + 0.010176879353821278, + 0.021978553384542465, + 0.06420586258172989, + 0.027728697285056114, + 0.03648859262466431, + 0.03041977435350418, + 0.03773907199501991, + 0.03403819724917412, + 0.031076177954673767, + 0.04231008142232895, + 0.011094444431364536, + 0.0045175496488809586, + 0.008926578797399998, + 0.004811735823750496, + 0.007858099415898323, + 0.007693079765886068, + 0.004165970254689455, + 0.01330923568457365, + 0.024490198120474815, + 0.018397202715277672, + 0.028211865574121475, + 0.01781216450035572, + 0.029045697301626205, + 0.023418720811605453, + 0.020644718781113625, + 0.020124712958931923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.014203292317688465 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot captures a towering, floor-to-ceiling bookshelf constructed of warm, honey-toned pine, lined with books in a sun-drenched log cabin. A silent, magical river of literature begins to cascade down from the upper shelves; hardcovers with frayed fabric spines and glossy paperbacks in a variety of colors tumble and slide against each other in not a chaotic collision but a fluid, flowing movement. The camera slowly tilts downward, carefully following the books as they fall onto the polished wooden floor, forming a winding stream that meanders through the room. Bright, golden afternoon sunlight streams in through a large window, illuminating the scene and capturing countless motes of dust dancing in the air. As the books drift across the floor, the pages of some open volumes flutter gently and deliberately, the thin paper catching the light, creating a serene and enchanting atmosphere of quiet, impossible magic within the peaceful, sun-drenched sanctuary.", + "seed": 9189, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_6/index_8167_seed_9189_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02355513200163841, + "vlm_logprob": -0.02355513200163841, + "vlm_yes_no_margin": 3.849999952316284, + "vlm_token_credit": -0.02355513200163841, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026796946302056313 + ], + "opposite_logprob_tokens": [ + -4.026796817779541 + ], + "logprob_reward": -0.026796946302056313, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.026796946302056313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14862626791000366, + 0.06422463059425354, + 0.15403130650520325, + 0.04836595430970192, + 0.1489345282316208, + 0.08226894587278366, + 0.07799060642719269, + 0.07183237373828888, + 0.11140324175357819, + 0.09232215583324432 + ], + "credit_map": [ + 0.016934625804424286, + 0.01828615926206112, + 0.019854549318552017, + 0.02210414968430996, + 0.017008665949106216, + 0.019262392073869705, + 0.012691006995737553, + 0.02248471789062023, + 0.004849675111472607, + 0.006131165660917759, + 0.005738983396440744, + 0.007852954789996147, + 0.008034974336624146, + 0.010649139061570168, + 0.010143694467842579, + 0.010824044235050678, + 0.016148416325449944, + 0.01914341188967228, + 0.018708549439907074, + 0.017301321029663086, + 0.01630409061908722, + 0.023200741037726402, + 0.021643508225679398, + 0.02158127725124359, + 0.003815396223217249, + 0.007545581553131342, + 0.00852610357105732, + 0.00479520671069622, + 0.00420655682682991, + 0.0074847061187028885, + 0.0034443619661033154, + 0.008548039942979813, + 0.010901214554905891, + 0.018251221626996994, + 0.019639933481812477, + 0.015870407223701477, + 0.02727600187063217, + 0.023855911567807198, + 0.014969698153436184, + 0.018170133233070374, + 0.006506144534796476, + 0.010050925426185131, + 0.012946078553795815, + 0.0073835477232933044, + 0.010138822719454765, + 0.01285062450915575, + 0.010104222223162651, + 0.012288580648601055, + 0.007691412698477507, + 0.01124020665884018, + 0.009846740402281284, + 0.01410767063498497, + 0.008028247393667698, + 0.010893400758504868, + 0.007634054869413376, + 0.008548873476684093, + 0.010137400589883327, + 0.007217278238385916, + 0.008770164102315903, + 0.007831203751266003, + 0.007964269258081913, + 0.011054781265556812, + 0.01045689545571804, + 0.008400380611419678, + 0.014505849219858646, + 0.01071156281977892, + 0.011816604994237423, + 0.01569097861647606, + 0.01836264505982399, + 0.023030104115605354, + 0.008515341207385063, + 0.008770150132477283, + 0.011242363601922989, + 0.010802354663610458, + 0.007624487392604351, + 0.010668762028217316, + 0.008889634162187576, + 0.013238869607448578, + 0.008741919882595539, + 0.02111375890672207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.026796946302056313 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638125978410244 + ], + "opposite_logprob_tokens": [ + -3.651381254196167 + ], + "logprob_reward": -0.02638125978410244, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638125978410244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16629132628440857, + 0.09405297040939331, + 0.06639345735311508, + 0.07145383954048157, + 0.10453785955905914, + 0.0922836884856224, + 0.04553103819489479, + 0.06309842318296432, + 0.11351212114095688, + 0.18284524977207184 + ], + "credit_map": [ + 0.019688118249177933, + 0.02566840872168541, + 0.011762592010200024, + 0.01603584736585617, + 0.025466270744800568, + 0.023652445524930954, + 0.014467094093561172, + 0.02955056168138981, + 0.006883283145725727, + 0.007492812350392342, + 0.009645013138651848, + 0.012610266916453838, + 0.013294327072799206, + 0.014208401553332806, + 0.013210502453148365, + 0.016708366572856903, + 0.004551048390567303, + 0.009819201193749905, + 0.004419656004756689, + 0.01185330655425787, + 0.008897831663489342, + 0.009062726981937885, + 0.010223695077002048, + 0.0075659919530153275, + 0.00642899377271533, + 0.008616614155471325, + 0.008650613017380238, + 0.011442355811595917, + 0.005921879317611456, + 0.013689654879271984, + 0.005655989516526461, + 0.01104774046689272, + 0.007781872525811195, + 0.015135983005166054, + 0.012045434676110744, + 0.010845615528523922, + 0.014805559068918228, + 0.016394607722759247, + 0.009720869362354279, + 0.017807919532060623, + 0.006718601565808058, + 0.010691645555198193, + 0.015065412037074566, + 0.0065744780004024506, + 0.011924679391086102, + 0.016775939613580704, + 0.007104164455085993, + 0.017428770661354065, + 0.005789709277451038, + 0.004813825245946646, + 0.006092258729040623, + 0.005808274261653423, + 0.00497407978400588, + 0.005956167820841074, + 0.0077045271173119545, + 0.004392197355628014, + 0.006645932327955961, + 0.007323635276407003, + 0.007392173632979393, + 0.006765326019376516, + 0.011753503233194351, + 0.008249022997915745, + 0.008524110540747643, + 0.006444717291742563, + 0.011641664430499077, + 0.0127982497215271, + 0.012052159756422043, + 0.02201756462454796, + 0.01946849934756756, + 0.012652909383177757, + 0.012067864648997784, + 0.010813203640282154, + 0.022154735401272774, + 0.009928619489073753, + 0.018718421459197998, + 0.02239236980676651, + 0.0115476930513978, + 0.030248522758483887, + 0.015084375627338886, + 0.05277051776647568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.02638125978410244 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029809776693582535 + ], + "opposite_logprob_tokens": [ + -3.5298097133636475 + ], + "logprob_reward": -0.029809776693582535, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029809776693582535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10732584446668625, + 0.07204589247703552, + 0.05840912088751793, + 0.08438710868358612, + 0.06002596765756607, + 0.2008034884929657, + 0.03766684979200363, + 0.03640763461589813, + 0.21631696820259094, + 0.12661112844944 + ], + "credit_map": [ + 0.012726900167763233, + 0.021323392167687416, + 0.007617405615746975, + 0.015487144701182842, + 0.01162523590028286, + 0.016380611807107925, + 0.00790140125900507, + 0.014263751916587353, + 0.006467576138675213, + 0.006544687785208225, + 0.004134188871830702, + 0.008158187381923199, + 0.0054710241965949535, + 0.014820175245404243, + 0.008601503446698189, + 0.017848554998636246, + 0.005570138804614544, + 0.010781343095004559, + 0.007347613573074341, + 0.010026111267507076, + 0.004365810193121433, + 0.008136768825352192, + 0.0062681883573532104, + 0.005913146771490574, + 0.005086215678602457, + 0.011656669899821281, + 0.005802674684673548, + 0.010670064948499203, + 0.009108837693929672, + 0.01609349064528942, + 0.0080898841843009, + 0.017879273742437363, + 0.00429649418219924, + 0.010082571767270565, + 0.006355996709316969, + 0.007790111005306244, + 0.007486782968044281, + 0.007451871410012245, + 0.00794934295117855, + 0.008612798526883125, + 0.012274221517145634, + 0.029333893209695816, + 0.030295664444565773, + 0.01659465953707695, + 0.027220996096730232, + 0.024444274604320526, + 0.023989854380488396, + 0.0366499200463295, + 0.0032839523628354073, + 0.004331153817474842, + 0.0055326735600829124, + 0.005354363936930895, + 0.005604603793472052, + 0.004732846282422543, + 0.004752406384795904, + 0.004074851982295513, + 0.004386915825307369, + 0.0038265350740402937, + 0.004017812665551901, + 0.004067845642566681, + 0.007684273179620504, + 0.0038051526062190533, + 0.005587433930486441, + 0.0030316689517349005, + 0.02443504147231579, + 0.02354566566646099, + 0.02520407736301422, + 0.04121318459510803, + 0.0370306596159935, + 0.03129676356911659, + 0.014894824475049973, + 0.01869676634669304, + 0.019267410039901733, + 0.026729198172688484, + 0.009256339631974697, + 0.015929928049445152, + 0.011629894375801086, + 0.012754839845001698, + 0.021458612754940987, + 0.00958490651100874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.029809776693582535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588263869285583 + ], + "opposite_logprob_tokens": [ + -3.8955881595611572 + ], + "logprob_reward": -0.020588263869285583, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588263869285583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08070754259824753, + 0.0712655633687973, + 0.06497108936309814, + 0.04642082005739212, + 0.11257940530776978, + 0.06588657200336456, + 0.06407717615365982, + 0.07355009764432907, + 0.37247517704963684, + 0.04806657135486603 + ], + "credit_map": [ + 0.008949127979576588, + 0.012371229939162731, + 0.0052214739844202995, + 0.010840470902621746, + 0.011254683136940002, + 0.012679153122007847, + 0.0074665723368525505, + 0.011924829334020615, + 0.00996431801468134, + 0.008316653780639172, + 0.006769966334104538, + 0.007260019425302744, + 0.007727402728050947, + 0.011416113004088402, + 0.006841215770691633, + 0.01296987198293209, + 0.004614413715898991, + 0.008759984746575356, + 0.004111535381525755, + 0.00874683354049921, + 0.006436733528971672, + 0.009580171667039394, + 0.01062079332768917, + 0.012100625783205032, + 0.00445895828306675, + 0.006456006318330765, + 0.005920169409364462, + 0.005090033169835806, + 0.004157188348472118, + 0.008166052401065826, + 0.003626493038609624, + 0.008545917458832264, + 0.007350299507379532, + 0.014385394752025604, + 0.010569470003247261, + 0.011029873043298721, + 0.0240761861205101, + 0.017619390040636063, + 0.014862471260130405, + 0.012686328031122684, + 0.0038704751059412956, + 0.005880582612007856, + 0.012032843194901943, + 0.005253451410681009, + 0.0073404195718467236, + 0.010819466784596443, + 0.006828842218965292, + 0.013860493898391724, + 0.006088435649871826, + 0.008225164376199245, + 0.00889537949115038, + 0.010090166702866554, + 0.007280203513801098, + 0.00902192760258913, + 0.0077457730658352375, + 0.006730122957378626, + 0.010499541647732258, + 0.012080001644790173, + 0.00967531930655241, + 0.006467997096478939, + 0.010700426995754242, + 0.00971702765673399, + 0.008092914707958698, + 0.006316873710602522, + 0.04355154186487198, + 0.037922944873571396, + 0.04706468805670738, + 0.08526629209518433, + 0.03156290575861931, + 0.06308995932340622, + 0.034004662185907364, + 0.030012156814336777, + 0.005801589228212833, + 0.00691318791359663, + 0.004299236927181482, + 0.006398008204996586, + 0.005896132439374924, + 0.005660741589963436, + 0.00763274822384119, + 0.005464924033731222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.020588263869285583 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199413359165192 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.014199413359165192, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199413359165192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10170309245586395, + 0.07601963728666306, + 0.05058887600898743, + 0.03746558725833893, + 0.14813850820064545, + 0.04766588285565376, + 0.03753188997507095, + 0.05773008614778519, + 0.2895433008670807, + 0.15361319482326508 + ], + "credit_map": [ + 0.011424433439970016, + 0.01590600796043873, + 0.005256826989352703, + 0.008939271792769432, + 0.015593011863529682, + 0.015667855739593506, + 0.012189192697405815, + 0.01672648824751377, + 0.007877576164901257, + 0.005998065695166588, + 0.005345619283616543, + 0.006097578909248114, + 0.006467102095484734, + 0.015421486459672451, + 0.008140616118907928, + 0.020671594887971878, + 0.004196590278297663, + 0.006778676528483629, + 0.004316598642617464, + 0.006793263368308544, + 0.007009393069893122, + 0.007226034067571163, + 0.008235854096710682, + 0.0060324654914438725, + 0.0027956657577306032, + 0.005364986136555672, + 0.0042175715789198875, + 0.003791800467297435, + 0.005527763161808252, + 0.00569329084828496, + 0.0037113050930202007, + 0.0063632032833993435, + 0.010786137543618679, + 0.01778409630060196, + 0.015051654540002346, + 0.019430937245488167, + 0.02236487902700901, + 0.025147389620542526, + 0.01907907985150814, + 0.01849433220922947, + 0.003189243609085679, + 0.007777833379805088, + 0.006647362839430571, + 0.005517918150871992, + 0.005790799856185913, + 0.006579318083822727, + 0.004631347954273224, + 0.007532054092735052, + 0.00470192264765501, + 0.004185905680060387, + 0.005563430022448301, + 0.005166850984096527, + 0.003342854091897607, + 0.004890670999884605, + 0.004333517514169216, + 0.005346739199012518, + 0.009037711657583714, + 0.0071716755628585815, + 0.007631690241396427, + 0.007055858615785837, + 0.0070606404915452, + 0.007387797348201275, + 0.007139519322663546, + 0.0052451929077506065, + 0.03570397570729256, + 0.03300391137599945, + 0.03384961932897568, + 0.053329598158597946, + 0.042455773800611496, + 0.04301026463508606, + 0.02149815484881401, + 0.026692014187574387, + 0.0207852553576231, + 0.018692519515752792, + 0.013166532851755619, + 0.020545056089758873, + 0.009423503652215004, + 0.021902412176132202, + 0.015617407858371735, + 0.03348051384091377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.014199413359165192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot captures a towering, floor-to-ceiling bookshelf constructed of warm, honey-toned pine, lined with books in a sun-drenched log cabin. A silent, magical river of literature begins to cascade down from the upper shelves; hardcovers with frayed fabric spines and glossy paperbacks in a variety of colors tumble and slide against each other in not a chaotic collision but a fluid, flowing movement. The camera slowly tilts downward, carefully following the books as they fall onto the polished wooden floor, forming a winding stream that meanders through the room. Bright, golden afternoon sunlight streams in through a large window, illuminating the scene and capturing countless motes of dust dancing in the air. As the books drift across the floor, the pages of some open volumes flutter gently and deliberately, the thin paper catching the light, creating a serene and enchanting atmosphere of quiet, impossible magic within the peaceful, sun-drenched sanctuary.", + "seed": 9190, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_6/index_8167_seed_9190_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.030095115303993225, + "vlm_logprob": -0.030095115303993225, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.030095115303993225, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02491888776421547 + ], + "opposite_logprob_tokens": [ + -4.024919033050537 + ], + "logprob_reward": -0.02491888776421547, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02491888776421547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1801677793264389, + 0.058154452592134476, + 0.03158144652843475, + 0.15481260418891907, + 0.021768389269709587, + 0.0672551766037941, + 0.09329169988632202, + 0.2523094117641449, + 0.09444165229797363, + 0.04621734097599983 + ], + "credit_map": [ + 0.017335187643766403, + 0.01940106227993965, + 0.010020683519542217, + 0.014363041147589684, + 0.020003441721200943, + 0.03183406963944435, + 0.015888812020421028, + 0.05132151395082474, + 0.008082099258899689, + 0.008391117677092552, + 0.004354686010628939, + 0.005080322735011578, + 0.005927965976297855, + 0.007418822031468153, + 0.006261306349188089, + 0.012638140469789505, + 0.0023338382598012686, + 0.0028607798740267754, + 0.004481761250644922, + 0.006377527955919504, + 0.0026868090499192476, + 0.0033436515368521214, + 0.0055490173399448395, + 0.003948066383600235, + 0.008830000646412373, + 0.00750599754974246, + 0.00890218373388052, + 0.009566428139805794, + 0.005724382121115923, + 0.05065692961215973, + 0.006831308826804161, + 0.056795381009578705, + 0.0015790568431839347, + 0.002331107622012496, + 0.002544066170230508, + 0.002308581955730915, + 0.0036198548041284084, + 0.0031736260280013084, + 0.003002856858074665, + 0.0032092405017465353, + 0.004803014453500509, + 0.005017216317355633, + 0.015512976795434952, + 0.00882112979888916, + 0.0064486730843782425, + 0.009792249649763107, + 0.009277484379708767, + 0.007582442834973335, + 0.004771724343299866, + 0.0035099575761705637, + 0.0076322853565216064, + 0.004566533491015434, + 0.005875447299331427, + 0.0396796278655529, + 0.005325070582330227, + 0.021931063383817673, + 0.027436858043074608, + 0.012166064232587814, + 0.07472312450408936, + 0.01581253670156002, + 0.01321613509207964, + 0.025064637884497643, + 0.015458761714398861, + 0.06843134015798569, + 0.010588702745735645, + 0.012290661223232746, + 0.014606572687625885, + 0.008734539151191711, + 0.011530210264027119, + 0.014083308167755604, + 0.012748095206916332, + 0.009859575890004635, + 0.005469599738717079, + 0.006819250527769327, + 0.006352706346660852, + 0.0050390129908919334, + 0.005741196684539318, + 0.005871785804629326, + 0.0058670127764344215, + 0.0050567821599543095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.02491888776421547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043105874210596085 + ], + "opposite_logprob_tokens": [ + -3.1681058406829834 + ], + "logprob_reward": -0.043105874210596085, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043105874210596085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1979122906923294, + 0.1117241159081459, + 0.03386969119310379, + 0.11155357211828232, + 0.022954266518354416, + 0.08207441866397858, + 0.25835344195365906, + 0.09121517837047577, + 0.043202195316553116, + 0.04714084416627884 + ], + "credit_map": [ + 0.019339513033628464, + 0.020637046545743942, + 0.01406659372150898, + 0.023365581408143044, + 0.019272079691290855, + 0.03153996914625168, + 0.019044574350118637, + 0.05064692720770836, + 0.015597580000758171, + 0.011883717961609364, + 0.008408849127590656, + 0.009500631131231785, + 0.008721177466213703, + 0.014036997221410275, + 0.017390182241797447, + 0.02618497796356678, + 0.0025615044869482517, + 0.004466209560632706, + 0.004030227195471525, + 0.005675794556736946, + 0.0028852031100541353, + 0.004386199172586203, + 0.0049184346571564674, + 0.004946117289364338, + 0.005908020306378603, + 0.005831690970808268, + 0.006435221992433071, + 0.00639320770278573, + 0.005718854255974293, + 0.03537449240684509, + 0.006372322328388691, + 0.03951975330710411, + 0.0018684384413063526, + 0.0032381510827690363, + 0.0020414243917912245, + 0.0028739047702401876, + 0.005584579426795244, + 0.003347633173689246, + 0.0020646946504712105, + 0.0019354377873241901, + 0.0070980992168188095, + 0.008950075134634972, + 0.014722790569067001, + 0.008961508050560951, + 0.00788737740367651, + 0.013132311403751373, + 0.010947521775960922, + 0.010374733246862888, + 0.01312499400228262, + 0.01008691918104887, + 0.020815353840589523, + 0.012136886827647686, + 0.0153976334258914, + 0.11061977595090866, + 0.014497614465653896, + 0.061674244701862335, + 0.01036873459815979, + 0.006573258433490992, + 0.016438331454992294, + 0.007994910702109337, + 0.0052114431746304035, + 0.012995959259569645, + 0.008500789292156696, + 0.023131737485527992, + 0.00424666004255414, + 0.005830348469316959, + 0.006629700306802988, + 0.0039106253534555435, + 0.006652624811977148, + 0.006118397228419781, + 0.004822597838938236, + 0.004991237539798021, + 0.004999167751520872, + 0.004980160854756832, + 0.009583747014403343, + 0.005114695057272911, + 0.003768372582271695, + 0.007933571934700012, + 0.005746185313910246, + 0.0050149401649832726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.043105874210596085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307344928383827 + ], + "opposite_logprob_tokens": [ + -3.1680734157562256 + ], + "logprob_reward": -0.04307344928383827, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307344928383827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07511630654335022, + 0.07829426229000092, + 0.04727402329444885, + 0.06040530651807785, + 0.01827085204422474, + 0.15865178406238556, + 0.3383699953556061, + 0.09841509908437729, + 0.05389253422617912, + 0.07130984216928482 + ], + "credit_map": [ + 0.0059348600916564465, + 0.009596298448741436, + 0.007679782807826996, + 0.016613125801086426, + 0.0066454727202653885, + 0.008764448575675488, + 0.008635139092803001, + 0.011247177608311176, + 0.007614171598106623, + 0.006077952217310667, + 0.00875244103372097, + 0.007664266042411327, + 0.00955980084836483, + 0.01131790317595005, + 0.012384030967950821, + 0.014923697337508202, + 0.003121579997241497, + 0.005007103085517883, + 0.00353249441832304, + 0.005361571442335844, + 0.0031920941546559334, + 0.008034992031753063, + 0.006931054871529341, + 0.012093129567801952, + 0.003736260114237666, + 0.004814228508621454, + 0.004281582776457071, + 0.006161060184240341, + 0.005539929494261742, + 0.014922990463674068, + 0.004229667130857706, + 0.01671958900988102, + 0.00184425781480968, + 0.002002380322664976, + 0.002760449191555381, + 0.0018226426327601075, + 0.0033651713747531176, + 0.002619164064526558, + 0.0016850928077474236, + 0.002171693602576852, + 0.01382060069590807, + 0.01426085364073515, + 0.020291607826948166, + 0.02325754426419735, + 0.015689846128225327, + 0.017805732786655426, + 0.03223487734794617, + 0.0212907325476408, + 0.01655205525457859, + 0.012174675241112709, + 0.026037612929940224, + 0.018970128148794174, + 0.020176634192466736, + 0.1399858593940735, + 0.021430082619190216, + 0.08304296433925629, + 0.009082278236746788, + 0.006273719947785139, + 0.02506914921104908, + 0.012051794677972794, + 0.006814275868237019, + 0.010423953644931316, + 0.009687851183116436, + 0.019012080505490303, + 0.0053226943127810955, + 0.007030951324850321, + 0.007008951157331467, + 0.00592419970780611, + 0.0070158448070287704, + 0.007097854744642973, + 0.00760939996689558, + 0.0068826391361653805, + 0.006719358265399933, + 0.009478487074375153, + 0.012510689906775951, + 0.010837060399353504, + 0.006906768772751093, + 0.007108490448445082, + 0.010581930167973042, + 0.007167048752307892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.04307344928383827 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023298323154449463 + ], + "opposite_logprob_tokens": [ + -3.7732982635498047 + ], + "logprob_reward": -0.023298323154449463, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023298323154449463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039866428822278976, + 0.10920745879411697, + 0.03571955859661102, + 0.10113571584224701, + 0.023027518764138222, + 0.04386770352721214, + 0.4161321818828583, + 0.09873649477958679, + 0.09500644356012344, + 0.03730042278766632 + ], + "credit_map": [ + 0.0046136826276779175, + 0.006982189137488604, + 0.002911027055233717, + 0.006415591109544039, + 0.0040682293474674225, + 0.005255057010799646, + 0.004013034049421549, + 0.005607617553323507, + 0.008492824621498585, + 0.011479721404612064, + 0.012449678033590317, + 0.012034491635859013, + 0.016209939494729042, + 0.015260526910424232, + 0.016297198832035065, + 0.01698308438062668, + 0.0037033462431281805, + 0.005252373404800892, + 0.004937971942126751, + 0.004821447655558586, + 0.003214742988348007, + 0.0054167527705430984, + 0.004778564907610416, + 0.003594357054680586, + 0.0057343789376318455, + 0.00484874751418829, + 0.006347562186419964, + 0.004702228587120771, + 0.0034714003559201956, + 0.033790647983551025, + 0.004269907251000404, + 0.03797084838151932, + 0.0027709307614713907, + 0.002094583585858345, + 0.004304832313209772, + 0.002504668664187193, + 0.003935385495424271, + 0.003090599784627557, + 0.002354156691581011, + 0.0019723617006093264, + 0.004469819366931915, + 0.005198232363909483, + 0.005637778900563717, + 0.005102258175611496, + 0.004726368933916092, + 0.006987334229052067, + 0.006096514407545328, + 0.005649396684020758, + 0.019844448193907738, + 0.015324774198234081, + 0.03285471349954605, + 0.019595718011260033, + 0.02038467675447464, + 0.18171195685863495, + 0.023569488897919655, + 0.1028464287519455, + 0.009755305014550686, + 0.007529369555413723, + 0.027497321367263794, + 0.009472710080444813, + 0.006149217486381531, + 0.014560352079570293, + 0.010361654683947563, + 0.013410565443336964, + 0.008704411797225475, + 0.012554425746202469, + 0.012832014821469784, + 0.008807129226624966, + 0.012016212567687035, + 0.01068229228258133, + 0.01229120884090662, + 0.017118752002716064, + 0.003498754696920514, + 0.004201832693070173, + 0.006565331947058439, + 0.0047714728862047195, + 0.0044722468592226505, + 0.0051668379455804825, + 0.0046519008465111256, + 0.003972043748944998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 3 + ], + "selected_reward": -0.023298323154449463 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-drenched log cabin featuring a floor-to-ceiling bookshelf made of warm, honey-toned pine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do hardcover books with frayed fabric spines and glossy paperbooks in various colors cascade down from the upper shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling books move in a fluid, flowing manner rather than a chaotic collision?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle static shot and then slowly tilt downward to follow the books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do some open volumes on the floor have pages that flutter gently and deliberately while catching the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079042106866837 + ], + "opposite_logprob_tokens": [ + -4.141078948974609 + ], + "logprob_reward": -0.016079042106866837, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079042106866837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25312539935112, + 0.04481988772749901, + 0.059506841003894806, + 0.08959890902042389, + 0.024356525391340256, + 0.03835510462522507, + 0.04289922118186951, + 0.2138231247663498, + 0.05245847627520561, + 0.18105646967887878 + ], + "credit_map": [ + 0.025226091966032982, + 0.02354486659169197, + 0.016909053549170494, + 0.02093411609530449, + 0.025439435616135597, + 0.04582769423723221, + 0.022309206426143646, + 0.07293494790792465, + 0.003152307588607073, + 0.005430598743259907, + 0.004139846656471491, + 0.005894884001463652, + 0.0065162936225533485, + 0.005305921658873558, + 0.007978799752891064, + 0.006401242688298225, + 0.003938197158277035, + 0.008422886952757835, + 0.008968876674771309, + 0.0056179785169661045, + 0.005350390449166298, + 0.008108364418148994, + 0.006806965451687574, + 0.012293187901377678, + 0.005813649855554104, + 0.007268297486007214, + 0.005268908105790615, + 0.0063782683573663235, + 0.008152482099831104, + 0.02265198715031147, + 0.008498032577335835, + 0.025567293167114258, + 0.00295647862367332, + 0.0029734796844422817, + 0.0025254301726818085, + 0.003725481918081641, + 0.003547373227775097, + 0.0032072626054286957, + 0.0032518322113901377, + 0.0021691895090043545, + 0.0031802605371922255, + 0.004502583295106888, + 0.004733640234917402, + 0.003588475752621889, + 0.006170539651066065, + 0.006942447740584612, + 0.004831467755138874, + 0.004405695013701916, + 0.0043097687885165215, + 0.004226980730891228, + 0.005597887560725212, + 0.003738602390512824, + 0.004281292203813791, + 0.009856730699539185, + 0.005062580574303865, + 0.005825383123010397, + 0.02052406780421734, + 0.012839882634580135, + 0.05220121890306473, + 0.02254832535982132, + 0.014086729846894741, + 0.025720851495862007, + 0.019659018144011497, + 0.04624302685260773, + 0.005628958344459534, + 0.007107769604772329, + 0.007075959350913763, + 0.004418494645506144, + 0.006613202393054962, + 0.008018635213375092, + 0.007422999013215303, + 0.00617246562615037, + 0.017887325957417488, + 0.02034919708967209, + 0.030535802245140076, + 0.02132706344127655, + 0.021497802808880806, + 0.03358570486307144, + 0.020994052290916443, + 0.01487953495234251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.016079042106866837 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_7/index_tensor([604, 604])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_7/index_tensor([604, 604])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aeeaafba6cc784484e940f7513d6c7bfbfa53830 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_7/index_tensor([604, 604])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera begins with a medium shot, slowly and steadily tilting downward, gradually focusing from the top of a surreal space constructed of delicate pencil crosshatch lines and sketched shading to the center of the scene. In this monochrome world, like a Leonardo da Vinci manuscript brought to life, a clockwork knight crafted entirely of polished brass sits at a workbench, also outlined in pencil lines. The knight's body is composed of countless precisely meshing gears, levers, and springs. Under a single, bright light source emanating from one side of the frame, the brass surface reflects a warm and sharp luster, creating a stark contrast against the black and white background. With a mechanical, focused gesture, it methodically disassembles its left forearm. Its right hand, mechanically grasping a tiny screwdriver, delicately loosens a screw securing the outer casing, its movements steady and deliberate. As the camera continues to descend and zoom in, the picture turns into an extreme close-up. We can clearly see the inside of the disassembled robotic arm, complex copper wires wrapped around tiny springs, several gears of different sizes neatly placed on the workbench, and the threads of each screw and the spiral of the spring are clearly visible. The whole scene is filled with an atmosphere of extreme tranquility, concentration and classical mechanical aesthetics.", + "seed": 3408, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_7/index_604_seed_3408_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.23619208224117755, + "vlm_logprob": -0.23619208224117755, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.23619208224117755, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0563666820526123 + ], + "opposite_logprob_tokens": [ + -0.4313667118549347 + ], + "logprob_reward": -1.0563666820526123, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0563666820526123, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11937607824802399, + 0.02159181237220764, + 0.01874551922082901, + 0.07302674651145935, + 0.02763952501118183, + 0.07289358973503113, + 0.027194933965802193, + 0.08357949554920197, + 0.40423572063446045, + 0.15171660482883453 + ], + "credit_map": [ + 0.020502084866166115, + 0.01594882644712925, + 0.012785548344254494, + 0.01881914772093296, + 0.01355243194848299, + 0.008960834704339504, + 0.013051855377852917, + 0.015755344182252884, + 0.0023579373955726624, + 0.004166884813457727, + 0.0025795428082346916, + 0.002307005925104022, + 0.002051594899967313, + 0.0026911436580121517, + 0.002517193555831909, + 0.0029205074533820152, + 0.0022658256348222494, + 0.0026868574786931276, + 0.0015674160094931722, + 0.002478142036125064, + 0.0017861839151009917, + 0.003102888585999608, + 0.0020305218640714884, + 0.0028276839293539524, + 0.005823368672281504, + 0.0053731827065348625, + 0.004721655044704676, + 0.007503855973482132, + 0.02190924808382988, + 0.011209186166524887, + 0.005973820574581623, + 0.010512434877455235, + 0.0039281235076487064, + 0.004287454765290022, + 0.0026868307031691074, + 0.0032561267726123333, + 0.0032508287113159895, + 0.004341545049101114, + 0.0026986123993992805, + 0.00319000449962914, + 0.012119757942855358, + 0.004923972766846418, + 0.0043923454359173775, + 0.006340536288917065, + 0.009350365959107876, + 0.012827327474951744, + 0.012368620373308659, + 0.010570666752755642, + 0.0030851431656628847, + 0.002641804749146104, + 0.0030101723968982697, + 0.003203333355486393, + 0.002903630258515477, + 0.0044729518704116344, + 0.004370357375591993, + 0.0035075400955975056, + 0.008480818942189217, + 0.006861007306724787, + 0.00866816658526659, + 0.010522043332457542, + 0.012171094305813313, + 0.014191652648150921, + 0.013219941407442093, + 0.00946477148681879, + 0.039018746465444565, + 0.03391330689191818, + 0.03411862999200821, + 0.05102358013391495, + 0.07355018705129623, + 0.08984466642141342, + 0.046160612255334854, + 0.036605965346097946, + 0.018012678250670433, + 0.01359124481678009, + 0.01802849769592285, + 0.017085254192352295, + 0.023615548387169838, + 0.025876222178339958, + 0.020811259746551514, + 0.014695893973112106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -1.0563666820526123 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052642300724983215 + ], + "opposite_logprob_tokens": [ + -5.2552642822265625 + ], + "logprob_reward": -0.0052642300724983215, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052642300724983215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045637473464012146, + 0.029230833053588867, + 0.029991887509822845, + 0.04511616379022598, + 0.024604931473731995, + 0.08167590200901031, + 0.0552983433008194, + 0.10614112764596939, + 0.4682280421257019, + 0.11407532542943954 + ], + "credit_map": [ + 0.00684161763638258, + 0.004915730562061071, + 0.00508565129712224, + 0.006842837203294039, + 0.005627669859677553, + 0.005358065478503704, + 0.005897073540836573, + 0.005068825092166662, + 0.0029253747779875994, + 0.0034866887144744396, + 0.0020623370073735714, + 0.002779331523925066, + 0.005041245371103287, + 0.005035469774156809, + 0.0034268375020474195, + 0.004473548382520676, + 0.003512095892801881, + 0.0043553682044148445, + 0.0036360048688948154, + 0.002625448862090707, + 0.002493813633918762, + 0.003637281246483326, + 0.005831246729940176, + 0.00390062527731061, + 0.005249600391834974, + 0.004571468569338322, + 0.0038906466215848923, + 0.0046249316073954105, + 0.006094792392104864, + 0.008688926696777344, + 0.00489936675876379, + 0.007096428889781237, + 0.0024218151811510324, + 0.002951467176899314, + 0.0021684879902750254, + 0.0021929878275841475, + 0.0029316043946892023, + 0.005089370999485254, + 0.003351791761815548, + 0.003497404046356678, + 0.017650574445724487, + 0.006688171532005072, + 0.00646724971011281, + 0.007497488986700773, + 0.012929641641676426, + 0.012971706688404083, + 0.009625806473195553, + 0.0078452592715621, + 0.006004410330206156, + 0.008869154378771782, + 0.006310607306659222, + 0.007015980314463377, + 0.007218655664473772, + 0.007678547874093056, + 0.006719793193042278, + 0.005481190979480743, + 0.00995714869350195, + 0.008287299424409866, + 0.014805467799305916, + 0.009314610622823238, + 0.022991782054305077, + 0.019387051463127136, + 0.011615404859185219, + 0.009782347828149796, + 0.041520871222019196, + 0.03694121167063713, + 0.03849026933312416, + 0.056944191455841064, + 0.08602365851402283, + 0.10551771521568298, + 0.05885535851120949, + 0.043934740126132965, + 0.014381296932697296, + 0.011816774494946003, + 0.013075397349894047, + 0.013987780548632145, + 0.015120269730687141, + 0.015378578566014767, + 0.019044535234570503, + 0.011270683258771896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.0052642300724983215 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866204410791397 + ], + "opposite_logprob_tokens": [ + -3.048661947250366 + ], + "logprob_reward": -0.04866204410791397, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866204410791397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10385613888502121, + 0.046821728348731995, + 0.10046888142824173, + 0.07590389996767044, + 0.05024163797497749, + 0.1619700938463211, + 0.2150154560804367, + 0.03615542873740196, + 0.14272578060626984, + 0.06684091687202454 + ], + "credit_map": [ + 0.01643856056034565, + 0.011843493208289146, + 0.013358755968511105, + 0.010716121643781662, + 0.01391318254172802, + 0.011111604049801826, + 0.014263840392231941, + 0.012210574932396412, + 0.004387118387967348, + 0.007979726418852806, + 0.006297824438661337, + 0.00615084171295166, + 0.004072968848049641, + 0.00694146566092968, + 0.005736168008297682, + 0.005255614407360554, + 0.013682570308446884, + 0.019689355045557022, + 0.011794396676123142, + 0.009809963405132294, + 0.008102304302155972, + 0.012806919403374195, + 0.010139396414160728, + 0.014443979598581791, + 0.007785385474562645, + 0.008628616109490395, + 0.00810129102319479, + 0.01455732062458992, + 0.00742042250931263, + 0.011969360522925854, + 0.007634137757122517, + 0.009807366877794266, + 0.004725215956568718, + 0.00517130084335804, + 0.0042338864877820015, + 0.00574973551556468, + 0.007049508858472109, + 0.011599710211157799, + 0.0052941772155463696, + 0.006418103352189064, + 0.018572835251688957, + 0.016940509900450706, + 0.011436241678893566, + 0.01758917048573494, + 0.01567300781607628, + 0.02222623862326145, + 0.02878650277853012, + 0.030745593830943108, + 0.019558683037757874, + 0.025404103100299835, + 0.023895347490906715, + 0.020390478894114494, + 0.035028256475925446, + 0.03175986185669899, + 0.03343512490391731, + 0.025543587282299995, + 0.00424552196636796, + 0.005069348961114883, + 0.004968527238816023, + 0.0040339878760278225, + 0.005267065949738026, + 0.0056895664893090725, + 0.004252778831869364, + 0.002628634450957179, + 0.012517252005636692, + 0.011244364082813263, + 0.012924677692353725, + 0.017513692378997803, + 0.02499415911734104, + 0.02933199144899845, + 0.020880360156297684, + 0.013319283723831177, + 0.008423742838203907, + 0.008886308409273624, + 0.0075826141983270645, + 0.008205259218811989, + 0.010998363606631756, + 0.00943449791520834, + 0.00711838249117136, + 0.0061917477287352085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.04866204410791397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620303601026535 + ], + "opposite_logprob_tokens": [ + -2.81203031539917 + ], + "logprob_reward": -0.0620303601026535, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620303601026535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13627587258815765, + 0.08804520219564438, + 0.0828620195388794, + 0.07206602394580841, + 0.06810237467288971, + 0.14586924016475677, + 0.07545924186706543, + 0.07870149612426758, + 0.13715174794197083, + 0.11546670645475388 + ], + "credit_map": [ + 0.01724468357861042, + 0.01735556311905384, + 0.02079368382692337, + 0.014742540195584297, + 0.01892934925854206, + 0.008390231058001518, + 0.022874314337968826, + 0.015945514664053917, + 0.013266037218272686, + 0.018639937043190002, + 0.007734806276857853, + 0.009844356216490269, + 0.010787133127450943, + 0.010231951251626015, + 0.007950306870043278, + 0.009590677917003632, + 0.009641658514738083, + 0.014762377366423607, + 0.0070917923003435135, + 0.008736278861761093, + 0.008185542188584805, + 0.014453145675361156, + 0.010355507954955101, + 0.009635714814066887, + 0.006589306518435478, + 0.0068275504745543, + 0.007091091945767403, + 0.009621960110962391, + 0.010243771597743034, + 0.014975202269852161, + 0.007763307075947523, + 0.008953840471804142, + 0.008105945773422718, + 0.006927003152668476, + 0.006893799174576998, + 0.010206811130046844, + 0.00984896533191204, + 0.010183295235037804, + 0.00820167362689972, + 0.007734884042292833, + 0.021168921142816544, + 0.01514091994613409, + 0.010887430980801582, + 0.014956733211874962, + 0.013988647609949112, + 0.0214364193379879, + 0.028232792392373085, + 0.020057378336787224, + 0.006919363047927618, + 0.009166284464299679, + 0.007426634896546602, + 0.010186185128986835, + 0.010487137362360954, + 0.011996951885521412, + 0.01310835499316454, + 0.006168333813548088, + 0.007093984168022871, + 0.0077715786173939705, + 0.009061240591108799, + 0.01009560376405716, + 0.011684700846672058, + 0.01290726475417614, + 0.008047784678637981, + 0.012039342895150185, + 0.02339189499616623, + 0.014483978040516376, + 0.013591107912361622, + 0.011515817604959011, + 0.021919937804341316, + 0.02672247216105461, + 0.01130806002765894, + 0.014218483120203018, + 0.0067741479724645615, + 0.007259184028953314, + 0.00916228350251913, + 0.021477317437529564, + 0.020670784637331963, + 0.02843524143099785, + 0.013563261367380619, + 0.008124490268528461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.0620303601026535 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637094870209694 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008637094870209694, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637094870209694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06217173859477043, + 0.0292014479637146, + 0.042217861860990524, + 0.0574510432779789, + 0.031625401228666306, + 0.16647548973560333, + 0.11791973561048508, + 0.098745197057724, + 0.3211420178413391, + 0.0730501040816307 + ], + "credit_map": [ + 0.007464575115591288, + 0.008586879819631577, + 0.00749982288107276, + 0.010124658234417439, + 0.007122793234884739, + 0.005625502206385136, + 0.00909015815705061, + 0.006657350808382034, + 0.002617935184389353, + 0.004295355174690485, + 0.0026890633162111044, + 0.0024728160351514816, + 0.004663195926696062, + 0.005411889869719744, + 0.003440828761085868, + 0.0036103639286011457, + 0.0039698276668787, + 0.004639058839529753, + 0.00379234179854393, + 0.0059144869446754456, + 0.004750065505504608, + 0.007780563551932573, + 0.006179829593747854, + 0.00519168796017766, + 0.006209040060639381, + 0.006980300415307283, + 0.005651148501783609, + 0.005037510767579079, + 0.00785965658724308, + 0.00727031659334898, + 0.009294386021792889, + 0.009148684330284595, + 0.003320275340229273, + 0.003487425623461604, + 0.004318807739764452, + 0.002976333722472191, + 0.0030485084280371666, + 0.00571672897785902, + 0.004487453028559685, + 0.00426986999809742, + 0.025848250836133957, + 0.018172921612858772, + 0.013719629496335983, + 0.015149475075304508, + 0.02555832825601101, + 0.02346334420144558, + 0.023295216262340546, + 0.02126830630004406, + 0.011794486083090305, + 0.01697254553437233, + 0.010399127379059792, + 0.015041632577776909, + 0.014711440540850163, + 0.01711079105734825, + 0.01917894557118416, + 0.012710757553577423, + 0.008453466929495335, + 0.0077205137349665165, + 0.010031551122665405, + 0.013625224120914936, + 0.014776003547012806, + 0.0174405574798584, + 0.015302661806344986, + 0.01139521412551403, + 0.027698000892996788, + 0.02540019527077675, + 0.027233196422457695, + 0.0402587428689003, + 0.05725366994738579, + 0.06905652582645416, + 0.04574437066912651, + 0.028497319668531418, + 0.007628527469933033, + 0.008280676789581776, + 0.005973337218165398, + 0.007638364564627409, + 0.011067344807088375, + 0.010287412442266941, + 0.013086282648146152, + 0.00908815860748291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.008637094870209694 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a medium shot, slowly and steadily tilting downward, gradually focusing from the top of a surreal space constructed of delicate pencil crosshatch lines and sketched shading to the center of the scene. In this monochrome world, like a Leonardo da Vinci manuscript brought to life, a clockwork knight crafted entirely of polished brass sits at a workbench, also outlined in pencil lines. The knight's body is composed of countless precisely meshing gears, levers, and springs. Under a single, bright light source emanating from one side of the frame, the brass surface reflects a warm and sharp luster, creating a stark contrast against the black and white background. With a mechanical, focused gesture, it methodically disassembles its left forearm. Its right hand, mechanically grasping a tiny screwdriver, delicately loosens a screw securing the outer casing, its movements steady and deliberate. As the camera continues to descend and zoom in, the picture turns into an extreme close-up. We can clearly see the inside of the disassembled robotic arm, complex copper wires wrapped around tiny springs, several gears of different sizes neatly placed on the workbench, and the threads of each screw and the spiral of the spring are clearly visible. The whole scene is filled with an atmosphere of extreme tranquility, concentration and classical mechanical aesthetics.", + "seed": 3409, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_7/index_604_seed_3409_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.20912647638469933, + "vlm_logprob": -0.20912647638469933, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.20912647638469933, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8999871015548706 + ], + "opposite_logprob_tokens": [ + -0.5249871015548706 + ], + "logprob_reward": -0.8999871015548706, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8999871015548706, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.5537164807319641, + 0.05201496556401253, + 0.054538793861866, + 0.048844654113054276, + 0.024556446820497513, + 0.14693117141723633, + 0.020445967093110085, + 0.029957571998238564, + 0.03763941302895546, + 0.03135454282164574 + ], + "credit_map": [ + 0.12294542789459229, + 0.04993828013539314, + 0.07066891342401505, + 0.026581931859254837, + 0.02861909195780754, + 0.19672128558158875, + 0.029873114079236984, + 0.02836839109659195, + 0.007654021494090557, + 0.006597165483981371, + 0.006744562182575464, + 0.005244454834610224, + 0.007426016498357058, + 0.005184028297662735, + 0.008442919701337814, + 0.004721795208752155, + 0.007266867905855179, + 0.00478406110778451, + 0.006883896421641111, + 0.005838707089424133, + 0.00524872774258256, + 0.008445912972092628, + 0.008723178878426552, + 0.0073474375531077385, + 0.007787639740854502, + 0.00586538715288043, + 0.005658301059156656, + 0.0037043443880975246, + 0.008729856461286545, + 0.0054920632392168045, + 0.005282481200993061, + 0.0063245780766010284, + 0.004312447737902403, + 0.0020033572800457478, + 0.003087094286456704, + 0.003701902227476239, + 0.0037491153925657272, + 0.0019518296467140317, + 0.0024148535449057817, + 0.0033358491491526365, + 0.017561234533786774, + 0.019857799634337425, + 0.018697131425142288, + 0.012112096883356571, + 0.0216831024736166, + 0.013237757608294487, + 0.021550804376602173, + 0.022231239825487137, + 0.00253328587859869, + 0.001828901469707489, + 0.002616981277242303, + 0.003469472983852029, + 0.002367394044995308, + 0.0024971936363726854, + 0.0026175901293754578, + 0.0025151476729661226, + 0.0031110786367207766, + 0.0025439898017793894, + 0.0057737682946026325, + 0.004217430483549833, + 0.0040898676961660385, + 0.00470847450196743, + 0.002308640396222472, + 0.00320432148873806, + 0.004869903903454542, + 0.00450683431699872, + 0.004829510115087032, + 0.0046633146703243256, + 0.0034681528341025114, + 0.003803438041359186, + 0.0065738302655518055, + 0.004924427252262831, + 0.003694219281896949, + 0.00431846221908927, + 0.004260886460542679, + 0.004664807114750147, + 0.004277938045561314, + 0.002351891715079546, + 0.004401038866490126, + 0.003385299351066351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.8999871015548706 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076382951810956 + ], + "opposite_logprob_tokens": [ + -4.882638454437256 + ], + "logprob_reward": -0.0076382951810956, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076382951810956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3333503305912018, + 0.08416878432035446, + 0.06416187435388565, + 0.07836762070655823, + 0.07875283062458038, + 0.11691286414861679, + 0.0444464236497879, + 0.0727275162935257, + 0.060930781066417694, + 0.0661810040473938 + ], + "credit_map": [ + 0.10671540349721909, + 0.02392810769379139, + 0.06311959773302078, + 0.01474468968808651, + 0.026471268385648727, + 0.06096208095550537, + 0.01878955215215683, + 0.01861961930990219, + 0.01298607513308525, + 0.01085993368178606, + 0.013650448992848396, + 0.007137617561966181, + 0.010355517268180847, + 0.011979960836470127, + 0.009923038072884083, + 0.00727619556710124, + 0.008370399475097656, + 0.005897976458072662, + 0.00971945933997631, + 0.006277532782405615, + 0.009369174018502235, + 0.00869616400450468, + 0.009488177485764027, + 0.006342992652207613, + 0.007632959634065628, + 0.011578785255551338, + 0.011073686182498932, + 0.011691910214722157, + 0.009532827883958817, + 0.009251032955944538, + 0.010140949860215187, + 0.007465470116585493, + 0.010942868888378143, + 0.009720460511744022, + 0.007522268686443567, + 0.014958684332668781, + 0.0074548921547830105, + 0.0062942118383944035, + 0.011019258759915829, + 0.010840185917913914, + 0.01486146915704012, + 0.013993903994560242, + 0.014501823112368584, + 0.01432550698518753, + 0.015030108392238617, + 0.010198053903877735, + 0.019263630732893944, + 0.014738362282514572, + 0.005682833027094603, + 0.0037218371871858835, + 0.006473564077168703, + 0.005989145487546921, + 0.007636272348463535, + 0.004531452897936106, + 0.0061607942916452885, + 0.0042505268938839436, + 0.00938380416482687, + 0.006318063009530306, + 0.010263510048389435, + 0.011964165605604649, + 0.008854673244059086, + 0.009193722158670425, + 0.006724331993609667, + 0.01002524234354496, + 0.008203498087823391, + 0.006731516681611538, + 0.010807987302541733, + 0.00789322704076767, + 0.0058874692767858505, + 0.008549069054424763, + 0.008285986259579659, + 0.004572027828544378, + 0.0070144664496183395, + 0.011387044563889503, + 0.008127069100737572, + 0.00941831897944212, + 0.008252580650150776, + 0.005713499151170254, + 0.007367715239524841, + 0.008900311775505543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0076382951810956 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700254887342453 + ], + "opposite_logprob_tokens": [ + -2.6950254440307617 + ], + "logprob_reward": -0.0700254887342453, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700254887342453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2079838663339615, + 0.048179980367422104, + 0.05397713929414749, + 0.08697720617055893, + 0.1771378368139267, + 0.22469286620616913, + 0.03197850286960602, + 0.0626692846417427, + 0.06613502651453018, + 0.04026821255683899 + ], + "credit_map": [ + 0.05086307227611542, + 0.019570577889680862, + 0.021113529801368713, + 0.01534199621528387, + 0.025522613897919655, + 0.042026013135910034, + 0.020041795447468758, + 0.013504265807569027, + 0.008017662912607193, + 0.005711749196052551, + 0.008040107786655426, + 0.0034979223273694515, + 0.00639718770980835, + 0.007531305775046349, + 0.004927865229547024, + 0.004056178964674473, + 0.009916062466800213, + 0.00477381469681859, + 0.007598011754453182, + 0.004611575044691563, + 0.007020632736384869, + 0.009471047669649124, + 0.006385346408933401, + 0.00420064851641655, + 0.014846297912299633, + 0.009302384220063686, + 0.00995582900941372, + 0.008550022728741169, + 0.010959855280816555, + 0.008587237447500229, + 0.010922078974545002, + 0.013853499665856361, + 0.025318624451756477, + 0.017443744465708733, + 0.015720663592219353, + 0.042594850063323975, + 0.01090192049741745, + 0.010059110820293427, + 0.024357574060559273, + 0.030741335824131966, + 0.03248615190386772, + 0.022277096286416054, + 0.025331899523735046, + 0.03976532816886902, + 0.029470300301909447, + 0.02060391940176487, + 0.03177459165453911, + 0.02298359014093876, + 0.004281564150005579, + 0.002596353879198432, + 0.0037107134703546762, + 0.002867372240871191, + 0.006867436692118645, + 0.0039576259441673756, + 0.004313955549150705, + 0.00338348257355392, + 0.008524790406227112, + 0.008165503852069378, + 0.008218593895435333, + 0.011040867306292057, + 0.006606972776353359, + 0.007213654927909374, + 0.0058411420322954655, + 0.007057758513838053, + 0.00822562724351883, + 0.008314617909491062, + 0.007956734858453274, + 0.005997119937092066, + 0.006357713136821985, + 0.0077718282118439674, + 0.01276577077805996, + 0.008745619095861912, + 0.004186266101896763, + 0.005007537081837654, + 0.007299636956304312, + 0.006922201719135046, + 0.004284876398742199, + 0.0030614130664616823, + 0.005632060579955578, + 0.0038742218166589737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0700254887342453 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203528866171837 + ], + "opposite_logprob_tokens": [ + -2.812035322189331 + ], + "logprob_reward": -0.06203528866171837, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203528866171837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10500594973564148, + 0.07265505939722061, + 0.09379784762859344, + 0.10554800927639008, + 0.06195889785885811, + 0.16111183166503906, + 0.06144791841506958, + 0.03963769972324371, + 0.24773789942264557, + 0.051098864525556564 + ], + "credit_map": [ + 0.013225690461695194, + 0.014299899339675903, + 0.015903659164905548, + 0.009612933732569218, + 0.0148242749273777, + 0.016302261501550674, + 0.012991133145987988, + 0.007846095599234104, + 0.013132119551301003, + 0.00940337497740984, + 0.009022354148328304, + 0.008095421828329563, + 0.009995704516768456, + 0.007807593327015638, + 0.010142644867300987, + 0.005055845715105534, + 0.0139952776953578, + 0.008447220548987389, + 0.01718776673078537, + 0.007818714715540409, + 0.00818849541246891, + 0.018519334495067596, + 0.010117143392562866, + 0.00952389370650053, + 0.015883514657616615, + 0.012444679625332355, + 0.01732487417757511, + 0.011631714180111885, + 0.015733128413558006, + 0.010310005396604538, + 0.01030650082975626, + 0.011913594789803028, + 0.00799589604139328, + 0.010378994978964329, + 0.004942574072629213, + 0.009909505024552345, + 0.007766537833958864, + 0.005998715292662382, + 0.006437321659177542, + 0.008529353886842728, + 0.02185157686471939, + 0.012629336677491665, + 0.015214474871754646, + 0.01469852402806282, + 0.03168260306119919, + 0.015914931893348694, + 0.011003882624208927, + 0.03811649605631828, + 0.008017275482416153, + 0.0052647371776402, + 0.009156938642263412, + 0.005756721366196871, + 0.010072476230561733, + 0.009052951820194721, + 0.0073128435760736465, + 0.006813977845013142, + 0.0048269606195390224, + 0.003600698197260499, + 0.005675320979207754, + 0.003761364845559001, + 0.006149883382022381, + 0.005652609281241894, + 0.005240595433861017, + 0.0047302693128585815, + 0.03021945431828499, + 0.03790070116519928, + 0.03157060965895653, + 0.020584721118211746, + 0.025094956159591675, + 0.03526272997260094, + 0.040162306278944016, + 0.02694242261350155, + 0.005595778115093708, + 0.010326432064175606, + 0.007155515719205141, + 0.006685705855488777, + 0.0037847142666578293, + 0.004419028293341398, + 0.005448836833238602, + 0.007682854309678078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.06203528866171837 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946207791566849 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005946207791566849, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946207791566849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20496104657649994, + 0.03899887204170227, + 0.03249572589993477, + 0.060967471450567245, + 0.04697174206376076, + 0.18163171410560608, + 0.021584102883934975, + 0.10600662231445312, + 0.10268270969390869, + 0.20369994640350342 + ], + "credit_map": [ + 0.03992552310228348, + 0.021107202395796776, + 0.022166648879647255, + 0.017924949526786804, + 0.02422306314110756, + 0.0389593131840229, + 0.02365117147564888, + 0.017003178596496582, + 0.0060541690327227116, + 0.004662069492042065, + 0.006627816241234541, + 0.003179246326908469, + 0.0033503274898976088, + 0.007475219666957855, + 0.004164042416960001, + 0.003485980676487088, + 0.00512376893311739, + 0.0029666360933333635, + 0.006271401420235634, + 0.0019505125237628818, + 0.004257409833371639, + 0.0034571094438433647, + 0.004307933617383242, + 0.004160952288657427, + 0.00863559264689684, + 0.006564473267644644, + 0.008194123394787312, + 0.008147479966282845, + 0.007281421683728695, + 0.006528808269649744, + 0.007754168938845396, + 0.007861403748393059, + 0.010532693937420845, + 0.0045282007195055485, + 0.005876604467630386, + 0.004182294476777315, + 0.005575492512434721, + 0.0034769412595778704, + 0.005786275491118431, + 0.007013241294771433, + 0.02326224185526371, + 0.01575441285967827, + 0.017976362258195877, + 0.019278034567832947, + 0.028311006724834442, + 0.01599740795791149, + 0.02542751096189022, + 0.03562474250793457, + 0.0026163794100284576, + 0.0018585638608783484, + 0.0020829045679420233, + 0.003705183509737253, + 0.003203572938218713, + 0.0023122974671423435, + 0.0026855168398469687, + 0.0031196847558021545, + 0.009819061495363712, + 0.008170755580067635, + 0.014383494853973389, + 0.010577232576906681, + 0.01707286946475506, + 0.012578406371176243, + 0.013810309581458569, + 0.019594497978687286, + 0.011536003090441227, + 0.013124498538672924, + 0.01135179866105318, + 0.01036424096673727, + 0.01058874186128378, + 0.012994186021387577, + 0.01896614395081997, + 0.013757089152932167, + 0.02229994721710682, + 0.05270445719361305, + 0.030404748395085335, + 0.023581871762871742, + 0.016192596405744553, + 0.011235463432967663, + 0.021179091185331345, + 0.026101769879460335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.005946207791566849 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a medium shot, slowly and steadily tilting downward, gradually focusing from the top of a surreal space constructed of delicate pencil crosshatch lines and sketched shading to the center of the scene. In this monochrome world, like a Leonardo da Vinci manuscript brought to life, a clockwork knight crafted entirely of polished brass sits at a workbench, also outlined in pencil lines. The knight's body is composed of countless precisely meshing gears, levers, and springs. Under a single, bright light source emanating from one side of the frame, the brass surface reflects a warm and sharp luster, creating a stark contrast against the black and white background. With a mechanical, focused gesture, it methodically disassembles its left forearm. Its right hand, mechanically grasping a tiny screwdriver, delicately loosens a screw securing the outer casing, its movements steady and deliberate. As the camera continues to descend and zoom in, the picture turns into an extreme close-up. We can clearly see the inside of the disassembled robotic arm, complex copper wires wrapped around tiny springs, several gears of different sizes neatly placed on the workbench, and the threads of each screw and the spiral of the spring are clearly visible. The whole scene is filled with an atmosphere of extreme tranquility, concentration and classical mechanical aesthetics.", + "seed": 3410, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_7/index_604_seed_3410_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08585504731163382, + "vlm_logprob": -0.08585504731163382, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.08585504731163382, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28559404611587524 + ], + "opposite_logprob_tokens": [ + -1.4105939865112305 + ], + "logprob_reward": -0.28559404611587524, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28559404611587524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36393141746520996, + 0.05309159308671951, + 0.03256073594093323, + 0.05542883276939392, + 0.03464198857545853, + 0.046154871582984924, + 0.0546819306910038, + 0.24105998873710632, + 0.08004990965127945, + 0.03839872032403946 + ], + "credit_map": [ + 0.07525860518217087, + 0.027773162350058556, + 0.07968591153621674, + 0.03506692498922348, + 0.027984824031591415, + 0.0423746332526207, + 0.03710174933075905, + 0.03868560865521431, + 0.010653280653059483, + 0.0057554952800273895, + 0.013388183899223804, + 0.004428625572472811, + 0.004939961712807417, + 0.005003941245377064, + 0.003955420106649399, + 0.004966687876731157, + 0.005965580232441425, + 0.0034567825496196747, + 0.004173212684690952, + 0.0031895730644464493, + 0.004127327352762222, + 0.00508473115041852, + 0.003879155032336712, + 0.0026843734085559845, + 0.005553786642849445, + 0.005930434446781874, + 0.005525852087885141, + 0.004456492606550455, + 0.006193280685693026, + 0.011275150813162327, + 0.007020213641226292, + 0.009473620913922787, + 0.003968851640820503, + 0.004669192247092724, + 0.0030437721870839596, + 0.003979840315878391, + 0.004669341258704662, + 0.005130966659635305, + 0.00465051457285881, + 0.004529512487351894, + 0.0072358944453299046, + 0.004763751290738583, + 0.0033170178066939116, + 0.003838065080344677, + 0.0056624035350978374, + 0.005654251668602228, + 0.0050270347855985165, + 0.010656454600393772, + 0.0046875737607479095, + 0.0056743137538433075, + 0.007627790793776512, + 0.007206357084214687, + 0.009016181342303753, + 0.005896748509258032, + 0.008246361277997494, + 0.0063266027718782425, + 0.028841543942689896, + 0.01662280596792698, + 0.03139504790306091, + 0.04753824695944786, + 0.019291339442133904, + 0.023803317919373512, + 0.04216093569993973, + 0.03140674903988838, + 0.015521809458732605, + 0.0030039604753255844, + 0.007774702273309231, + 0.005719766486436129, + 0.006232315208762884, + 0.017583100125193596, + 0.011139439418911934, + 0.013074814341962337, + 0.0034457650035619736, + 0.003529930952936411, + 0.005058296024799347, + 0.0036624399945139885, + 0.005377332679927349, + 0.004228444769978523, + 0.006399689707905054, + 0.006696823053061962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.28559404611587524 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748385261744261 + ], + "opposite_logprob_tokens": [ + -5.006748199462891 + ], + "logprob_reward": -0.006748385261744261, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748385261744261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12445250898599625, + 0.12077561020851135, + 0.046840619295835495, + 0.14048801362514496, + 0.05937911570072174, + 0.05339166894555092, + 0.07271862030029297, + 0.20977084338665009, + 0.07009171694517136, + 0.10209131985902786 + ], + "credit_map": [ + 0.013616785407066345, + 0.01673213392496109, + 0.016225280240178108, + 0.013485895469784737, + 0.01821352168917656, + 0.016277354210615158, + 0.01343545876443386, + 0.016466092318296432, + 0.01017423439770937, + 0.019681766629219055, + 0.008606272749602795, + 0.02056192047894001, + 0.021553227677941322, + 0.014920712448656559, + 0.014236398041248322, + 0.01104107964783907, + 0.008161680772900581, + 0.005286270286887884, + 0.004761453252285719, + 0.004965058993548155, + 0.007463542744517326, + 0.006861812900751829, + 0.004659587051719427, + 0.004681214690208435, + 0.02145414613187313, + 0.01910221017897129, + 0.012275585904717445, + 0.011338607408106327, + 0.025549180805683136, + 0.015572134405374527, + 0.021617276594042778, + 0.013578853569924831, + 0.006571002770215273, + 0.004433554131537676, + 0.005777452606707811, + 0.008285903371870518, + 0.006594575475901365, + 0.007285878527909517, + 0.013121028430759907, + 0.007309720851480961, + 0.00501606147736311, + 0.005517153535038233, + 0.0051599969156086445, + 0.009783352725207806, + 0.008557732217013836, + 0.006290137302130461, + 0.005308846011757851, + 0.007758387830108404, + 0.00645869504660368, + 0.006728104781359434, + 0.010240876115858555, + 0.009742168709635735, + 0.010273839347064495, + 0.011423069052398205, + 0.008077404461801052, + 0.009774463251233101, + 0.024925358593463898, + 0.014690196141600609, + 0.02969527244567871, + 0.03510052338242531, + 0.017070608213543892, + 0.02585490420460701, + 0.033100567758083344, + 0.029333412647247314, + 0.010900067165493965, + 0.0044466229155659676, + 0.008349427953362465, + 0.005585661623626947, + 0.0060428534634411335, + 0.013854349963366985, + 0.009522981010377407, + 0.011389750987291336, + 0.012222583405673504, + 0.009248060174286366, + 0.018316546455025673, + 0.009205622598528862, + 0.012191515415906906, + 0.012501421384513378, + 0.013719780370593071, + 0.014685794711112976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.006748385261744261 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295939236879349 + ], + "opposite_logprob_tokens": [ + -2.237959384918213 + ], + "logprob_reward": -0.11295939236879349, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295939236879349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1485479176044464, + 0.04848780483007431, + 0.03787020966410637, + 0.19954365491867065, + 0.13606783747673035, + 0.07679358124732971, + 0.11720162630081177, + 0.14272210001945496, + 0.04841947183012962, + 0.04434574767947197 + ], + "credit_map": [ + 0.014415507204830647, + 0.014345784671604633, + 0.01960768550634384, + 0.01929629221558571, + 0.016561519354581833, + 0.019806362688541412, + 0.018148887902498245, + 0.02636587619781494, + 0.0073864939622581005, + 0.006383924745023251, + 0.009037008509039879, + 0.005742909386754036, + 0.005440313834697008, + 0.005580703727900982, + 0.004619063343852758, + 0.004297388251870871, + 0.004301413893699646, + 0.003601460950449109, + 0.00337044196203351, + 0.004178755450993776, + 0.006616315338760614, + 0.006606888957321644, + 0.005381379276514053, + 0.003813553601503372, + 0.05765970051288605, + 0.02198476903140545, + 0.0189268309623003, + 0.011921574361622334, + 0.02894102782011032, + 0.024546487256884575, + 0.01731070689857006, + 0.01825256273150444, + 0.014862031675875187, + 0.012278950773179531, + 0.011286135762929916, + 0.017372354865074158, + 0.016318419948220253, + 0.01585310697555542, + 0.027234163135290146, + 0.020862674340605736, + 0.006388409528881311, + 0.007135821972042322, + 0.00599523913115263, + 0.018193894997239113, + 0.010680804960429668, + 0.009886086918413639, + 0.007999681867659092, + 0.010513639077544212, + 0.013615586794912815, + 0.013037953525781631, + 0.015729472041130066, + 0.015800360590219498, + 0.021328924223780632, + 0.015299797058105469, + 0.01194869726896286, + 0.010440838523209095, + 0.01565505936741829, + 0.009284060448408127, + 0.01918402872979641, + 0.0256357304751873, + 0.011227192357182503, + 0.01721358858048916, + 0.024462485685944557, + 0.02005995251238346, + 0.00958932377398014, + 0.0028325545135885477, + 0.006928319111466408, + 0.003469750750809908, + 0.005406549666076899, + 0.011829584836959839, + 0.004103841260075569, + 0.004259546287357807, + 0.004563366062939167, + 0.004181758034974337, + 0.008025161921977997, + 0.004154391586780548, + 0.0050003561191260815, + 0.006020909640938044, + 0.006844756193459034, + 0.0055550467222929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.11295939236879349 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978058110922575 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.00978058110922575, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978058110922575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15753339231014252, + 0.06175529956817627, + 0.05973996967077255, + 0.11138506233692169, + 0.10946864634752274, + 0.16886097192764282, + 0.08196775615215302, + 0.0714370384812355, + 0.14935481548309326, + 0.028497060760855675 + ], + "credit_map": [ + 0.015626560896635056, + 0.0174819715321064, + 0.02193046361207962, + 0.020686980336904526, + 0.018152063712477684, + 0.020561596378684044, + 0.019367456436157227, + 0.02372630126774311, + 0.011924371123313904, + 0.008484176360070705, + 0.006451557856053114, + 0.006001407280564308, + 0.008981854654848576, + 0.0063184406608343124, + 0.00804261863231659, + 0.005550875328481197, + 0.0042861457914114, + 0.0071114711463451385, + 0.005403268616646528, + 0.008103686384856701, + 0.010688375681638718, + 0.008741379715502262, + 0.009113354608416557, + 0.006292287260293961, + 0.019421175122261047, + 0.011865827254951, + 0.011236131191253662, + 0.00855268444865942, + 0.01910267397761345, + 0.013158624991774559, + 0.015361906029284, + 0.012686047703027725, + 0.010163696482777596, + 0.011800098232924938, + 0.00875923689454794, + 0.01660751923918724, + 0.01373258139938116, + 0.01343522034585476, + 0.020412899553775787, + 0.014557396061718464, + 0.026241738349199295, + 0.017353681847453117, + 0.010969855822622776, + 0.021897947415709496, + 0.022204037755727768, + 0.01374428253620863, + 0.017493456602096558, + 0.03895598649978638, + 0.009132076986134052, + 0.008818740956485271, + 0.009159881621599197, + 0.008119620382785797, + 0.011637317016720772, + 0.011993090622127056, + 0.010236451402306557, + 0.012870576232671738, + 0.007280366960912943, + 0.004944831132888794, + 0.010593552142381668, + 0.013829508796334267, + 0.005735511891543865, + 0.007024793885648251, + 0.012442833743989468, + 0.009585640393197536, + 0.01986832544207573, + 0.009417431429028511, + 0.02439017780125141, + 0.02122425101697445, + 0.012760377489030361, + 0.023185983300209045, + 0.020215146243572235, + 0.018293114379048347, + 0.002483166055753827, + 0.00233570858836174, + 0.0036130936350673437, + 0.0027773010078817606, + 0.004166515078395605, + 0.004342769738286734, + 0.005056384950876236, + 0.0037221231032162905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.00978058110922575 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192831702530384 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192831702530384, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192831702530384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08621489256620407, + 0.06574791669845581, + 0.0332450233399868, + 0.10822910815477371, + 0.06044657900929451, + 0.07808932662010193, + 0.06546204537153244, + 0.3217797577381134, + 0.11582555621862411, + 0.06495976448059082 + ], + "credit_map": [ + 0.010331672616302967, + 0.008663935586810112, + 0.01251349039375782, + 0.009137414395809174, + 0.010939252562820911, + 0.011065308004617691, + 0.010346105322241783, + 0.013217717409133911, + 0.009179259650409222, + 0.009238501079380512, + 0.010211695916950703, + 0.009093618020415306, + 0.008280659094452858, + 0.007168044336140156, + 0.006781408563256264, + 0.005794730503112078, + 0.0057203397154808044, + 0.0035331447143107653, + 0.0034042943734675646, + 0.004992389585822821, + 0.004006687086075544, + 0.005386165343225002, + 0.0031866924837231636, + 0.003015312133356929, + 0.024861052632331848, + 0.014745688997209072, + 0.008799200877547264, + 0.007484246976673603, + 0.02221386507153511, + 0.01082665380090475, + 0.009816976264119148, + 0.009481435641646385, + 0.005855572875589132, + 0.006709336303174496, + 0.004478498362004757, + 0.007449334487318993, + 0.009399482049047947, + 0.006934040226042271, + 0.01027553528547287, + 0.009344784542918205, + 0.006955144461244345, + 0.00728481262922287, + 0.006406589411199093, + 0.016744336113333702, + 0.00942239724099636, + 0.008078515529632568, + 0.010062038898468018, + 0.013135495595633984, + 0.005502783227711916, + 0.005582367070019245, + 0.009267226792871952, + 0.01056094840168953, + 0.006039133761078119, + 0.013231641612946987, + 0.00814112089574337, + 0.007136831525713205, + 0.03843235969543457, + 0.022983577102422714, + 0.0475381538271904, + 0.05738326534628868, + 0.026254909113049507, + 0.04105526953935623, + 0.047829754650592804, + 0.04030248895287514, + 0.012431048788130283, + 0.007060748990625143, + 0.016348406672477722, + 0.010933661833405495, + 0.008396915160119534, + 0.013175844214856625, + 0.02159172296524048, + 0.025887222960591316, + 0.006148329470306635, + 0.0064742714166641235, + 0.010873685590922832, + 0.005937405861914158, + 0.009823351167142391, + 0.008679796941578388, + 0.008843057788908482, + 0.008179871365427971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.014192831702530384 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a medium shot, slowly and steadily tilting downward, gradually focusing from the top of a surreal space constructed of delicate pencil crosshatch lines and sketched shading to the center of the scene. In this monochrome world, like a Leonardo da Vinci manuscript brought to life, a clockwork knight crafted entirely of polished brass sits at a workbench, also outlined in pencil lines. The knight's body is composed of countless precisely meshing gears, levers, and springs. Under a single, bright light source emanating from one side of the frame, the brass surface reflects a warm and sharp luster, creating a stark contrast against the black and white background. With a mechanical, focused gesture, it methodically disassembles its left forearm. Its right hand, mechanically grasping a tiny screwdriver, delicately loosens a screw securing the outer casing, its movements steady and deliberate. As the camera continues to descend and zoom in, the picture turns into an extreme close-up. We can clearly see the inside of the disassembled robotic arm, complex copper wires wrapped around tiny springs, several gears of different sizes neatly placed on the workbench, and the threads of each screw and the spiral of the spring are clearly visible. The whole scene is filled with an atmosphere of extreme tranquility, concentration and classical mechanical aesthetics.", + "seed": 3411, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_7/index_604_seed_3411_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2067251607775688, + "vlm_logprob": -0.2067251607775688, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.2067251607775688, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9768635034561157 + ], + "opposite_logprob_tokens": [ + -0.4768635034561157 + ], + "logprob_reward": -0.9768635034561157, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9768635034561157, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18831995129585266, + 0.08642896264791489, + 0.03247843682765961, + 0.03710602968931198, + 0.14535899460315704, + 0.0672183707356453, + 0.05261332541704178, + 0.24513588845729828, + 0.04707940295338631, + 0.09826058149337769 + ], + "credit_map": [ + 0.036038778722286224, + 0.021710170432925224, + 0.019497882574796677, + 0.0304432213306427, + 0.01870453916490078, + 0.01904098317027092, + 0.019561152905225754, + 0.023323221132159233, + 0.01251272764056921, + 0.010025435127317905, + 0.011733903549611568, + 0.011831438168883324, + 0.01191762089729309, + 0.007571098860353231, + 0.009656183421611786, + 0.011180559173226357, + 0.006382792256772518, + 0.004270096309483051, + 0.002768290229141712, + 0.004664119798690081, + 0.0035197946708649397, + 0.003508408088237047, + 0.003073133295401931, + 0.0042918045073747635, + 0.005813308525830507, + 0.00535851763561368, + 0.004620232619345188, + 0.004866677336394787, + 0.004580455832183361, + 0.0032632953952997923, + 0.003660737071186304, + 0.004942806903272867, + 0.04436253011226654, + 0.014077321626245975, + 0.020259611308574677, + 0.01068672351539135, + 0.01134799886494875, + 0.014440570026636124, + 0.011232374235987663, + 0.018951868638396263, + 0.011131655424833298, + 0.008575980551540852, + 0.00910473708063364, + 0.008446873165667057, + 0.006197392009198666, + 0.006023057270795107, + 0.008849230594933033, + 0.008889446035027504, + 0.005644191987812519, + 0.009689598344266415, + 0.006240345537662506, + 0.005193473305553198, + 0.00476507842540741, + 0.00442172447219491, + 0.008529221639037132, + 0.00812968797981739, + 0.015134243294596672, + 0.013877472840249538, + 0.020490730181336403, + 0.012713853269815445, + 0.008470939472317696, + 0.010920570231974125, + 0.08490829914808273, + 0.07861980050802231, + 0.006205399986356497, + 0.005267304368317127, + 0.0052376724779605865, + 0.006188889034092426, + 0.004775368142873049, + 0.006650158669799566, + 0.006356824189424515, + 0.006397784221917391, + 0.01848740689456463, + 0.007737432140856981, + 0.014171216636896133, + 0.011623530648648739, + 0.011656032875180244, + 0.012037012726068497, + 0.010527870617806911, + 0.012020080350339413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.9768635034561157 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646931774914265 + ], + "opposite_logprob_tokens": [ + -5.3796467781066895 + ], + "logprob_reward": -0.004646931774914265, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646931774914265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16714513301849365, + 0.09378816187381744, + 0.045159995555877686, + 0.032968927174806595, + 0.06853906810283661, + 0.06461504846811295, + 0.13156278431415558, + 0.0488286167383194, + 0.13574543595314026, + 0.21164679527282715 + ], + "credit_map": [ + 0.03295254707336426, + 0.027142835780978203, + 0.02551925927400589, + 0.020747335627675056, + 0.015635155141353607, + 0.016342345625162125, + 0.01535779982805252, + 0.013447838835418224, + 0.01661769300699234, + 0.010604946874082088, + 0.013134938664734364, + 0.01377701386809349, + 0.011952090077102184, + 0.008073592558503151, + 0.01068845670670271, + 0.008939431048929691, + 0.00849111471325159, + 0.006171476561576128, + 0.004658087622374296, + 0.006734626833349466, + 0.004546631593257189, + 0.003971306141465902, + 0.00444289343431592, + 0.006143859587609768, + 0.006206472869962454, + 0.0039354427717626095, + 0.005744150839745998, + 0.003288687439635396, + 0.003643634030595422, + 0.003023711033165455, + 0.0027108979411423206, + 0.004415930714458227, + 0.014517511241137981, + 0.008551894687116146, + 0.0094036515802145, + 0.009343456476926804, + 0.006817331537604332, + 0.0059346300549805164, + 0.006629207171499729, + 0.007341383025050163, + 0.006848102435469627, + 0.009172419086098671, + 0.008867155760526657, + 0.006813523359596729, + 0.006681552156805992, + 0.005711388308554888, + 0.009143002331256866, + 0.011377902701497078, + 0.01163035724312067, + 0.011009281501173973, + 0.03620080649852753, + 0.012255510315299034, + 0.006529583130031824, + 0.009142668917775154, + 0.01656748354434967, + 0.02822709083557129, + 0.005374709144234657, + 0.007616665214300156, + 0.006472107023000717, + 0.0040811812505126, + 0.0033888956531882286, + 0.0033584304619580507, + 0.010813510045409203, + 0.007723117247223854, + 0.013061282224953175, + 0.013759179040789604, + 0.02514038048684597, + 0.016917528584599495, + 0.010975093580782413, + 0.013747503980994225, + 0.01797552779316902, + 0.02416893094778061, + 0.023439820855855942, + 0.017694072797894478, + 0.03267865628004074, + 0.026540599763393402, + 0.034698277711868286, + 0.0231112539768219, + 0.027791118249297142, + 0.025692999362945557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.004646931774914265 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370029106736183 + ], + "opposite_logprob_tokens": [ + -3.4087002277374268 + ], + "logprob_reward": -0.03370029106736183, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370029106736183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12753646075725555, + 0.049432653933763504, + 0.03223979473114014, + 0.05435420572757721, + 0.14381422102451324, + 0.09605380147695541, + 0.04773230105638504, + 0.1611756980419159, + 0.1450149267911911, + 0.14264601469039917 + ], + "credit_map": [ + 0.02120331861078739, + 0.020077882334589958, + 0.01372563187032938, + 0.015308387577533722, + 0.015742547810077667, + 0.017077937722206116, + 0.012674088589847088, + 0.011726672761142254, + 0.009581553749740124, + 0.006384527776390314, + 0.005212891846895218, + 0.005886688828468323, + 0.006740394979715347, + 0.004391580820083618, + 0.005540757440030575, + 0.005694263149052858, + 0.006348830182105303, + 0.0037053206469863653, + 0.0036470373161137104, + 0.004289740230888128, + 0.004000755958259106, + 0.002982108388096094, + 0.0035003828816115856, + 0.0037656177300959826, + 0.006138262804597616, + 0.010877884924411774, + 0.005177535582333803, + 0.009339631535112858, + 0.006472896318882704, + 0.004711620509624481, + 0.0059133777394890785, + 0.005722991656512022, + 0.03421821817755699, + 0.018353106454014778, + 0.019917134195566177, + 0.019618649035692215, + 0.013381553813815117, + 0.013239312916994095, + 0.013790851458907127, + 0.01129539031535387, + 0.008546003140509129, + 0.009476905688643456, + 0.00863613747060299, + 0.008164162747561932, + 0.013240624219179153, + 0.009116454981267452, + 0.015695372596383095, + 0.02317814901471138, + 0.004050953313708305, + 0.006834638305008411, + 0.006566199008375406, + 0.0059981816448271275, + 0.004515453707426786, + 0.004973211325705051, + 0.008041700348258018, + 0.0067519620060920715, + 0.009903056547045708, + 0.00991741381585598, + 0.012552152387797832, + 0.012389830313622952, + 0.009385189972817898, + 0.008404922671616077, + 0.05070137977600098, + 0.04792174696922302, + 0.01222397480159998, + 0.015433424152433872, + 0.030256761237978935, + 0.016220849007368088, + 0.010618673637509346, + 0.011710877530276775, + 0.012103235349059105, + 0.03644712641835213, + 0.013809950090944767, + 0.014632152393460274, + 0.016865478828549385, + 0.021325642243027687, + 0.02309957891702652, + 0.016258351504802704, + 0.014921522699296474, + 0.021733347326517105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.03370029106736183 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779046289622784 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009779046289622784, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779046289622784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05287552252411842, + 0.026790998876094818, + 0.022684108465909958, + 0.04453013464808464, + 0.04190250113606453, + 0.037279825657606125, + 0.05255070701241493, + 0.1514599323272705, + 0.07377927005290985, + 0.4961470067501068 + ], + "credit_map": [ + 0.009174035862088203, + 0.009906640276312828, + 0.005036929156631231, + 0.0077724214643239975, + 0.005597053561359644, + 0.005106345284730196, + 0.005651003215461969, + 0.0046310885809361935, + 0.003812892595306039, + 0.0036732845474034548, + 0.002624128945171833, + 0.004301436711102724, + 0.0037258497904986143, + 0.002818632870912552, + 0.0028359100688248873, + 0.002998863346874714, + 0.0034422320313751698, + 0.002980548655614257, + 0.0018875149544328451, + 0.0030829026363790035, + 0.003025064244866371, + 0.0023907741997390985, + 0.0031640585511922836, + 0.0027110138908028603, + 0.005583567079156637, + 0.007377952802926302, + 0.006223093252629042, + 0.0055993469431996346, + 0.006052996497601271, + 0.003760184859856963, + 0.00482889823615551, + 0.005104098003357649, + 0.00948392041027546, + 0.005060700234025717, + 0.005637580994516611, + 0.004158218391239643, + 0.0053721205331385136, + 0.003408604534342885, + 0.004241509363055229, + 0.004539846442639828, + 0.004979098215699196, + 0.004822966642677784, + 0.007178259082138538, + 0.004668078850954771, + 0.003159320680424571, + 0.0035254580434411764, + 0.0044506690464913845, + 0.004495974630117416, + 0.004173974506556988, + 0.00385644449852407, + 0.014806510880589485, + 0.005226875655353069, + 0.003532855771481991, + 0.004075623117387295, + 0.006414953153580427, + 0.010463470593094826, + 0.00967992004007101, + 0.009488793089985847, + 0.013944562524557114, + 0.010062899440526962, + 0.006959397811442614, + 0.009169901721179485, + 0.048009615391492844, + 0.044144853949546814, + 0.007815956138074398, + 0.008406020700931549, + 0.01008335780352354, + 0.009569610469043255, + 0.008167666383087635, + 0.006790058221668005, + 0.012043061666190624, + 0.010903540998697281, + 0.07441886514425278, + 0.04209213703870773, + 0.06987851858139038, + 0.06030179560184479, + 0.07016568630933762, + 0.05611521005630493, + 0.06215338781476021, + 0.061021383851766586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.009779046289622784 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in monochrome with pencil crosshatch lines and sketched shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a clockwork knight crafted of polished brass sitting at a workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight methodically disassembling its left forearm using a tiny screwdriver?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a medium shot and slowly tilt downward while zooming in to an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the internal components of the disassembled arm, including copper wires, springs, and gears, clearly visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636031299829483 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.008636031299829483, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636031299829483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09460274875164032, + 0.06600288301706314, + 0.03196943923830986, + 0.02223358489573002, + 0.059890054166316986, + 0.057579874992370605, + 0.15968771278858185, + 0.10215332359075546, + 0.16875801980495453, + 0.2371223270893097 + ], + "credit_map": [ + 0.01799124851822853, + 0.012957638129591942, + 0.009756221435964108, + 0.01365337148308754, + 0.010379806160926819, + 0.010890532284975052, + 0.008771130815148354, + 0.010202803649008274, + 0.01073178369551897, + 0.009209620766341686, + 0.007509015500545502, + 0.007940161041915417, + 0.007546840235590935, + 0.006315252278000116, + 0.008467848412692547, + 0.00828236062079668, + 0.004505799151957035, + 0.003868214786052704, + 0.0028106786776334047, + 0.0035442980006337166, + 0.0037698494270443916, + 0.003829667344689369, + 0.00416370015591383, + 0.005477230064570904, + 0.0023790423292666674, + 0.00368242128752172, + 0.0022803049068897963, + 0.0024654162116348743, + 0.0034053493291139603, + 0.0023597993422299623, + 0.0020996651146560907, + 0.0035615882370620966, + 0.012491593137383461, + 0.006611688062548637, + 0.008037333376705647, + 0.00705646863207221, + 0.005451787728816271, + 0.0049643246456980705, + 0.006154913455247879, + 0.009121947921812534, + 0.00582076283171773, + 0.006145237013697624, + 0.005653027910739183, + 0.005623818375170231, + 0.006139808334410191, + 0.005328959319740534, + 0.009372559376060963, + 0.013495702296495438, + 0.012809227220714092, + 0.020414669066667557, + 0.03502247855067253, + 0.017955468967556953, + 0.009780260734260082, + 0.010977262631058693, + 0.026143362745642662, + 0.026584971696138382, + 0.007021785248070955, + 0.006145243998616934, + 0.010956862941384315, + 0.00676511786878109, + 0.0047900318168103695, + 0.003942674957215786, + 0.03210284933447838, + 0.030428754165768623, + 0.014053070917725563, + 0.013986879028379917, + 0.020687445998191833, + 0.03673630580306053, + 0.01187564991414547, + 0.01204698160290718, + 0.025674190372228622, + 0.03369749337434769, + 0.03552040457725525, + 0.022622134536504745, + 0.030445646494627, + 0.029270915314555168, + 0.03593546524643898, + 0.02337736263871193, + 0.027239808812737465, + 0.032710593193769455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.008636031299829483 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_8/index_tensor([5077, 5077])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_8/index_tensor([5077, 5077])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7ea932bf9fa5da62a6f0c74d329647ce0e365ac1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_8/index_tensor([5077, 5077])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low, static, wide-angle shot captures the iconic silver and blue M3 GTR parked on the cracked, oil-stained concrete floor of a dimly lit, cavernous underpass at dusk. The car's metallic silver paint gleams with a cold sheen under the harsh overhead fluorescent lights, casting long, sharp shadows and creating brilliant, mirror-like highlights across its aggressive widebody kit, prominent rear wing, and angular blue vinyl livery. After the engine is shut off, an echoing silence suddenly descends, and a violent, high-frequency tremor instantly rips through the body, causing the large rear spoiler to visibly tremble. This intense, under-sprung vibration transforms into a heavier, more pronounced tremor as the entire chassis deliberately rocks on its stiff racing suspension, its weight shifting from side to side in a series of gradually diminishing swaying movements. The movement gradually subsides, and the car slowly and heavily sinks into a final, profound silence. A palpable mechanical tension dissipates into the cool, quiet air, leaving only the machine's silent, majestic presence in the desolate urban environment.", + "seed": 14575, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_8/index_5077_seed_14575_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05471524968743324, + "vlm_logprob": -0.05471524968743324, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.05471524968743324, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02231837995350361 + ], + "opposite_logprob_tokens": [ + -4.397318363189697 + ], + "logprob_reward": -0.02231837995350361, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02231837995350361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28068727254867554, + 0.08491995185613632, + 0.07156168669462204, + 0.044991638511419296, + 0.0659087598323822, + 0.08792824298143387, + 0.048976920545101166, + 0.07906705141067505, + 0.07185839861631393, + 0.16410009562969208 + ], + "credit_map": [ + 0.04334116354584694, + 0.015510677359998226, + 0.022118307650089264, + 0.03144713118672371, + 0.011540418490767479, + 0.04475327208638191, + 0.022644953802227974, + 0.08933135122060776, + 0.010711613111197948, + 0.0056602731347084045, + 0.009669418446719646, + 0.018433045595884323, + 0.005687527358531952, + 0.011678673326969147, + 0.010409596376121044, + 0.012669802643358707, + 0.007917198352515697, + 0.0040793404914438725, + 0.007104257121682167, + 0.014704450964927673, + 0.004426255356520414, + 0.012153050862252712, + 0.008389328606426716, + 0.012787805870175362, + 0.005582058802247047, + 0.0026717104483395815, + 0.0059541966766119, + 0.006085291039198637, + 0.0025878583546727896, + 0.0067912014201283455, + 0.005264684092253447, + 0.010054636746644974, + 0.00800164695829153, + 0.004404847510159016, + 0.006025062873959541, + 0.006361103616654873, + 0.004597902297973633, + 0.01267373375594616, + 0.006590292323380709, + 0.017254170030355453, + 0.011639820411801338, + 0.004495643079280853, + 0.006727773696184158, + 0.010989032685756683, + 0.005049548577517271, + 0.017591193318367004, + 0.008731887675821781, + 0.022703347727656364, + 0.009252605028450489, + 0.004937207326292992, + 0.0055477009154856205, + 0.005662394221872091, + 0.003452194621786475, + 0.007888924330472946, + 0.004871326964348555, + 0.0073645710945129395, + 0.010692237876355648, + 0.005293007008731365, + 0.008021555840969086, + 0.008254886604845524, + 0.005122384987771511, + 0.014670150354504585, + 0.007021704223006964, + 0.019991127774119377, + 0.010414092801511288, + 0.009400852955877781, + 0.006621172185987234, + 0.006594035308808088, + 0.006076188292354345, + 0.009784555062651634, + 0.0067845000885427, + 0.016182996332645416, + 0.02006969042122364, + 0.017501313239336014, + 0.03590632230043411, + 0.018381385132670403, + 0.011921048164367676, + 0.026045482605695724, + 0.01854301057755947, + 0.015731852501630783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.02231837995350361 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014218805357813835 + ], + "opposite_logprob_tokens": [ + -4.264218807220459 + ], + "logprob_reward": -0.014218805357813835, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014218805357813835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2972519099712372, + 0.19093608856201172, + 0.12629172205924988, + 0.029591189697384834, + 0.05290982127189636, + 0.03080151230096817, + 0.032180171459913254, + 0.046866919845342636, + 0.06588489562273026, + 0.12728576362133026 + ], + "credit_map": [ + 0.039181455969810486, + 0.012829678133130074, + 0.02788526751101017, + 0.036563895642757416, + 0.01406820211559534, + 0.04521862789988518, + 0.021337149664759636, + 0.1001676395535469, + 0.01780860126018524, + 0.009864216670393944, + 0.01964723877608776, + 0.027977783232927322, + 0.014570116065442562, + 0.038028739392757416, + 0.021065400913357735, + 0.04197399318218231, + 0.016089189797639847, + 0.007314597722142935, + 0.012597433291375637, + 0.019530590623617172, + 0.009206164628267288, + 0.02710554003715515, + 0.014062232337892056, + 0.020385975018143654, + 0.0033080819994211197, + 0.0025587002746760845, + 0.003841168712824583, + 0.006060770712792873, + 0.002328821225091815, + 0.0037205922417342663, + 0.004270594101399183, + 0.0035024601966142654, + 0.007797859609127045, + 0.004203295800834894, + 0.005887210834771395, + 0.005559856072068214, + 0.004499264992773533, + 0.010492938570678234, + 0.0055006579495966434, + 0.008968737907707691, + 0.0036250760313123465, + 0.0019185796845704317, + 0.00309551227837801, + 0.003268955973908305, + 0.0019381205784156919, + 0.005388068500906229, + 0.004009000491350889, + 0.007558201439678669, + 0.0036578397266566753, + 0.002586635760962963, + 0.003329084487631917, + 0.004482405725866556, + 0.0020380974747240543, + 0.004985562991350889, + 0.003079964080825448, + 0.008020581677556038, + 0.006916590966284275, + 0.003510967828333378, + 0.004941408056765795, + 0.004772806540131569, + 0.002692601177841425, + 0.008964146487414837, + 0.004236391745507717, + 0.010832006111741066, + 0.006267337594181299, + 0.005462558940052986, + 0.013022130355238914, + 0.004789649974554777, + 0.005264992825686932, + 0.007377396337687969, + 0.007247009314596653, + 0.0164538212120533, + 0.011621011421084404, + 0.007389422506093979, + 0.03928816318511963, + 0.00824936106801033, + 0.008720037527382374, + 0.011935900896787643, + 0.006199357565492392, + 0.033882517367601395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.014218805357813835 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207069382071495 + ], + "opposite_logprob_tokens": [ + -2.812070608139038 + ], + "logprob_reward": -0.06207069382071495, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207069382071495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18912744522094727, + 0.13110777735710144, + 0.10607821494340897, + 0.04028947651386261, + 0.02998337894678116, + 0.081645168364048, + 0.03582758456468582, + 0.05900175869464874, + 0.09921696782112122, + 0.22772230207920074 + ], + "credit_map": [ + 0.024541422724723816, + 0.010119215585291386, + 0.028906667605042458, + 0.02554713375866413, + 0.009477272629737854, + 0.024691781029105186, + 0.022305596619844437, + 0.043538350611925125, + 0.014667333103716373, + 0.006560224574059248, + 0.014644509181380272, + 0.022352900356054306, + 0.008993894793093204, + 0.026836073026061058, + 0.016186365857720375, + 0.02086648717522621, + 0.015644047409296036, + 0.006780971772968769, + 0.011116431094706059, + 0.014816764742136002, + 0.006494111847132444, + 0.020332423970103264, + 0.010927051305770874, + 0.01996641419827938, + 0.0052010249346494675, + 0.002833951497450471, + 0.005992373917251825, + 0.007539988495409489, + 0.002661854960024357, + 0.0046651544980704784, + 0.005557180382311344, + 0.005837948992848396, + 0.003416561521589756, + 0.0024448977783322334, + 0.004180221818387508, + 0.00410608621314168, + 0.002863936359062791, + 0.003978992346674204, + 0.004344358574599028, + 0.004648326430469751, + 0.010819144546985626, + 0.005024260375648737, + 0.006109421607106924, + 0.006720697972923517, + 0.004345389548689127, + 0.013864003121852875, + 0.007191551849246025, + 0.02757069282233715, + 0.005141958594322205, + 0.0025513668078929186, + 0.004009730648249388, + 0.0032007405534386635, + 0.0020775403827428818, + 0.006167068611830473, + 0.0028998779598623514, + 0.009779299609363079, + 0.00980567280203104, + 0.005300904158502817, + 0.008512593805789948, + 0.006203745491802692, + 0.0035370648838579655, + 0.00953039899468422, + 0.005911005660891533, + 0.010200373828411102, + 0.010576230473816395, + 0.013087645173072815, + 0.018963972106575966, + 0.010988084599375725, + 0.0073041412979364395, + 0.009034505113959312, + 0.011876523494720459, + 0.01738586463034153, + 0.02064422331750393, + 0.015343860723078251, + 0.052134864032268524, + 0.024234110489487648, + 0.020511766895651817, + 0.02381489984691143, + 0.028833962976932526, + 0.0422046072781086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.06207069382071495 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293117702007294 + ], + "opposite_logprob_tokens": [ + -2.237931251525879 + ], + "logprob_reward": -0.11293117702007294, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293117702007294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26534491777420044, + 0.05833706259727478, + 0.02319801226258278, + 0.04240827262401581, + 0.07251954823732376, + 0.03699583187699318, + 0.0412282831966877, + 0.05002877116203308, + 0.2488899827003479, + 0.16104936599731445 + ], + "credit_map": [ + 0.03368760645389557, + 0.011480850167572498, + 0.03068510629236698, + 0.021233398467302322, + 0.015190890990197659, + 0.04361961781978607, + 0.022541729733347893, + 0.08690569549798965, + 0.008165986277163029, + 0.0021673496812582016, + 0.006142436992377043, + 0.005547306500375271, + 0.0028398740105330944, + 0.007996157743036747, + 0.004399139899760485, + 0.021078815683722496, + 0.0029520683456212282, + 0.0014855313347652555, + 0.0028909966349601746, + 0.0017472110921517015, + 0.0010225105797871947, + 0.004659720230847597, + 0.002208857564255595, + 0.006231116130948067, + 0.005913491826504469, + 0.002658925950527191, + 0.004242028575390577, + 0.005515467841178179, + 0.0029221398290246725, + 0.009462145157158375, + 0.005000789649784565, + 0.006693280767649412, + 0.012302117422223091, + 0.0043272823095321655, + 0.007395726628601551, + 0.008174029178917408, + 0.00513311754912138, + 0.015084254555404186, + 0.007916607894003391, + 0.012186410836875439, + 0.004975779913365841, + 0.002457957249134779, + 0.003961815964430571, + 0.004697064869105816, + 0.002475126413628459, + 0.006600594148039818, + 0.005894137546420097, + 0.005933356937021017, + 0.005154022481292486, + 0.002882069442421198, + 0.003944346681237221, + 0.0038010606076568365, + 0.002760302508249879, + 0.006369470153003931, + 0.004541832022368908, + 0.011775179766118526, + 0.007273644208908081, + 0.002784885698929429, + 0.00392700033262372, + 0.004465373698621988, + 0.0023038790095597506, + 0.008904202841222286, + 0.004203864373266697, + 0.016165921464562416, + 0.03049403242766857, + 0.026076272130012512, + 0.03176164999604225, + 0.01992117241024971, + 0.01542273722589016, + 0.023244908079504967, + 0.032196044921875, + 0.06977318227291107, + 0.02089005522429943, + 0.020193133503198624, + 0.02092568948864937, + 0.01860724575817585, + 0.01307681668549776, + 0.031285353004932404, + 0.017995329573750496, + 0.018075747415423393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.11293117702007294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203719228506088 + ], + "opposite_logprob_tokens": [ + -2.812037229537964 + ], + "logprob_reward": -0.06203719228506088, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203719228506088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27081939578056335, + 0.11132553964853287, + 0.025891635566949844, + 0.04788070172071457, + 0.04481977969408035, + 0.02195821702480316, + 0.030925925821065903, + 0.07945240288972855, + 0.1703958660364151, + 0.19653049111366272 + ], + "credit_map": [ + 0.04231267794966698, + 0.015160744078457355, + 0.020026179030537605, + 0.01898413524031639, + 0.01003286987543106, + 0.044898152351379395, + 0.013845709152519703, + 0.10555891692638397, + 0.01304854080080986, + 0.00451635243371129, + 0.009125346317887306, + 0.013637618161737919, + 0.005900923162698746, + 0.019789939746260643, + 0.009890524670481682, + 0.03541629761457443, + 0.00402110954746604, + 0.0013212092453613877, + 0.0025480221956968307, + 0.0022099700290709734, + 0.0012020221911370754, + 0.004888751544058323, + 0.0024394879583269358, + 0.00726106483489275, + 0.006396129727363586, + 0.003063077572733164, + 0.004975635092705488, + 0.006221629213541746, + 0.0038515820633620024, + 0.010698025114834309, + 0.005550974048674107, + 0.007123644929379225, + 0.007008251268416643, + 0.002300470834597945, + 0.004794296808540821, + 0.005277564749121666, + 0.0027663742657750845, + 0.009226162917912006, + 0.0048040226101875305, + 0.00864263717085123, + 0.0028049075044691563, + 0.00154297542758286, + 0.0025740554556250572, + 0.003219036618247628, + 0.001494168071076274, + 0.0036875391378998756, + 0.0033655334264039993, + 0.0032699997536838055, + 0.005432233680039644, + 0.002529056277126074, + 0.00297811278142035, + 0.0024268836714327335, + 0.0017240733141079545, + 0.005001784767955542, + 0.002435557311400771, + 0.008398224599659443, + 0.012174204923212528, + 0.003992615267634392, + 0.005835470277816057, + 0.004949809983372688, + 0.002762613818049431, + 0.01259949803352356, + 0.004470107611268759, + 0.032668083906173706, + 0.022383632138371468, + 0.0246301107108593, + 0.017428996041417122, + 0.018537523224949837, + 0.01432758942246437, + 0.014617708511650562, + 0.02222277782857418, + 0.03624752163887024, + 0.024633103981614113, + 0.016312509775161743, + 0.014820457436144352, + 0.02366376854479313, + 0.014852608554065228, + 0.03614958003163338, + 0.028109516948461533, + 0.03798893466591835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.06203719228506088 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, static, wide-angle shot captures the iconic silver and blue M3 GTR parked on the cracked, oil-stained concrete floor of a dimly lit, cavernous underpass at dusk. The car's metallic silver paint gleams with a cold sheen under the harsh overhead fluorescent lights, casting long, sharp shadows and creating brilliant, mirror-like highlights across its aggressive widebody kit, prominent rear wing, and angular blue vinyl livery. After the engine is shut off, an echoing silence suddenly descends, and a violent, high-frequency tremor instantly rips through the body, causing the large rear spoiler to visibly tremble. This intense, under-sprung vibration transforms into a heavier, more pronounced tremor as the entire chassis deliberately rocks on its stiff racing suspension, its weight shifting from side to side in a series of gradually diminishing swaying movements. The movement gradually subsides, and the car slowly and heavily sinks into a final, profound silence. A palpable mechanical tension dissipates into the cool, quiet air, leaving only the machine's silent, majestic presence in the desolate urban environment.", + "seed": 14576, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_8/index_5077_seed_14576_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09615297056734562, + "vlm_logprob": -0.09615297056734562, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.09615297056734562, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04228790104389191 + ], + "opposite_logprob_tokens": [ + -3.542287826538086 + ], + "logprob_reward": -0.04228790104389191, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.04228790104389191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14264944195747375, + 0.06981033831834793, + 0.26017096638679504, + 0.14550116658210754, + 0.06522763520479202, + 0.051927559077739716, + 0.06024574115872383, + 0.06425271183252335, + 0.06107519939541817, + 0.07913924008607864 + ], + "credit_map": [ + 0.026125449687242508, + 0.009630539454519749, + 0.01432107575237751, + 0.01354952622205019, + 0.014675985090434551, + 0.019538532942533493, + 0.027402542531490326, + 0.017405793070793152, + 0.011456928215920925, + 0.0059585305862128735, + 0.005045800004154444, + 0.006601973436772823, + 0.006588175892829895, + 0.008686658926308155, + 0.013710535131394863, + 0.011761732399463654, + 0.04248703643679619, + 0.023822205141186714, + 0.04602667689323425, + 0.024603497236967087, + 0.022509202361106873, + 0.049018312245607376, + 0.02202586643397808, + 0.02967817895114422, + 0.022846192121505737, + 0.01999749056994915, + 0.013651877641677856, + 0.02157623879611492, + 0.01766899600625038, + 0.010143551975488663, + 0.0278874933719635, + 0.011729313991963863, + 0.00887847039848566, + 0.00475310068577528, + 0.007401098031550646, + 0.0057454160414636135, + 0.005843769293278456, + 0.012347474694252014, + 0.008446254767477512, + 0.011812051758170128, + 0.007648251485079527, + 0.005032348446547985, + 0.007371316198259592, + 0.0036441695410758257, + 0.0051935045048594475, + 0.007347457110881805, + 0.009577528573572636, + 0.006112983915954828, + 0.010249000042676926, + 0.006822317838668823, + 0.005584610626101494, + 0.005537688732147217, + 0.0050981175154447556, + 0.006624170113354921, + 0.013547860085964203, + 0.006781972479075193, + 0.010060874745249748, + 0.0069273775443434715, + 0.005840229336172342, + 0.006944285240024328, + 0.005622210446745157, + 0.008925027213990688, + 0.013673119246959686, + 0.0062595875933766365, + 0.00947364792227745, + 0.005161385051906109, + 0.005163498688489199, + 0.005510168150067329, + 0.007768799085170031, + 0.010343922302126884, + 0.011150537058711052, + 0.006503241136670113, + 0.014196760952472687, + 0.0064384546130895615, + 0.007997513748705387, + 0.008485391736030579, + 0.007167450152337551, + 0.012636084109544754, + 0.016336536034941673, + 0.005881055258214474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.04228790104389191 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014234436675906181 + ], + "opposite_logprob_tokens": [ + -4.26423454284668 + ], + "logprob_reward": -0.014234436675906181, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014234436675906181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09850548207759857, + 0.052180856466293335, + 0.13346272706985474, + 0.17662137746810913, + 0.14560545980930328, + 0.04721449688076973, + 0.056725095957517624, + 0.08110328018665314, + 0.0966787189245224, + 0.11190244555473328 + ], + "credit_map": [ + 0.01540082972496748, + 0.013148898258805275, + 0.009007508866488934, + 0.011762971989810467, + 0.010837973095476627, + 0.009379614144563675, + 0.016506485641002655, + 0.012461196631193161, + 0.008346944116055965, + 0.0043844301253557205, + 0.005448739044368267, + 0.007675546687096357, + 0.0037815705873072147, + 0.006203201599419117, + 0.010136864148080349, + 0.006203554105013609, + 0.018648797646164894, + 0.01582927815616131, + 0.01635439693927765, + 0.011953532695770264, + 0.013843593187630177, + 0.018641993403434753, + 0.024478957056999207, + 0.01371216680854559, + 0.02832646481692791, + 0.017029820010066032, + 0.017593298107385635, + 0.0152235496789217, + 0.018368225544691086, + 0.034180574119091034, + 0.024827783927321434, + 0.021071678027510643, + 0.021442091092467308, + 0.01156996563076973, + 0.018338775262236595, + 0.013515215367078781, + 0.010937290266156197, + 0.02539692260324955, + 0.022415395826101303, + 0.02198980376124382, + 0.008362825028598309, + 0.005737867671996355, + 0.0038205431774258614, + 0.005229540169239044, + 0.00500895082950592, + 0.0037849436048418283, + 0.01125063095241785, + 0.004019195679575205, + 0.009174801409244537, + 0.00573581550270319, + 0.0077785286121070385, + 0.005876136012375355, + 0.004859421402215958, + 0.008281568996608257, + 0.008979462087154388, + 0.006039358209818602, + 0.013839718885719776, + 0.00917040091007948, + 0.009584104642271996, + 0.007737342733889818, + 0.0073824659921228886, + 0.007166690193116665, + 0.016373585909605026, + 0.009848972782492638, + 0.015574730932712555, + 0.006502942182123661, + 0.010729970410466194, + 0.006632979493588209, + 0.011815302073955536, + 0.01886315643787384, + 0.0140281543135643, + 0.012531484477221966, + 0.022039176896214485, + 0.007984256371855736, + 0.010937380604445934, + 0.010889837518334389, + 0.008496521040797234, + 0.013098260387778282, + 0.02877713181078434, + 0.00967987161129713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.014234436675906181 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135230541229248 + ], + "opposite_logprob_tokens": [ + -1.3135230541229248 + ], + "logprob_reward": -0.3135230541229248, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135230541229248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17896965146064758, + 0.08214327692985535, + 0.1694018691778183, + 0.1241348460316658, + 0.15259158611297607, + 0.07883724570274353, + 0.021884312853217125, + 0.05253390967845917, + 0.04201063886284828, + 0.09749259799718857 + ], + "credit_map": [ + 0.02861771732568741, + 0.021649915724992752, + 0.021224170923233032, + 0.017760755494236946, + 0.014860671013593674, + 0.02725870907306671, + 0.024124344810843468, + 0.02347337268292904, + 0.013846942223608494, + 0.007476353086531162, + 0.009335021488368511, + 0.008483472280204296, + 0.006507509388029575, + 0.013658621348440647, + 0.013724192976951599, + 0.009111161343753338, + 0.023779550567269325, + 0.014109466224908829, + 0.01657439023256302, + 0.01450927834957838, + 0.01724107377231121, + 0.02381284348666668, + 0.04034803807735443, + 0.01902722753584385, + 0.01907593384385109, + 0.0116544459015131, + 0.016545170918107033, + 0.009342804551124573, + 0.012164712883532047, + 0.019549714401364326, + 0.01791449449956417, + 0.017887568101286888, + 0.021740904077887535, + 0.011039750650525093, + 0.021333899348974228, + 0.011925916187465191, + 0.01294624526053667, + 0.02841946855187416, + 0.018989838659763336, + 0.02619555965065956, + 0.013949692249298096, + 0.008703027851879597, + 0.008962664753198624, + 0.0070191905833780766, + 0.0066931466571986675, + 0.011559953913092613, + 0.014988943003118038, + 0.006960621569305658, + 0.003115496365353465, + 0.0021098447032272816, + 0.0015539820306003094, + 0.002498040907084942, + 0.0024330667220056057, + 0.0022668999154120684, + 0.004335430916398764, + 0.0035715498961508274, + 0.008839749731123447, + 0.004921307787299156, + 0.004179822281002998, + 0.004682532511651516, + 0.00477061839774251, + 0.005360807292163372, + 0.01385904848575592, + 0.0059200190007686615, + 0.006979535799473524, + 0.004299882799386978, + 0.004547822754830122, + 0.003386803437024355, + 0.003598229493945837, + 0.008799861185252666, + 0.006257709581404924, + 0.004140792414546013, + 0.018021205440163612, + 0.0074617923237383366, + 0.010135426186025143, + 0.011703110300004482, + 0.0076723359525203705, + 0.01314244233071804, + 0.021273506805300713, + 0.008082782849669456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.3135230541229248 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062061842530965805 + ], + "opposite_logprob_tokens": [ + -2.8120617866516113 + ], + "logprob_reward": -0.062061842530965805, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062061842530965805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08354268223047256, + 0.044295649975538254, + 0.18621444702148438, + 0.09069982171058655, + 0.1463388055562973, + 0.04759300872683525, + 0.08686988800764084, + 0.12910571694374084, + 0.07362961769104004, + 0.11171037703752518 + ], + "credit_map": [ + 0.012651323340833187, + 0.00997863058000803, + 0.010957317426800728, + 0.007850212045013905, + 0.008087540976703167, + 0.007617252413183451, + 0.012034891173243523, + 0.014365514740347862, + 0.0059128813445568085, + 0.005948783829808235, + 0.003542146412655711, + 0.005941166076809168, + 0.0040126279927790165, + 0.003085802774876356, + 0.0084826136007905, + 0.0073696281760931015, + 0.03255593031644821, + 0.018242334946990013, + 0.02115386724472046, + 0.017438123002648354, + 0.018622389063239098, + 0.028587281703948975, + 0.03005223348736763, + 0.019562311470508575, + 0.012616469524800777, + 0.009001626633107662, + 0.008750203996896744, + 0.013691206462681293, + 0.010092289187014103, + 0.008136821910738945, + 0.01977177895605564, + 0.008639427833259106, + 0.02354351617395878, + 0.016768183559179306, + 0.012085545808076859, + 0.018481340259313583, + 0.013871847651898861, + 0.012421024963259697, + 0.03349841386079788, + 0.01566893234848976, + 0.007210844196379185, + 0.004507866688072681, + 0.004067395813763142, + 0.0052796402014791965, + 0.0045540533028542995, + 0.004128387663513422, + 0.013578807935118675, + 0.004266011994332075, + 0.014063438400626183, + 0.010060124099254608, + 0.006756125018000603, + 0.00961335189640522, + 0.0077200219966471195, + 0.008435186930000782, + 0.021715836599469185, + 0.008505803532898426, + 0.018528703600168228, + 0.0138290049508214, + 0.01353532262146473, + 0.012304535135626793, + 0.01272459514439106, + 0.015242512337863445, + 0.02719244174659252, + 0.015748608857393265, + 0.011482387781143188, + 0.009220421314239502, + 0.006667908281087875, + 0.008646891452372074, + 0.00628212746232748, + 0.008490483276546001, + 0.014866488985717297, + 0.007972906343638897, + 0.020905954763293266, + 0.00827858317643404, + 0.010661771520972252, + 0.014757421799004078, + 0.008327118121087551, + 0.015184841118752956, + 0.025668414309620857, + 0.0079262750223279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.062061842530965805 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486576184630394 + ], + "opposite_logprob_tokens": [ + -3.0486576557159424 + ], + "logprob_reward": -0.0486576184630394, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486576184630394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14569053053855896, + 0.04025133326649666, + 0.10929480195045471, + 0.21099813282489777, + 0.08975694328546524, + 0.04341539368033409, + 0.0472884327173233, + 0.07796388119459152, + 0.15285249054431915, + 0.0824880301952362 + ], + "credit_map": [ + 0.022594720125198364, + 0.010476779192686081, + 0.012952948920428753, + 0.013788015581667423, + 0.01355687528848648, + 0.019417904317378998, + 0.028344132006168365, + 0.024559149518609047, + 0.0071106250397861, + 0.00439927913248539, + 0.003337704576551914, + 0.005418137181550264, + 0.004042609594762325, + 0.0031685216818004847, + 0.008527274243533611, + 0.004247184377163649, + 0.018516255542635918, + 0.011609477922320366, + 0.011979824863374233, + 0.01342787779867649, + 0.012839422561228275, + 0.011345995590090752, + 0.01898878999054432, + 0.010587157681584358, + 0.028196852654218674, + 0.024619068950414658, + 0.016589757055044174, + 0.031224731355905533, + 0.02533290907740593, + 0.023286057636141777, + 0.040863584727048874, + 0.020885175094008446, + 0.01487780548632145, + 0.008845641277730465, + 0.007940463721752167, + 0.010200101882219315, + 0.009475981816649437, + 0.006029796786606312, + 0.023544758558273315, + 0.008842402137815952, + 0.006313175428658724, + 0.004432787653058767, + 0.003038659691810608, + 0.004615294747054577, + 0.006519197020679712, + 0.005635822657495737, + 0.008471734821796417, + 0.004388718865811825, + 0.0069601875729858875, + 0.0046777124516665936, + 0.006881260313093662, + 0.0049718101508915424, + 0.004047355614602566, + 0.006742491852492094, + 0.006740032695233822, + 0.0062675802037119865, + 0.01234615407884121, + 0.006843083538115025, + 0.007801243104040623, + 0.006574125494807959, + 0.009045371785759926, + 0.007191102486103773, + 0.01872030273079872, + 0.00944249052554369, + 0.02240883745253086, + 0.01470714807510376, + 0.01634453609585762, + 0.013797961175441742, + 0.02142278663814068, + 0.020424097776412964, + 0.027782516553997993, + 0.015964603051543236, + 0.011974518187344074, + 0.008595360442996025, + 0.009760265238583088, + 0.012201167643070221, + 0.006600562483072281, + 0.012869036756455898, + 0.012263222597539425, + 0.008223903365433216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.0486576184630394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, static, wide-angle shot captures the iconic silver and blue M3 GTR parked on the cracked, oil-stained concrete floor of a dimly lit, cavernous underpass at dusk. The car's metallic silver paint gleams with a cold sheen under the harsh overhead fluorescent lights, casting long, sharp shadows and creating brilliant, mirror-like highlights across its aggressive widebody kit, prominent rear wing, and angular blue vinyl livery. After the engine is shut off, an echoing silence suddenly descends, and a violent, high-frequency tremor instantly rips through the body, causing the large rear spoiler to visibly tremble. This intense, under-sprung vibration transforms into a heavier, more pronounced tremor as the entire chassis deliberately rocks on its stiff racing suspension, its weight shifting from side to side in a series of gradually diminishing swaying movements. The movement gradually subsides, and the car slowly and heavily sinks into a final, profound silence. A palpable mechanical tension dissipates into the cool, quiet air, leaving only the machine's silent, majestic presence in the desolate urban environment.", + "seed": 14577, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_8/index_5077_seed_14577_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07263335138559342, + "vlm_logprob": -0.07263335138559342, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.07263335138559342, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037434838712215424 + ], + "opposite_logprob_tokens": [ + -3.7874348163604736 + ], + "logprob_reward": -0.037434838712215424, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.037434838712215424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20635810494422913, + 0.12621185183525085, + 0.061459414660930634, + 0.035111527889966965, + 0.07151611894369125, + 0.09470925480127335, + 0.1627727746963501, + 0.06998135149478912, + 0.10792279243469238, + 0.06395681947469711 + ], + "credit_map": [ + 0.025878116488456726, + 0.014622333459556103, + 0.024896012619137764, + 0.013425816781818867, + 0.03246501088142395, + 0.02866468019783497, + 0.030013661831617355, + 0.03639248013496399, + 0.013822861015796661, + 0.006172372493892908, + 0.013191563077270985, + 0.007687080185860395, + 0.013370403088629246, + 0.01952580362558365, + 0.02898971550166607, + 0.023452043533325195, + 0.006478932686150074, + 0.0037805940955877304, + 0.0057525974698364735, + 0.004014568869024515, + 0.00834673922508955, + 0.011305303312838078, + 0.013161098584532738, + 0.008619583211839199, + 0.003696546657010913, + 0.0025954637676477432, + 0.004862458445131779, + 0.002544716466218233, + 0.004740090109407902, + 0.006071394309401512, + 0.0050314851105213165, + 0.005569374188780785, + 0.01812741719186306, + 0.005413268227130175, + 0.00841840822249651, + 0.005018504802137613, + 0.010529291816055775, + 0.009824465960264206, + 0.007161426357924938, + 0.007023337297141552, + 0.015158231370151043, + 0.006393977906554937, + 0.014860955998301506, + 0.008132093586027622, + 0.012305575422942638, + 0.01319873332977295, + 0.011667008511722088, + 0.012992682866752148, + 0.014690034091472626, + 0.010715138167142868, + 0.010594676248729229, + 0.011469160206615925, + 0.01104586012661457, + 0.015024026855826378, + 0.03824399411678314, + 0.050989892333745956, + 0.008634749799966812, + 0.009463250637054443, + 0.01143174059689045, + 0.009707105346024036, + 0.005918258801102638, + 0.009148601442575455, + 0.005959764588624239, + 0.009717882610857487, + 0.015409293584525585, + 0.0073606958612799644, + 0.010174250230193138, + 0.012439262121915817, + 0.008588959462940693, + 0.029786352068185806, + 0.008245163597166538, + 0.01591881550848484, + 0.009646165184676647, + 0.004981670994311571, + 0.013182766735553741, + 0.011342816054821014, + 0.004615719895809889, + 0.006911302916705608, + 0.0076141334138810635, + 0.005662250332534313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.037434838712215424 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014226209372282028 + ], + "opposite_logprob_tokens": [ + -4.26422643661499 + ], + "logprob_reward": -0.014226209372282028, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014226209372282028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15810814499855042, + 0.1308136284351349, + 0.07464800775051117, + 0.09514882415533066, + 0.03784669563174248, + 0.06184094026684761, + 0.12994077801704407, + 0.05936636030673981, + 0.12254331260919571, + 0.1297433078289032 + ], + "credit_map": [ + 0.027154400944709778, + 0.012003505602478981, + 0.015363023616373539, + 0.012276588007807732, + 0.024306196719408035, + 0.020398957654833794, + 0.026635628193616867, + 0.019969850778579712, + 0.015542314387857914, + 0.006047587841749191, + 0.011885955929756165, + 0.008474759757518768, + 0.011842207983136177, + 0.020609358325600624, + 0.030784817412495613, + 0.02562662959098816, + 0.007341538090258837, + 0.004077021963894367, + 0.007510767783969641, + 0.00571408960968256, + 0.0095740407705307, + 0.008675412274897099, + 0.015381459146738052, + 0.016373679041862488, + 0.006882172543555498, + 0.005117342807352543, + 0.008722729049623013, + 0.006522114388644695, + 0.009702872484922409, + 0.011428632773458958, + 0.02294679917395115, + 0.023826155811548233, + 0.004499285481870174, + 0.0027609060052782297, + 0.004488582257181406, + 0.003124645445495844, + 0.005092164967209101, + 0.006821138318628073, + 0.006508265621960163, + 0.0045517063699662685, + 0.007843433879315853, + 0.004367702174931765, + 0.010214793495833874, + 0.004784457851201296, + 0.007922475226223469, + 0.008690288290381432, + 0.007123212795704603, + 0.010894577018916607, + 0.010459478944540024, + 0.008159460499882698, + 0.012323737144470215, + 0.009451162070035934, + 0.008161826059222221, + 0.017012665048241615, + 0.027674032375216484, + 0.03669840097427368, + 0.0077813128009438515, + 0.006807759869843721, + 0.006753299850970507, + 0.009504016488790512, + 0.0053771925158798695, + 0.008230083622038364, + 0.006817913614213467, + 0.008094779215753078, + 0.012478734366595745, + 0.018125714734196663, + 0.015522805973887444, + 0.016328725963830948, + 0.011767123825848103, + 0.011881699785590172, + 0.019854217767715454, + 0.016584284603595734, + 0.017353449016809464, + 0.01132976170629263, + 0.026539357379078865, + 0.021816136315464973, + 0.009980263188481331, + 0.013443462550640106, + 0.016404997557401657, + 0.012875888496637344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.014226209372282028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16041673719882965 + ], + "opposite_logprob_tokens": [ + -1.9104167222976685 + ], + "logprob_reward": -0.16041673719882965, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16041673719882965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15247225761413574, + 0.07921791076660156, + 0.07067632675170898, + 0.033761367201805115, + 0.1253446489572525, + 0.054616235196590424, + 0.20337048172950745, + 0.08651651442050934, + 0.12059300392866135, + 0.07343126833438873 + ], + "credit_map": [ + 0.019242776557803154, + 0.012725325301289558, + 0.013593044131994247, + 0.01369966845959425, + 0.03381446748971939, + 0.019784679636359215, + 0.021190449595451355, + 0.0184218417853117, + 0.009250248782336712, + 0.005548521876335144, + 0.0072951107285916805, + 0.006537541747093201, + 0.011501260101795197, + 0.015077979303896427, + 0.014419825747609138, + 0.009587419219315052, + 0.010700886137783527, + 0.005207783542573452, + 0.007018738426268101, + 0.005573789123445749, + 0.012029968202114105, + 0.013582480140030384, + 0.009059651754796505, + 0.0075030275620520115, + 0.0034412583336234093, + 0.0022079136688262224, + 0.0035778102464973927, + 0.0027276279870420694, + 0.004603114444762468, + 0.007467500399798155, + 0.00611584447324276, + 0.0036202992778271437, + 0.03065727837383747, + 0.008888090029358864, + 0.017817026004195213, + 0.009341890923678875, + 0.01680637337267399, + 0.01699235662817955, + 0.012422611005604267, + 0.012419029138982296, + 0.008757954463362694, + 0.0034496390726417303, + 0.006990964524447918, + 0.004768661689013243, + 0.0046858713030815125, + 0.007565915584564209, + 0.006879999302327633, + 0.011517226696014404, + 0.014166592620313168, + 0.012846014462411404, + 0.015182936564087868, + 0.01151940505951643, + 0.012231535278260708, + 0.020352395251393318, + 0.05394161120057106, + 0.06312999874353409, + 0.010552022606134415, + 0.009935584850609303, + 0.010548282414674759, + 0.012384959496557713, + 0.00866921991109848, + 0.012328214943408966, + 0.009938734583556652, + 0.012159491889178753, + 0.014488409273326397, + 0.014059155248105526, + 0.014726007357239723, + 0.013431649655103683, + 0.012087786570191383, + 0.022662032395601273, + 0.013527707196772099, + 0.01561026182025671, + 0.00945615116506815, + 0.006404661573469639, + 0.014392221346497536, + 0.012214764952659607, + 0.006355727091431618, + 0.0078930938616395, + 0.009019679389894009, + 0.007694973610341549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.16041673719882965 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890478566288948 + ], + "opposite_logprob_tokens": [ + -2.464047908782959 + ], + "logprob_reward": -0.0890478566288948, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890478566288948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33747968077659607, + 0.05466868355870247, + 0.06628965586423874, + 0.05091651901602745, + 0.02296026051044464, + 0.051637500524520874, + 0.08060160279273987, + 0.05196079611778259, + 0.23795856535434723, + 0.045526739209890366 + ], + "credit_map": [ + 0.0639362707734108, + 0.022240659222006798, + 0.04607599601149559, + 0.023927561938762665, + 0.03135555982589722, + 0.045136354863643646, + 0.03844940662384033, + 0.06635783612728119, + 0.0071236807852983475, + 0.004420376382768154, + 0.00491016311571002, + 0.004692533984780312, + 0.004432591609656811, + 0.006714079063385725, + 0.009552987292408943, + 0.012822271324694157, + 0.007060478441417217, + 0.004761745221912861, + 0.010415763594210148, + 0.0058921040035784245, + 0.00788387842476368, + 0.010252593085169792, + 0.009785789065063, + 0.010237304493784904, + 0.006652017589658499, + 0.002329680137336254, + 0.005155783612281084, + 0.00349379307590425, + 0.004738850053399801, + 0.006011716090142727, + 0.010340950451791286, + 0.012193726375699043, + 0.0028709734324365854, + 0.0021997662261128426, + 0.002933494048193097, + 0.002487853402271867, + 0.0029403632506728172, + 0.002697802847251296, + 0.0027414741925895214, + 0.004088531713932753, + 0.006493743974715471, + 0.0033607608638703823, + 0.00527548510581255, + 0.0038650580681860447, + 0.00987267680466175, + 0.006347718648612499, + 0.005503423511981964, + 0.010918629355728626, + 0.011509709060192108, + 0.0073851654306054115, + 0.009072811342775822, + 0.008206001482903957, + 0.015680914744734764, + 0.011013173498213291, + 0.007734996732324362, + 0.00999883096665144, + 0.006732879672199488, + 0.00474211573600769, + 0.006170730106532574, + 0.005283679813146591, + 0.0049362643621861935, + 0.011267664842307568, + 0.004440618213266134, + 0.008386846631765366, + 0.027162116020917892, + 0.027273545041680336, + 0.03313241899013519, + 0.055147819221019745, + 0.017668429762125015, + 0.019036229699850082, + 0.029402270913124084, + 0.02913574129343033, + 0.005936193745583296, + 0.004652660805732012, + 0.007662408985197544, + 0.007443179376423359, + 0.005148631054908037, + 0.003975379280745983, + 0.005388124380260706, + 0.005320160649716854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0890478566288948 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204111501574516 + ], + "opposite_logprob_tokens": [ + -2.8120410442352295 + ], + "logprob_reward": -0.06204111501574516, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204111501574516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34931957721710205, + 0.043833471834659576, + 0.04945024475455284, + 0.05470733717083931, + 0.04408932104706764, + 0.05735760182142258, + 0.10338544100522995, + 0.042697250843048096, + 0.18396233022212982, + 0.07119739055633545 + ], + "credit_map": [ + 0.06195253133773804, + 0.01983780786395073, + 0.05044586583971977, + 0.023325586691498756, + 0.03758059814572334, + 0.038902487605810165, + 0.045225590467453, + 0.07204913347959518, + 0.0055221738293766975, + 0.0030952822417020798, + 0.007544153835624456, + 0.0032594550866633654, + 0.004770149942487478, + 0.005566198378801346, + 0.005004066042602062, + 0.009071996435523033, + 0.006089536007493734, + 0.002976867603138089, + 0.006083374843001366, + 0.00380691047757864, + 0.004743937402963638, + 0.006847334560006857, + 0.007685478311032057, + 0.011216811835765839, + 0.005457918159663677, + 0.0021353766787797213, + 0.005994829814881086, + 0.0035959184169769287, + 0.004431913141161203, + 0.007466710172593594, + 0.011907869949936867, + 0.01371680572628975, + 0.00677398731932044, + 0.0036874832585453987, + 0.0035062418319284916, + 0.0038460211362689734, + 0.009508815594017506, + 0.006191026419401169, + 0.005717603489756584, + 0.004858145955950022, + 0.01104435883462429, + 0.004303739406168461, + 0.0078117805533111095, + 0.005061571951955557, + 0.005356506444513798, + 0.00762546481564641, + 0.00607866607606411, + 0.010075521655380726, + 0.005845234729349613, + 0.003505374537780881, + 0.006481235846877098, + 0.00498576695099473, + 0.009934627451002598, + 0.007214047480374575, + 0.03048902377486229, + 0.034930139780044556, + 0.004556899890303612, + 0.0041316538117825985, + 0.004034011624753475, + 0.007967169396579266, + 0.002938765799626708, + 0.006878528278321028, + 0.00545032462105155, + 0.0067399004474282265, + 0.02137521095573902, + 0.022605258971452713, + 0.02566428855061531, + 0.041868992149829865, + 0.013460726477205753, + 0.01731487549841404, + 0.019009677693247795, + 0.022663328796625137, + 0.01001631561666727, + 0.00813974253833294, + 0.01415136456489563, + 0.009664284065365791, + 0.006005324888974428, + 0.00828515738248825, + 0.007699264213442802, + 0.007235944736748934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.06204111501574516 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, static, wide-angle shot captures the iconic silver and blue M3 GTR parked on the cracked, oil-stained concrete floor of a dimly lit, cavernous underpass at dusk. The car's metallic silver paint gleams with a cold sheen under the harsh overhead fluorescent lights, casting long, sharp shadows and creating brilliant, mirror-like highlights across its aggressive widebody kit, prominent rear wing, and angular blue vinyl livery. After the engine is shut off, an echoing silence suddenly descends, and a violent, high-frequency tremor instantly rips through the body, causing the large rear spoiler to visibly tremble. This intense, under-sprung vibration transforms into a heavier, more pronounced tremor as the entire chassis deliberately rocks on its stiff racing suspension, its weight shifting from side to side in a series of gradually diminishing swaying movements. The movement gradually subsides, and the car slowly and heavily sinks into a final, profound silence. A palpable mechanical tension dissipates into the cool, quiet air, leaving only the machine's silent, majestic presence in the desolate urban environment.", + "seed": 14578, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_8/index_5077_seed_14578_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09838956259191037, + "vlm_logprob": -0.09838956259191037, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.09838956259191037, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037392351776361465 + ], + "opposite_logprob_tokens": [ + -3.7873923778533936 + ], + "logprob_reward": -0.037392351776361465, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.037392351776361465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24891123175621033, + 0.10216018557548523, + 0.08910220861434937, + 0.12298817932605743, + 0.0432172529399395, + 0.06550293415784836, + 0.07743579149246216, + 0.06694746762514114, + 0.10618941485881805, + 0.07754534482955933 + ], + "credit_map": [ + 0.03770584613084793, + 0.032231830060482025, + 0.02722788415849209, + 0.03429505601525307, + 0.021222781389951706, + 0.02230227179825306, + 0.03804740309715271, + 0.03587815910577774, + 0.014035898260772228, + 0.011454596184194088, + 0.013366150669753551, + 0.015124557539820671, + 0.011080675758421421, + 0.01132513489574194, + 0.015052956528961658, + 0.010720224119722843, + 0.01369565725326538, + 0.012634365819394588, + 0.010037105530500412, + 0.01137333270162344, + 0.010597866959869862, + 0.00843492429703474, + 0.012115373276174068, + 0.010213585570454597, + 0.019650479778647423, + 0.018559889867901802, + 0.016146207228302956, + 0.016121063381433487, + 0.010747038759291172, + 0.012613574042916298, + 0.014808566309511662, + 0.014341358095407486, + 0.0068718669936060905, + 0.0054703387431800365, + 0.004858084488660097, + 0.005643325392156839, + 0.003955557011067867, + 0.004162718541920185, + 0.006314690224826336, + 0.0059406692162156105, + 0.00626664562150836, + 0.006458678282797337, + 0.007770454045385122, + 0.008021816611289978, + 0.008398783393204212, + 0.006061439868062735, + 0.012214756570756435, + 0.010310359299182892, + 0.015249115414917469, + 0.010435841977596283, + 0.011944636702537537, + 0.008650793693959713, + 0.00590283190831542, + 0.009123415686190128, + 0.007084884215146303, + 0.00904427282512188, + 0.006337588187307119, + 0.008973667398095131, + 0.009618674404919147, + 0.009687433950603008, + 0.006884817034006119, + 0.005908226128667593, + 0.009194431826472282, + 0.010342631489038467, + 0.010691025294363499, + 0.011748631484806538, + 0.016934573650360107, + 0.012232464738190174, + 0.01103824470192194, + 0.012021135538816452, + 0.01734730415046215, + 0.014176039025187492, + 0.007350742816925049, + 0.0060430592857301235, + 0.0168988686054945, + 0.007838052697479725, + 0.010232371278107166, + 0.0076422300189733505, + 0.008684813976287842, + 0.012855201959609985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.037392351776361465 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609945483505726 + ], + "opposite_logprob_tokens": [ + -4.141099452972412 + ], + "logprob_reward": -0.01609945483505726, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609945483505726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23052768409252167, + 0.0998324304819107, + 0.06438083946704865, + 0.07814246416091919, + 0.046251747757196426, + 0.08030813932418823, + 0.07147123664617538, + 0.07293695211410522, + 0.10039601475000381, + 0.155752494931221 + ], + "credit_map": [ + 0.027514632791280746, + 0.027864420786499977, + 0.03306043893098831, + 0.03795899823307991, + 0.027538785710930824, + 0.019284160807728767, + 0.030843092128634453, + 0.02646315097808838, + 0.013258790597319603, + 0.010047182440757751, + 0.01495976373553276, + 0.017376692965626717, + 0.014445222914218903, + 0.010124454274773598, + 0.011037660762667656, + 0.00858265720307827, + 0.00791031215339899, + 0.006238806992769241, + 0.008844477124512196, + 0.009906596504151821, + 0.009433342143893242, + 0.006825624033808708, + 0.00879538431763649, + 0.0064262947998940945, + 0.012081480585038662, + 0.01468278095126152, + 0.007902778685092926, + 0.009734280407428741, + 0.007556635420769453, + 0.008898274041712284, + 0.00863194465637207, + 0.008654288947582245, + 0.005591064225882292, + 0.003926287405192852, + 0.007153204642236233, + 0.008927551098167896, + 0.004820911679416895, + 0.004757900256663561, + 0.006276440806686878, + 0.004798384383320808, + 0.007809842005372047, + 0.006517400499433279, + 0.0113193579018116, + 0.011660123243927956, + 0.009474532678723335, + 0.007004636339843273, + 0.015463980846107006, + 0.011058257892727852, + 0.01702580787241459, + 0.010049525648355484, + 0.008229312486946583, + 0.007526958826929331, + 0.0061536128632724285, + 0.007962147705256939, + 0.005711244884878397, + 0.008812624029815197, + 0.008718929253518581, + 0.011134592816233635, + 0.01115505862981081, + 0.009461085312068462, + 0.006104576867073774, + 0.007061549928039312, + 0.009301397018134594, + 0.009999765083193779, + 0.009517689235508442, + 0.013324564322829247, + 0.010142695158720016, + 0.014842026866972446, + 0.011462169699370861, + 0.010723927989602089, + 0.016009332612156868, + 0.01437360793352127, + 0.01183952484279871, + 0.012641817331314087, + 0.038888800889253616, + 0.020883651450276375, + 0.015057924203574657, + 0.01774086058139801, + 0.016620228067040443, + 0.022079678252339363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.01609945483505726 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562460601329803 + ], + "opposite_logprob_tokens": [ + -1.6006245613098145 + ], + "logprob_reward": -0.22562460601329803, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562460601329803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2506961524486542, + 0.09040281176567078, + 0.0579567551612854, + 0.03691506013274193, + 0.04823052138090134, + 0.07518336176872253, + 0.02765207178890705, + 0.0995272621512413, + 0.06020037457346916, + 0.25323569774627686 + ], + "credit_map": [ + 0.0372089259326458, + 0.04205887019634247, + 0.031120430678129196, + 0.03259376809000969, + 0.0223851278424263, + 0.024315182119607925, + 0.031234832480549812, + 0.02977898344397545, + 0.01416882872581482, + 0.01188603974878788, + 0.009168860502541065, + 0.0159598495811224, + 0.008054540492594242, + 0.008948815986514091, + 0.01161983422935009, + 0.010596036911010742, + 0.00868930947035551, + 0.008542150259017944, + 0.005826955195516348, + 0.007841961458325386, + 0.0062643191777169704, + 0.0058062938041985035, + 0.0076142712496221066, + 0.007371493615210056, + 0.004964860621839762, + 0.004940544720739126, + 0.0050508566200733185, + 0.005110680591315031, + 0.0037376917898654938, + 0.00355272414162755, + 0.005511065945029259, + 0.00404663011431694, + 0.006671865936368704, + 0.008815770037472248, + 0.004347209352999926, + 0.006045707501471043, + 0.0036552739329636097, + 0.0056634340435266495, + 0.006069762632250786, + 0.006961495615541935, + 0.009264880791306496, + 0.01007949747145176, + 0.008202162571251392, + 0.011216913349926472, + 0.00823007058352232, + 0.0073317415080964565, + 0.011847655288875103, + 0.0090104341506958, + 0.003394500585272908, + 0.003018440678715706, + 0.00385058531537652, + 0.0049402001313865185, + 0.002964983694255352, + 0.002893137978389859, + 0.003638316411525011, + 0.002951905597001314, + 0.012185539118945599, + 0.02043132483959198, + 0.014118527993559837, + 0.014580811373889446, + 0.008629905059933662, + 0.010873247869312763, + 0.00862790085375309, + 0.010079996660351753, + 0.007009181659668684, + 0.006515687797218561, + 0.006927083246409893, + 0.008107168599963188, + 0.006038677878677845, + 0.006255741231143475, + 0.010380894877016544, + 0.008965934626758099, + 0.0180969201028347, + 0.012458283454179764, + 0.062462981790304184, + 0.03355909883975983, + 0.03003041446208954, + 0.02951396070420742, + 0.026718100532889366, + 0.04039590433239937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.22562460601329803 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281636476516724 + ], + "opposite_logprob_tokens": [ + -2.0178163051605225 + ], + "logprob_reward": -0.14281636476516724, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14281636476516724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28105226159095764, + 0.06998412311077118, + 0.04490917548537254, + 0.02903803437948227, + 0.041757285594940186, + 0.0930621325969696, + 0.05808572471141815, + 0.10728947073221207, + 0.13079886138439178, + 0.14402291178703308 + ], + "credit_map": [ + 0.0336398184299469, + 0.04923585057258606, + 0.04235570877790451, + 0.05382170155644417, + 0.0174272358417511, + 0.027164479717612267, + 0.02823389507830143, + 0.029173582792282104, + 0.009762842208147049, + 0.01140505913645029, + 0.010783305391669273, + 0.015532607212662697, + 0.00483536534011364, + 0.005733622703701258, + 0.005987253040075302, + 0.005944068543612957, + 0.00535499956458807, + 0.005744307301938534, + 0.0051265014335513115, + 0.0062575372867286205, + 0.004018974490463734, + 0.0037033571861684322, + 0.008096370846033096, + 0.006607126910239458, + 0.004805347882211208, + 0.004480438306927681, + 0.003043845761567354, + 0.004126363433897495, + 0.002110081724822521, + 0.002685839543119073, + 0.0035112258046865463, + 0.004274889826774597, + 0.005579371005296707, + 0.004728867672383785, + 0.004565064329653978, + 0.006648477166891098, + 0.004268052522093058, + 0.004416349343955517, + 0.005082814488559961, + 0.006468289531767368, + 0.007849195040762424, + 0.006883563008159399, + 0.009391475468873978, + 0.013065749779343605, + 0.010060799308121204, + 0.008582035079598427, + 0.020780224353075027, + 0.016449090093374252, + 0.006807239260524511, + 0.005633429158478975, + 0.008617684245109558, + 0.00985026452690363, + 0.005145135801285505, + 0.005658814683556557, + 0.008500121533870697, + 0.007873034104704857, + 0.011955074034631252, + 0.00882946327328682, + 0.01356092095375061, + 0.014687693677842617, + 0.01610199734568596, + 0.013537190854549408, + 0.013815195299685001, + 0.014801939018070698, + 0.009030235931277275, + 0.00687779625877738, + 0.020410409197211266, + 0.021329376846551895, + 0.01447196677327156, + 0.012541737407445908, + 0.02694062329828739, + 0.01919671893119812, + 0.01011194009333849, + 0.00857791118323803, + 0.03150489553809166, + 0.020104339346289635, + 0.021422946825623512, + 0.015708066523075104, + 0.017056114971637726, + 0.0195367019623518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.14281636476516724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a low, static, wide-angle shot of a silver and blue M3 GTR parked in a dimly lit, cavernous underpass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the car's paint metallic silver with a cold sheen, featuring an aggressive widebody kit, prominent rear wing, and angular blue vinyl livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lighting consist of harsh overhead fluorescent lights casting long, sharp shadows and creating brilliant, mirror-like highlights on the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car's chassis visibly rock from side to side in a series of gradually diminishing swaying movements after the engine is shut off?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video end with the car slowly and heavily sinking into a final, profound silence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001503556966782 + ], + "opposite_logprob_tokens": [ + -2.6950149536132812 + ], + "logprob_reward": -0.07001503556966782, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001503556966782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1953166425228119, + 0.07584918290376663, + 0.06272188574075699, + 0.05276312306523323, + 0.0508006326854229, + 0.09467437118291855, + 0.05632789805531502, + 0.05840226635336876, + 0.15321025252342224, + 0.19993376731872559 + ], + "credit_map": [ + 0.03154761716723442, + 0.030388902872800827, + 0.021398302167654037, + 0.030888473615050316, + 0.013244982808828354, + 0.017619334161281586, + 0.029946649447083473, + 0.02028236724436283, + 0.01073282491415739, + 0.012278690934181213, + 0.00836306530982256, + 0.013348717242479324, + 0.005543902516365051, + 0.006550911348313093, + 0.009611132554709911, + 0.009419934824109077, + 0.006623642053455114, + 0.006174394860863686, + 0.006485283374786377, + 0.008512372151017189, + 0.005880220793187618, + 0.005436991807073355, + 0.011814301833510399, + 0.01179467886686325, + 0.007157346699386835, + 0.004977373871952295, + 0.006345566362142563, + 0.00949655007570982, + 0.00425854604691267, + 0.004303248133510351, + 0.008053609170019627, + 0.008170883171260357, + 0.006256811320781708, + 0.005093029234558344, + 0.006328805815428495, + 0.00853427778929472, + 0.005755433347076178, + 0.005585855804383755, + 0.006745439022779465, + 0.006500978488475084, + 0.008343820460140705, + 0.007348453160375357, + 0.009183412417769432, + 0.012435766868293285, + 0.010166442021727562, + 0.00911381933838129, + 0.021382395178079605, + 0.016700254753232002, + 0.007962333969771862, + 0.006182167213410139, + 0.006472030654549599, + 0.006958199664950371, + 0.0051282611675560474, + 0.005458120256662369, + 0.011181958951056004, + 0.006984828505665064, + 0.006803014315664768, + 0.005932494066655636, + 0.007806433830410242, + 0.00715235061943531, + 0.010744660161435604, + 0.006847790442407131, + 0.0061621214263141155, + 0.006953399628400803, + 0.01298894826322794, + 0.009207276627421379, + 0.021914362907409668, + 0.019815508276224136, + 0.019391505047678947, + 0.015357078984379768, + 0.031005917116999626, + 0.023529650643467903, + 0.012009522877633572, + 0.01176010724157095, + 0.03989683464169502, + 0.024415742605924606, + 0.03599795326590538, + 0.020649146288633347, + 0.028720159083604813, + 0.026484306901693344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.07001503556966782 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_9/index_tensor([6090, 6090])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_9/index_tensor([6090, 6090])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..06a4415cd81c0fa37d37ceb7da1314f3d1ce7395 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_9/index_tensor([6090, 6090])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, bird's-eye view, shot vertically from above, centers on a pristine white round plate with a slightly irregular crayon outline around its edge. The plate rests on a light brown tabletop with a distinctly hand-painted wood grain. Scattered haphazardly in the center are a few tiny, dry, light beige, irregular spherical food particles. Suddenly, a drop of crystal-clear water, seemingly animate, falls from outside the frame with pinpoint precision. As it touches one of the particles, a miracle occurs: instead of dissolving, the particle, like an enchanted paper flower, rapidly and magically blooms, transforming into a charming little fish with a lemon-yellow body and sky-blue fins. More drops of water follow, and other particles unfold one after another, transforming into five-pointed stars shimmering with a soft pink glow and a variety of colorful miniature marine creatures. The texture of the entire picture is full of crayon hand-painted brushstrokes, the color saturation is moderate, the tones are warm and dreamy, and there are no harsh shadows under the bright and soft diffuse light. Ultimately, a fairy tale microcosm full of wonder, innocence and infinite imagination is presented on the plate.", + "seed": 25725, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_9/index_6090_seed_25725_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.006818100623786449, + "vlm_logprob": -0.006818100623786449, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.006818100623786449, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009831931442022324 + ], + "opposite_logprob_tokens": [ + -4.88483190536499 + ], + "logprob_reward": -0.009831931442022324, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009831931442022324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17609409987926483, + 0.16349786520004272, + 0.2031017243862152, + 0.10299032926559448, + 0.11117855459451675, + 0.023299118503928185, + 0.09944149106740952, + 0.042188141494989395, + 0.03246380761265755, + 0.0457448847591877 + ], + "credit_map": [ + 0.014514325186610222, + 0.01740291528403759, + 0.008121555671095848, + 0.018113793805241585, + 0.04858766496181488, + 0.03528428077697754, + 0.02313205413520336, + 0.010937516577541828, + 0.008552191779017448, + 0.013264606706798077, + 0.004048686474561691, + 0.022982509806752205, + 0.06490626931190491, + 0.027069304138422012, + 0.018528390675783157, + 0.004145890008658171, + 0.012891465798020363, + 0.015072082169353962, + 0.005403976421803236, + 0.07596582174301147, + 0.02242368645966053, + 0.03619878739118576, + 0.02932588942348957, + 0.005820007994771004, + 0.012082324363291264, + 0.007817533798515797, + 0.0063404180109500885, + 0.01707235351204872, + 0.017705287784337997, + 0.01917547918856144, + 0.012265411205589771, + 0.010531522333621979, + 0.013671448454260826, + 0.01443924754858017, + 0.010317744687199593, + 0.01327542681246996, + 0.01215190626680851, + 0.02879607118666172, + 0.008704867213964462, + 0.009821847081184387, + 0.002475764136761427, + 0.00290810433216393, + 0.001989342737942934, + 0.0037798197008669376, + 0.0026819829363375902, + 0.004975475370883942, + 0.0026277625001966953, + 0.0018608681857585907, + 0.013042010366916656, + 0.011347807943820953, + 0.01023703534156084, + 0.011565175838768482, + 0.006801673211157322, + 0.02881648950278759, + 0.010127313435077667, + 0.00750399474054575, + 0.005301821045577526, + 0.004672843962907791, + 0.0034840283915400505, + 0.009483403526246548, + 0.0033955492544919252, + 0.007231722120195627, + 0.006047364789992571, + 0.002571410732343793, + 0.0039307293482124805, + 0.0037032063119113445, + 0.002924820175394416, + 0.0041323271580040455, + 0.00456342613324523, + 0.005280886311084032, + 0.005009787157177925, + 0.0029186250176280737, + 0.005108345299959183, + 0.010423962026834488, + 0.0026386925019323826, + 0.006723207887262106, + 0.004054352641105652, + 0.00638175942003727, + 0.007203351240605116, + 0.0032112146727740765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.009831931442022324 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052597238682210445 + ], + "opposite_logprob_tokens": [ + -5.2552595138549805 + ], + "logprob_reward": -0.0052597238682210445, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052597238682210445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1482856273651123, + 0.23040543496608734, + 0.11332827061414719, + 0.2231825441122055, + 0.06527575105428696, + 0.040690407156944275, + 0.05012281611561775, + 0.062142789363861084, + 0.015238219872117043, + 0.051328178495168686 + ], + "credit_map": [ + 0.011957324109971523, + 0.02278747968375683, + 0.009593025781214237, + 0.018179139122366905, + 0.03875373676419258, + 0.010617160238325596, + 0.021244486793875694, + 0.015153275802731514, + 0.01200984325259924, + 0.02052190899848938, + 0.005837093107402325, + 0.024972565472126007, + 0.09094911068677902, + 0.03821323812007904, + 0.031900495290756226, + 0.006001172121614218, + 0.009004119783639908, + 0.009307630360126495, + 0.0046468498185276985, + 0.034785620868206024, + 0.01177651435136795, + 0.02295154519379139, + 0.015682987868785858, + 0.005173001904040575, + 0.019990235567092896, + 0.014752708375453949, + 0.007748675066977739, + 0.07165948301553726, + 0.026504764333367348, + 0.05375025421380997, + 0.018277935683727264, + 0.010498492047190666, + 0.007372750900685787, + 0.006237431429326534, + 0.00631405133754015, + 0.006801345851272345, + 0.007202294655144215, + 0.019897814840078354, + 0.005043857265263796, + 0.006406203843653202, + 0.0042033432982862, + 0.005180461797863245, + 0.0033803414553403854, + 0.004933381453156471, + 0.004288021475076675, + 0.010571107268333435, + 0.004994542337954044, + 0.0031392069067806005, + 0.005471915937960148, + 0.005727640353143215, + 0.0037118960171937943, + 0.006654030177742243, + 0.005009079352021217, + 0.013039672747254372, + 0.006469982210546732, + 0.004038597922772169, + 0.00625229487195611, + 0.007120901253074408, + 0.0031630827579647303, + 0.01876964047551155, + 0.0049295504577457905, + 0.005845314357429743, + 0.012020224705338478, + 0.0040417821146547794, + 0.0014064381830394268, + 0.002124796388670802, + 0.0011824386892840266, + 0.0030438879039138556, + 0.0028210999444127083, + 0.0009770628530532122, + 0.0028493793215602636, + 0.0008331168792210519, + 0.007051768247038126, + 0.012129574082791805, + 0.002887221984565258, + 0.0070641557686030865, + 0.005069403443485498, + 0.007939852774143219, + 0.004694829229265451, + 0.004491369239985943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0052597238682210445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864844024181366 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.00864844024181366, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864844024181366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20636005699634552, + 0.2661478519439697, + 0.10861881822347641, + 0.107394739985466, + 0.02973269298672676, + 0.04909050092101097, + 0.04355742037296295, + 0.07056046277284622, + 0.040349699556827545, + 0.07818770408630371 + ], + "credit_map": [ + 0.01148366741836071, + 0.02981971576809883, + 0.010221400298178196, + 0.037817731499671936, + 0.0324065238237381, + 0.014854487031698227, + 0.05353429913520813, + 0.016222218051552773, + 0.012607816606760025, + 0.02686762809753418, + 0.0070655457675457, + 0.03195364400744438, + 0.10275127738714218, + 0.033927321434020996, + 0.04459044337272644, + 0.006384161300957203, + 0.008848448283970356, + 0.009544246830046177, + 0.005177038721740246, + 0.021262021735310555, + 0.018843235448002815, + 0.026818782091140747, + 0.010938762687146664, + 0.007186281494796276, + 0.012591171078383923, + 0.008623824454843998, + 0.006121616344898939, + 0.01038286741822958, + 0.01639905571937561, + 0.028532173484563828, + 0.015721557661890984, + 0.009022473357617855, + 0.0024229472037404776, + 0.003845732193440199, + 0.0015378482639789581, + 0.003713747253641486, + 0.004634723998606205, + 0.006472407840192318, + 0.0048974486999213696, + 0.0022078349720686674, + 0.004661956802010536, + 0.005780478939414024, + 0.004027769900858402, + 0.0060917348600924015, + 0.006750071421265602, + 0.01063257735222578, + 0.007537927012890577, + 0.003607985796406865, + 0.003364439122378826, + 0.00671657919883728, + 0.0029828306287527084, + 0.005794689524918795, + 0.006778276525437832, + 0.005376190412789583, + 0.01059130672365427, + 0.0019531071884557605, + 0.005999252200126648, + 0.00882185623049736, + 0.0036427003797143698, + 0.022743962705135345, + 0.005118285771459341, + 0.006617747247219086, + 0.014026697725057602, + 0.0035899640060961246, + 0.0031524845398962498, + 0.005764761473983526, + 0.002318523358553648, + 0.006960747763514519, + 0.005112461280077696, + 0.004378494806587696, + 0.008836855180561543, + 0.0038253699894994497, + 0.017148876562714577, + 0.01510734111070633, + 0.005259152967482805, + 0.009485883638262749, + 0.00723832193762064, + 0.010029918514192104, + 0.009464222006499767, + 0.004453992936760187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.00864844024181366 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036187181249260902 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036187181249260902, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036187181249260902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12175047397613525, + 0.17511148750782013, + 0.2283332496881485, + 0.2149990051984787, + 0.040002141147851944, + 0.07767688482999802, + 0.04628165066242218, + 0.02824539877474308, + 0.023980513215065002, + 0.04361914098262787 + ], + "credit_map": [ + 0.007611951325088739, + 0.014787913300096989, + 0.005492295138537884, + 0.013657601550221443, + 0.03992117568850517, + 0.0123640401288867, + 0.021122880280017853, + 0.006792609114199877, + 0.00959866028279066, + 0.014227922074496746, + 0.004341358318924904, + 0.02250896580517292, + 0.06785441190004349, + 0.03138832002878189, + 0.019699938595294952, + 0.005491927266120911, + 0.018035076558589935, + 0.014825225807726383, + 0.007763298228383064, + 0.07361285388469696, + 0.021775022149086, + 0.05416637286543846, + 0.028652049601078033, + 0.009503363631665707, + 0.020203879103064537, + 0.013967795297503471, + 0.007416644133627415, + 0.0654529556632042, + 0.022238096222281456, + 0.059274822473526, + 0.015249976888298988, + 0.011194846592843533, + 0.00401321891695261, + 0.004213958978652954, + 0.003477923572063446, + 0.005694074090570211, + 0.006201176438480616, + 0.008396998979151249, + 0.0050647626630961895, + 0.0029400254134088755, + 0.008023815229535103, + 0.007044935133308172, + 0.0040249633602797985, + 0.013951453380286694, + 0.008675592020154, + 0.01978725753724575, + 0.010478947311639786, + 0.00568991806358099, + 0.003997917287051678, + 0.008902398869395256, + 0.002767183119431138, + 0.005746791604906321, + 0.00705692870542407, + 0.007268812041729689, + 0.007791826967149973, + 0.002749793464317918, + 0.0032286185305565596, + 0.0032522857654839754, + 0.002082032384350896, + 0.00753539614379406, + 0.0022244006395339966, + 0.0034632054157555103, + 0.004795928485691547, + 0.0016635312931612134, + 0.00226484052836895, + 0.0037427006755024195, + 0.0019567839335650206, + 0.003289711195975542, + 0.004321493208408356, + 0.002642147010192275, + 0.004215775057673454, + 0.0015470614889636636, + 0.010691430419683456, + 0.00646565854549408, + 0.00329286209307611, + 0.004679184406995773, + 0.004876349121332169, + 0.0037566686514765024, + 0.005115849897265434, + 0.0047411383129656315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.0036187181249260902 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731689441949129 + ], + "opposite_logprob_tokens": [ + -5.0067315101623535 + ], + "logprob_reward": -0.006731689441949129, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731689441949129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16367371380329132, + 0.07747270166873932, + 0.13073331117630005, + 0.3582567274570465, + 0.048514775931835175, + 0.058807503432035446, + 0.029381852596998215, + 0.02592453360557556, + 0.017960520461201668, + 0.08927439153194427 + ], + "credit_map": [ + 0.012265676632523537, + 0.020186418667435646, + 0.005746704060584307, + 0.019009433686733246, + 0.04437510296702385, + 0.028092090040445328, + 0.023539310321211815, + 0.01045896951109171, + 0.0045426697470247746, + 0.008971098810434341, + 0.0024181839544326067, + 0.008543615229427814, + 0.024003012105822563, + 0.013334664516150951, + 0.01305453572422266, + 0.002604922978207469, + 0.012956703081727028, + 0.011177952401340008, + 0.0073042516596615314, + 0.026772979646921158, + 0.014014514163136482, + 0.037338752299547195, + 0.012112848460674286, + 0.00905530620366335, + 0.03308074548840523, + 0.023118574172258377, + 0.011744137853384018, + 0.1106746718287468, + 0.0387125127017498, + 0.09610611200332642, + 0.027824455872178078, + 0.01699552685022354, + 0.004311138764023781, + 0.005805594380944967, + 0.003097588662058115, + 0.010913451202213764, + 0.00526774488389492, + 0.011324183084070683, + 0.004147039260715246, + 0.0036480349954217672, + 0.007020549848675728, + 0.0075048403814435005, + 0.00481465645134449, + 0.007816566154360771, + 0.005942794028669596, + 0.013115273788571358, + 0.0067397612147033215, + 0.005853063426911831, + 0.002683234168216586, + 0.003249893430620432, + 0.0019300909480080009, + 0.0035846042446792126, + 0.0039220647886395454, + 0.007879341952502728, + 0.0038013469893485308, + 0.002331276424229145, + 0.0029322642367333174, + 0.003070808481425047, + 0.002445240505039692, + 0.004269314929842949, + 0.003165471600368619, + 0.003298404859378934, + 0.0037620067596435547, + 0.002981021534651518, + 0.0014458986697718501, + 0.0026150550693273544, + 0.001486005843617022, + 0.0031159138306975365, + 0.0020316848531365395, + 0.0024749201256781816, + 0.00325710978358984, + 0.001533932751044631, + 0.010162482038140297, + 0.030524013563990593, + 0.0038074192125350237, + 0.009300517849624157, + 0.01053636148571968, + 0.004904039669781923, + 0.015904182568192482, + 0.004135375842452049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.006731689441949129 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, bird's-eye view, shot vertically from above, centers on a pristine white round plate with a slightly irregular crayon outline around its edge. The plate rests on a light brown tabletop with a distinctly hand-painted wood grain. Scattered haphazardly in the center are a few tiny, dry, light beige, irregular spherical food particles. Suddenly, a drop of crystal-clear water, seemingly animate, falls from outside the frame with pinpoint precision. As it touches one of the particles, a miracle occurs: instead of dissolving, the particle, like an enchanted paper flower, rapidly and magically blooms, transforming into a charming little fish with a lemon-yellow body and sky-blue fins. More drops of water follow, and other particles unfold one after another, transforming into five-pointed stars shimmering with a soft pink glow and a variety of colorful miniature marine creatures. The texture of the entire picture is full of crayon hand-painted brushstrokes, the color saturation is moderate, the tones are warm and dreamy, and there are no harsh shadows under the bright and soft diffuse light. Ultimately, a fairy tale microcosm full of wonder, innocence and infinite imagination is presented on the plate.", + "seed": 25726, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_9/index_6090_seed_25726_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.005172650422900915, + "vlm_logprob": -0.005172650422900915, + "vlm_yes_no_margin": 5.4, + "vlm_token_credit": -0.005172650422900915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00443332688882947 + ], + "opposite_logprob_tokens": [ + -6.0044331550598145 + ], + "logprob_reward": -0.00443332688882947, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00443332688882947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33730554580688477, + 0.1646348536014557, + 0.15103670954704285, + 0.06970052421092987, + 0.03976862132549286, + 0.029887739568948746, + 0.036486681550741196, + 0.04742858186364174, + 0.05371972545981407, + 0.07003103941679001 + ], + "credit_map": [ + 0.022975455969572067, + 0.09608693420886993, + 0.01975141279399395, + 0.027854658663272858, + 0.11130305379629135, + 0.01784837618470192, + 0.02051514759659767, + 0.020970502868294716, + 0.016889730468392372, + 0.03305477276444435, + 0.009572066366672516, + 0.031885504722595215, + 0.03211177513003349, + 0.012861250899732113, + 0.02033195085823536, + 0.007927794940769672, + 0.013369145803153515, + 0.02469727024435997, + 0.011361783370375633, + 0.023053253069519997, + 0.021584194153547287, + 0.010651294142007828, + 0.039106033742427826, + 0.0072137173265218735, + 0.00478884344920516, + 0.013698408380150795, + 0.003218012163415551, + 0.006201900076121092, + 0.023757290095090866, + 0.004212514031678438, + 0.010834361426532269, + 0.0029891880694776773, + 0.003503829939290881, + 0.010162752121686935, + 0.0020492004696279764, + 0.006044310051947832, + 0.008303253911435604, + 0.001946171629242599, + 0.006376386154443026, + 0.0013827153015881777, + 0.002594288904219866, + 0.006864752620458603, + 0.0025090042036026716, + 0.004108634777367115, + 0.006053997203707695, + 0.0014777597971260548, + 0.004939658101648092, + 0.0013396417489275336, + 0.0036148291546851397, + 0.008078821934759617, + 0.002455645240843296, + 0.0057833073660731316, + 0.00711891008540988, + 0.00211196206510067, + 0.005278187803924084, + 0.0020450165029615164, + 0.00421416899189353, + 0.006362944841384888, + 0.005748317111283541, + 0.007356664631515741, + 0.00780265498906374, + 0.0031047577504068613, + 0.010718446224927902, + 0.0021206261590123177, + 0.004020328633487225, + 0.007041647098958492, + 0.006913062650710344, + 0.011384042911231518, + 0.008738546632230282, + 0.0031989715062081814, + 0.010033980943262577, + 0.002389145316556096, + 0.007627810351550579, + 0.019647520035505295, + 0.006182629615068436, + 0.01180334109812975, + 0.010051758028566837, + 0.004527811426669359, + 0.006982542108744383, + 0.0032076272182166576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00443332688882947 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410274276509881 + ], + "opposite_logprob_tokens": [ + -5.50410270690918 + ], + "logprob_reward": -0.00410274276509881, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410274276509881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45975247025489807, + 0.07926396280527115, + 0.04862721264362335, + 0.11334449052810669, + 0.014899300411343575, + 0.006897103972733021, + 0.010834437794983387, + 0.007896598428487778, + 0.007826054468750954, + 0.25065839290618896 + ], + "credit_map": [ + 0.01580963283777237, + 0.19950656592845917, + 0.0074926214292645454, + 0.019631048664450645, + 0.18256203830242157, + 0.014303038828074932, + 0.015092701651155949, + 0.005354815162718296, + 0.0034978168550878763, + 0.024563081562519073, + 0.0024997154250741005, + 0.010428780689835548, + 0.02516547217965126, + 0.0037847531493753195, + 0.007230239454656839, + 0.0020940981339663267, + 0.0038832109421491623, + 0.008178208954632282, + 0.00297076185233891, + 0.007371249608695507, + 0.00692846579477191, + 0.0029887997079640627, + 0.01321240235120058, + 0.0030941145960241556, + 0.004553944803774357, + 0.04705708846449852, + 0.003475409932434559, + 0.004958818666636944, + 0.042440518736839294, + 0.004064659122377634, + 0.005140619818121195, + 0.0016534430906176567, + 0.0018086658092215657, + 0.002221109578385949, + 0.002211677609011531, + 0.002304411493241787, + 0.0025746775791049004, + 0.0007880220073275268, + 0.002203836338594556, + 0.0007868995307944715, + 0.0006378749967552722, + 0.0010409672977402806, + 0.0006444249302148819, + 0.0013260567793622613, + 0.001238579279743135, + 0.00040758310933597386, + 0.0011632760288193822, + 0.00043834137613885105, + 0.001249385066330433, + 0.0012808492174372077, + 0.0015150833642110229, + 0.001494504394941032, + 0.0018540247110649943, + 0.0008477814844809473, + 0.001720838830806315, + 0.0008719704346731305, + 0.0007169372984208167, + 0.001173163764178753, + 0.0007857735618017614, + 0.0014045011484995484, + 0.0016472249990329146, + 0.00048285198863595724, + 0.0011955545051023364, + 0.0004905906971544027, + 0.0007783783366903663, + 0.0013032013084739447, + 0.00086908636149019, + 0.0010364896152168512, + 0.0014130623312667012, + 0.0005405399715527892, + 0.0013469599653035402, + 0.000538336462341249, + 0.07155709713697433, + 0.05549114570021629, + 0.0251380056142807, + 0.02230427972972393, + 0.030020251870155334, + 0.017223605886101723, + 0.017150754109025, + 0.011773237027227879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.00410274276509881 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742583122104406 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.006742583122104406, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742583122104406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3570590913295746, + 0.06100193411111832, + 0.13959026336669922, + 0.213230162858963, + 0.029609328135848045, + 0.016222676262259483, + 0.0249120332300663, + 0.03692213445901871, + 0.036508020013570786, + 0.08494435250759125 + ], + "credit_map": [ + 0.014340461231768131, + 0.15768782794475555, + 0.0053314403630793095, + 0.014374754391610622, + 0.13787557184696198, + 0.011286851018667221, + 0.011682001873850822, + 0.0044801668263971806, + 0.004326396156102419, + 0.015339069999754429, + 0.0030562300235033035, + 0.009641623124480247, + 0.01538281049579382, + 0.003411990124732256, + 0.007052712608128786, + 0.0027911013457924128, + 0.015992142260074615, + 0.017364846542477608, + 0.014278126880526543, + 0.020328685641288757, + 0.015328548848628998, + 0.011117133311927319, + 0.037675853818655014, + 0.007504926063120365, + 0.008471131324768066, + 0.09032827615737915, + 0.004534629173576832, + 0.008242386393249035, + 0.08150118589401245, + 0.007847304455935955, + 0.00896214134991169, + 0.003343121148645878, + 0.003463961649686098, + 0.004291743505746126, + 0.002867825096473098, + 0.00561808655038476, + 0.005215704441070557, + 0.0012458153069019318, + 0.005421738140285015, + 0.0014844532124698162, + 0.00153946946375072, + 0.0026369374245405197, + 0.0011503745336085558, + 0.00278283329680562, + 0.0029753590933978558, + 0.0008188409265130758, + 0.0035237965639680624, + 0.0007950655417516828, + 0.0017341981874778867, + 0.003497383091598749, + 0.002328934846445918, + 0.0044054207392036915, + 0.00564958481118083, + 0.0011922031408175826, + 0.005094008054584265, + 0.0010103002423420548, + 0.0029610462952405214, + 0.009175688959658146, + 0.0018391086487099528, + 0.004830675199627876, + 0.007466943934559822, + 0.0014482509577646852, + 0.007759647443890572, + 0.001440772321075201, + 0.005522918421775103, + 0.005429090466350317, + 0.0026698708534240723, + 0.0069852471351623535, + 0.005952014587819576, + 0.002118527190759778, + 0.005468109156936407, + 0.002362240571528673, + 0.02370082028210163, + 0.01744704321026802, + 0.008369808085262775, + 0.008170735090970993, + 0.010348369367420673, + 0.005731937009841204, + 0.006968540605157614, + 0.004207103978842497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.006742583122104406 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640998784452677 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004640998784452677, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640998784452677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07375098019838333, + 0.01511005125939846, + 0.0057455128990113735, + 0.043880265206098557, + 0.013877942226827145, + 0.009199691005051136, + 0.012940417975187302, + 0.025370437651872635, + 0.03020358271896839, + 0.7699211239814758 + ], + "credit_map": [ + 0.002835088176652789, + 0.03129193186759949, + 0.0016354683320969343, + 0.0036603217013180256, + 0.027759220451116562, + 0.00245658284984529, + 0.0029284434858709574, + 0.0011839166982099414, + 0.0009418845875188708, + 0.004411427304148674, + 0.000802473456133157, + 0.0015008600894361734, + 0.004402379039674997, + 0.0008899999083951116, + 0.0014644891489297152, + 0.0006965373759157956, + 0.0008474445785395801, + 0.0009022318408824503, + 0.0006561455084010959, + 0.0008032642072066665, + 0.0007555796182714403, + 0.00047479706699959934, + 0.0008852052269503474, + 0.00042084482265636325, + 0.0031362678855657578, + 0.017361391335725784, + 0.0026261082384735346, + 0.0017813941230997443, + 0.013578186742961407, + 0.0014937043888494372, + 0.0026547270826995373, + 0.0012484804028645158, + 0.0008156133117154241, + 0.00274164997972548, + 0.0007882764330133796, + 0.0015826969174668193, + 0.004118981305509806, + 0.0006675738259218633, + 0.0026248826179653406, + 0.0005382673116400838, + 0.0004851251433137804, + 0.0017442565876990557, + 0.0004767795617226511, + 0.00115298037417233, + 0.002840188331902027, + 0.0004662744468078017, + 0.0016924137016758323, + 0.00034167247940786183, + 0.0016601691022515297, + 0.0022739036940038204, + 0.0011635699775069952, + 0.0012285865377634764, + 0.0032005219254642725, + 0.0007373161497525871, + 0.001976760569959879, + 0.0006995892617851496, + 0.0019517255714163184, + 0.0028644613921642303, + 0.0029473856557160616, + 0.005113806109875441, + 0.004603575449436903, + 0.0013418892631307244, + 0.005538084544241428, + 0.001009510480798781, + 0.004559603054076433, + 0.010536820627748966, + 0.0016443801578134298, + 0.0036862895358353853, + 0.004708183463662863, + 0.0013322817394509912, + 0.002597526414319873, + 0.0011384966783225536, + 0.22328993678092957, + 0.17635580897331238, + 0.07873503863811493, + 0.06585431098937988, + 0.0881877914071083, + 0.05164513364434242, + 0.050598252564668655, + 0.0352548323571682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.004640998784452677 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943600554019213 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943600554019213, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943600554019213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06220699101686478, + 0.039281148463487625, + 0.06811802089214325, + 0.32024890184402466, + 0.03318730741739273, + 0.022809134796261787, + 0.03233993053436279, + 0.04716382175683975, + 0.028075212612748146, + 0.3465695381164551 + ], + "credit_map": [ + 0.005298513453453779, + 0.021423552185297012, + 0.003221072955057025, + 0.0041594235226511955, + 0.018156783655285835, + 0.003357807407155633, + 0.0035056564956903458, + 0.0030841820407658815, + 0.0024721103254705667, + 0.012303520925343037, + 0.001895493594929576, + 0.003332902444526553, + 0.011737510561943054, + 0.003428409807384014, + 0.0023319453466683626, + 0.0017792538274079561, + 0.009357594884932041, + 0.008758137002587318, + 0.00849166139960289, + 0.00847199559211731, + 0.0080904271453619, + 0.006601550616323948, + 0.014010530896484852, + 0.004336124751716852, + 0.017171356827020645, + 0.135619655251503, + 0.007129486184567213, + 0.013982000760734081, + 0.10684940218925476, + 0.011259420774877071, + 0.021212512627243996, + 0.007025063503533602, + 0.002321643056347966, + 0.007640197407454252, + 0.0024205869995057583, + 0.005378114990890026, + 0.00713015953078866, + 0.0013945925747975707, + 0.005893935915082693, + 0.0010080798529088497, + 0.0026950626634061337, + 0.0037141924258321524, + 0.0022143495734781027, + 0.002584547735750675, + 0.005309623666107655, + 0.0010419697500765324, + 0.0040723225101828575, + 0.0011770675191655755, + 0.0041380636394023895, + 0.004491456318646669, + 0.005425875540822744, + 0.0038133750203996897, + 0.0063123698346316814, + 0.0018013857770711184, + 0.004565277136862278, + 0.0017921291291713715, + 0.00524273794144392, + 0.007758510764688253, + 0.004675411619246006, + 0.006391487549990416, + 0.007102036848664284, + 0.0024836352095007896, + 0.011630752123892307, + 0.0018792483024299145, + 0.002947180764749646, + 0.005163394846022129, + 0.002839375054463744, + 0.005239882972091436, + 0.004241308197379112, + 0.0015727642457932234, + 0.004433393012732267, + 0.0016379131702706218, + 0.09823641926050186, + 0.08587450534105301, + 0.03045978583395481, + 0.032157402485609055, + 0.03969205543398857, + 0.021989505738019943, + 0.02333700656890869, + 0.014822871424257755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.005943600554019213 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, bird's-eye view, shot vertically from above, centers on a pristine white round plate with a slightly irregular crayon outline around its edge. The plate rests on a light brown tabletop with a distinctly hand-painted wood grain. Scattered haphazardly in the center are a few tiny, dry, light beige, irregular spherical food particles. Suddenly, a drop of crystal-clear water, seemingly animate, falls from outside the frame with pinpoint precision. As it touches one of the particles, a miracle occurs: instead of dissolving, the particle, like an enchanted paper flower, rapidly and magically blooms, transforming into a charming little fish with a lemon-yellow body and sky-blue fins. More drops of water follow, and other particles unfold one after another, transforming into five-pointed stars shimmering with a soft pink glow and a variety of colorful miniature marine creatures. The texture of the entire picture is full of crayon hand-painted brushstrokes, the color saturation is moderate, the tones are warm and dreamy, and there are no harsh shadows under the bright and soft diffuse light. Ultimately, a fairy tale microcosm full of wonder, innocence and infinite imagination is presented on the plate.", + "seed": 25727, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_9/index_6090_seed_25727_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010039116814732551, + "vlm_logprob": -0.010039116814732551, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.010039116814732551, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01045122928917408 + ], + "opposite_logprob_tokens": [ + -4.885451316833496 + ], + "logprob_reward": -0.01045122928917408, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01045122928917408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28925302624702454, + 0.05860383063554764, + 0.04635821282863617, + 0.060077279806137085, + 0.054966703057289124, + 0.05802146717905998, + 0.10974863171577454, + 0.0404006764292717, + 0.18166209757328033, + 0.10090815275907516 + ], + "credit_map": [ + 0.02438141033053398, + 0.041874960064888, + 0.020321812480688095, + 0.07699110358953476, + 0.057811491191387177, + 0.015195777639746666, + 0.03931731358170509, + 0.01335914246737957, + 0.0063073779456317425, + 0.00722175557166338, + 0.004424699582159519, + 0.01622108183801174, + 0.01295686885714531, + 0.002601741347461939, + 0.006606807000935078, + 0.002263499889522791, + 0.005199453327804804, + 0.006691711023449898, + 0.003600833471864462, + 0.0068651046603918076, + 0.01009994838386774, + 0.003013485111296177, + 0.007266673259437084, + 0.0036210045218467712, + 0.009015461429953575, + 0.006799217779189348, + 0.007388869766145945, + 0.009708772413432598, + 0.009742305614054203, + 0.0035218321718275547, + 0.01003188081085682, + 0.0038689388893544674, + 0.005932924337685108, + 0.008292539976537228, + 0.004071768373250961, + 0.007701312657445669, + 0.012170705012977123, + 0.002527195028960705, + 0.012095398269593716, + 0.002174857771024108, + 0.004410525318235159, + 0.006737475749105215, + 0.003788464702665806, + 0.011814509518444538, + 0.015121620148420334, + 0.0020778807811439037, + 0.011921180412173271, + 0.0021498107817023993, + 0.008702765218913555, + 0.012570634484291077, + 0.0063425954431295395, + 0.033637579530477524, + 0.012286283075809479, + 0.005016622133553028, + 0.023469973355531693, + 0.007722184062004089, + 0.00423729233443737, + 0.004492153879255056, + 0.004093409050256014, + 0.007497700396925211, + 0.006612669676542282, + 0.0033409763127565384, + 0.008084658533334732, + 0.0020418185740709305, + 0.036072347313165665, + 0.015117790549993515, + 0.04620371758937836, + 0.019362136721611023, + 0.017216376960277557, + 0.009950037114322186, + 0.02365913987159729, + 0.014080550521612167, + 0.026868321001529694, + 0.007848205044865608, + 0.01641906052827835, + 0.011982615105807781, + 0.00852981861680746, + 0.00834651943296194, + 0.014592720195651054, + 0.006320891436189413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.01045122928917408 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637348957359791 + ], + "opposite_logprob_tokens": [ + -4.8826375007629395 + ], + "logprob_reward": -0.007637348957359791, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637348957359791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1098574697971344, + 0.0448618046939373, + 0.05482551082968712, + 0.05685185268521309, + 0.048610620200634, + 0.03124806471168995, + 0.42933693528175354, + 0.11778930574655533, + 0.049293242394924164, + 0.05732523277401924 + ], + "credit_map": [ + 0.009885789826512337, + 0.01186778862029314, + 0.006881985347718, + 0.018799806013703346, + 0.019291894510388374, + 0.008212901651859283, + 0.024474825710058212, + 0.010442479513585567, + 0.0033956095576286316, + 0.006608809344470501, + 0.0031268487218767405, + 0.011406179517507553, + 0.009595709852874279, + 0.0024738290812820196, + 0.005959226284176111, + 0.0022955925669521093, + 0.005175050813704729, + 0.005303157493472099, + 0.0028991284780204296, + 0.008723157458007336, + 0.010244377888739109, + 0.004296987783163786, + 0.013975257985293865, + 0.004208393394947052, + 0.005394266452640295, + 0.007041319739073515, + 0.004544388502836227, + 0.012207260355353355, + 0.010203083977103233, + 0.004483571741729975, + 0.008338558487594128, + 0.004639406688511372, + 0.00357389566488564, + 0.006174808833748102, + 0.002294292440637946, + 0.007518436294049025, + 0.013042203150689602, + 0.0024841001722961664, + 0.010615355335175991, + 0.002907523652538657, + 0.0024065261241048574, + 0.0032498000655323267, + 0.0017775705782696605, + 0.006930624134838581, + 0.006336976774036884, + 0.00130182015709579, + 0.007372842635959387, + 0.0018719020299613476, + 0.01644899696111679, + 0.027450645342469215, + 0.01040470227599144, + 0.18429034948349, + 0.024978145956993103, + 0.009448510594666004, + 0.1470576971769333, + 0.009257872588932514, + 0.010209291242063046, + 0.028824225068092346, + 0.004873684141784906, + 0.03037143684923649, + 0.01632598415017128, + 0.003960505127906799, + 0.01848292350769043, + 0.004741264972835779, + 0.009105764329433441, + 0.00465783616527915, + 0.008886568248271942, + 0.005633096676319838, + 0.0069753085263073444, + 0.002835908904671669, + 0.008435575291514397, + 0.002763185417279601, + 0.010129258036613464, + 0.00545918894931674, + 0.00762215768918395, + 0.007438362576067448, + 0.0069167157635092735, + 0.004211150109767914, + 0.011583437211811543, + 0.003964961506426334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.007637348957359791 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789316914975643 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.009789316914975643, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789316914975643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15736925601959229, + 0.1130829006433487, + 0.06700412184000015, + 0.0738406702876091, + 0.08442480117082596, + 0.1504526287317276, + 0.11348461359739304, + 0.030106445774435997, + 0.11716556549072266, + 0.09306901693344116 + ], + "credit_map": [ + 0.009649170562624931, + 0.021037323400378227, + 0.0069990199990570545, + 0.023405419662594795, + 0.032586827874183655, + 0.008632288314402103, + 0.04377022758126259, + 0.01128897350281477, + 0.012009055353701115, + 0.012203958816826344, + 0.006442314945161343, + 0.016995839774608612, + 0.018736617639660835, + 0.00979877170175314, + 0.025845317170023918, + 0.011051016859710217, + 0.006895441096276045, + 0.006197384558618069, + 0.004650476388633251, + 0.009581197053194046, + 0.010583920404314995, + 0.009008800610899925, + 0.01006079651415348, + 0.010026093572378159, + 0.00753116374835372, + 0.009168810211122036, + 0.0044924793764948845, + 0.0118169030174613, + 0.015360836870968342, + 0.003883417695760727, + 0.016942977905273438, + 0.004644076805561781, + 0.007013679947704077, + 0.012251368723809719, + 0.00516158901154995, + 0.010646133683621883, + 0.022125152871012688, + 0.0037394315004348755, + 0.019718553870916367, + 0.0037688883021473885, + 0.011545316316187382, + 0.014681369997560978, + 0.012223458848893642, + 0.033107515424489975, + 0.03441591560840607, + 0.005435026716440916, + 0.03325458988547325, + 0.005789424758404493, + 0.005666980519890785, + 0.007120607420802116, + 0.004166812635958195, + 0.046045172959566116, + 0.00986475683748722, + 0.002522073220461607, + 0.035288743674755096, + 0.0028094640001654625, + 0.0031659388914704323, + 0.0042975596152246, + 0.00288307573646307, + 0.0042422618716955185, + 0.003621942363679409, + 0.0025552157312631607, + 0.006963120307773352, + 0.0023773303255438805, + 0.033061958849430084, + 0.00951873417943716, + 0.02372562326490879, + 0.01035371981561184, + 0.01312177162617445, + 0.005332853179425001, + 0.015581601299345493, + 0.006469287909567356, + 0.01814158447086811, + 0.00645713834092021, + 0.01315989252179861, + 0.012789960019290447, + 0.01069337222725153, + 0.005230476148426533, + 0.02066553384065628, + 0.0059310561046004295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.009789316914975643 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541357427835464 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541357427835464, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541357427835464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08356747031211853, + 0.02637362666428089, + 0.04088220372796059, + 0.031056933104991913, + 0.04777585342526436, + 0.1318274736404419, + 0.24379047751426697, + 0.08783991634845734, + 0.1508113145828247, + 0.15607474744319916 + ], + "credit_map": [ + 0.0065977382473647594, + 0.009298867546021938, + 0.006353544071316719, + 0.013049041852355003, + 0.014563939534127712, + 0.00539356516674161, + 0.021203327924013138, + 0.007107442244887352, + 0.002630701055750251, + 0.004169886000454426, + 0.0021821577101945877, + 0.005162715911865234, + 0.005100433714687824, + 0.0011272078845649958, + 0.004870567470788956, + 0.0011299579637125134, + 0.003079068847000599, + 0.00647186441347003, + 0.0028930706903338432, + 0.0066123888827860355, + 0.008833708241581917, + 0.00264107808470726, + 0.007583734579384327, + 0.002767289290204644, + 0.002526570111513138, + 0.004458248615264893, + 0.0019506927346810699, + 0.00821427907794714, + 0.006041225977241993, + 0.0016780184814706445, + 0.004218113608658314, + 0.001969784963876009, + 0.006660209037363529, + 0.00618935190141201, + 0.004173386842012405, + 0.00498571153730154, + 0.011761706322431564, + 0.0021760070230811834, + 0.008936834521591663, + 0.0028926508966833353, + 0.01096179150044918, + 0.012654741294682026, + 0.011986465193331242, + 0.02760442905128002, + 0.028843089938163757, + 0.00559368496760726, + 0.02859547547996044, + 0.005587805062532425, + 0.01023532822728157, + 0.015768742188811302, + 0.0075918277725577354, + 0.10099374502897263, + 0.013584692031145096, + 0.006153857335448265, + 0.08384152501821518, + 0.005620760843157768, + 0.006862394977360964, + 0.02308664843440056, + 0.004242321010679007, + 0.021681219339370728, + 0.010491271503269672, + 0.0031908652745187283, + 0.014838146045804024, + 0.0034470499958842993, + 0.025891084223985672, + 0.01728806272149086, + 0.024506257846951485, + 0.017891528084874153, + 0.026478363201022148, + 0.008575472980737686, + 0.02020922489464283, + 0.00997131410986185, + 0.047861695289611816, + 0.011664229445159435, + 0.034770697355270386, + 0.014888775534927845, + 0.015494190156459808, + 0.00829681009054184, + 0.014500492252409458, + 0.008597856387495995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.012541357427835464 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977633148431778 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.00977633148431778, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977633148431778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11196739226579666, + 0.02875175140798092, + 0.02232089266180992, + 0.03270641714334488, + 0.030270949006080627, + 0.02797183208167553, + 0.5097202062606812, + 0.04379913583397865, + 0.08945107460021973, + 0.10304033011198044 + ], + "credit_map": [ + 0.009880066849291325, + 0.015091418288648129, + 0.010495618917047977, + 0.020533431321382523, + 0.017687123268842697, + 0.009125464595854282, + 0.017706509679555893, + 0.011447766795754433, + 0.003021512646228075, + 0.004119460005313158, + 0.0026401763316243887, + 0.005831771530210972, + 0.005724870599806309, + 0.0018552429974079132, + 0.00427655503153801, + 0.0012821657583117485, + 0.0019733484368771315, + 0.0030507410410791636, + 0.001281740958802402, + 0.003628162667155266, + 0.0039848629385232925, + 0.0018295933259651065, + 0.00423255143687129, + 0.0023398923221975565, + 0.004455334506928921, + 0.003445704933255911, + 0.0024248147383332253, + 0.004496777895838022, + 0.004963194951415062, + 0.0035715969279408455, + 0.005157319828867912, + 0.0041916752234101295, + 0.0030210562981665134, + 0.005667827557772398, + 0.0018861307762563229, + 0.0037581322249025106, + 0.005564399994909763, + 0.001835015369579196, + 0.0062595875933766365, + 0.0022787998896092176, + 0.002898257924243808, + 0.0039062947034835815, + 0.0033706077374517918, + 0.004456563852727413, + 0.0053642624989151955, + 0.0012140482431277633, + 0.005378128495067358, + 0.001383668277412653, + 0.019135184586048126, + 0.031124526634812355, + 0.012018401175737381, + 0.22036126255989075, + 0.029711296781897545, + 0.011351102031767368, + 0.17605257034301758, + 0.009965878911316395, + 0.004837244749069214, + 0.008311078883707523, + 0.003468769136816263, + 0.006796251051127911, + 0.0055018640123307705, + 0.0023258684668689966, + 0.009185231290757656, + 0.0033728282433003187, + 0.020342860370874405, + 0.007153209764510393, + 0.018601302057504654, + 0.011421320959925652, + 0.01090551633387804, + 0.004164653830230236, + 0.012501777149736881, + 0.004360429011285305, + 0.027231868356466293, + 0.007382053416222334, + 0.02346440777182579, + 0.010968407616019249, + 0.00818935688585043, + 0.005962077062577009, + 0.012791501358151436, + 0.007050652988255024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.00977633148431778 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, bird's-eye view, shot vertically from above, centers on a pristine white round plate with a slightly irregular crayon outline around its edge. The plate rests on a light brown tabletop with a distinctly hand-painted wood grain. Scattered haphazardly in the center are a few tiny, dry, light beige, irregular spherical food particles. Suddenly, a drop of crystal-clear water, seemingly animate, falls from outside the frame with pinpoint precision. As it touches one of the particles, a miracle occurs: instead of dissolving, the particle, like an enchanted paper flower, rapidly and magically blooms, transforming into a charming little fish with a lemon-yellow body and sky-blue fins. More drops of water follow, and other particles unfold one after another, transforming into five-pointed stars shimmering with a soft pink glow and a variety of colorful miniature marine creatures. The texture of the entire picture is full of crayon hand-painted brushstrokes, the color saturation is moderate, the tones are warm and dreamy, and there are no harsh shadows under the bright and soft diffuse light. Ultimately, a fairy tale microcosm full of wonder, innocence and infinite imagination is presented on the plate.", + "seed": 25728, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000138/group_9/index_6090_seed_25728_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.008703501801937819, + "vlm_logprob": -0.008703501801937819, + "vlm_yes_no_margin": 4.924999952316284, + "vlm_token_credit": -0.008703501801937819, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006585561670362949 + ], + "opposite_logprob_tokens": [ + -5.381585597991943 + ], + "logprob_reward": -0.006585561670362949, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.006585561670362949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12445644289255142, + 0.017293693497776985, + 0.018221130594611168, + 0.028590230271220207, + 0.050334617495536804, + 0.46530720591545105, + 0.11954756826162338, + 0.09885234385728836, + 0.027543669566512108, + 0.0498531237244606 + ], + "credit_map": [ + 0.012413865886628628, + 0.01607057824730873, + 0.007753093261271715, + 0.018844544887542725, + 0.020838694646954536, + 0.029096923768520355, + 0.01267864741384983, + 0.006760094780474901, + 0.0017181320581585169, + 0.002333722310140729, + 0.0008226074278354645, + 0.00351630337536335, + 0.0028801215812563896, + 0.0031099857296794653, + 0.002186625264585018, + 0.0007261959835886955, + 0.0016822275938466191, + 0.002443174133077264, + 0.001161509077064693, + 0.0025373396929353476, + 0.004147364757955074, + 0.0031961945351213217, + 0.0021101715974509716, + 0.0009431483340449631, + 0.0020939225796610117, + 0.005177395883947611, + 0.0019068511901423335, + 0.00489460164681077, + 0.0066744606010615826, + 0.002550044795498252, + 0.003895603120326996, + 0.0013973499881103635, + 0.004061468411237001, + 0.008252347819507122, + 0.003622261341661215, + 0.004405843559652567, + 0.013122692704200745, + 0.00970970094203949, + 0.0047464799135923386, + 0.002413819544017315, + 0.02324969321489334, + 0.05606774240732193, + 0.01974332146346569, + 0.1099456176161766, + 0.09217112511396408, + 0.026653718203306198, + 0.1219792366027832, + 0.01549672894179821, + 0.03138493373990059, + 0.008170166052877903, + 0.007956752553582191, + 0.04090793430805206, + 0.011343371123075485, + 0.0066734300926327705, + 0.010140275582671165, + 0.0029706922359764576, + 0.0219771359115839, + 0.0059631806798279285, + 0.00721018249168992, + 0.009164161048829556, + 0.005143157206475735, + 0.041733428835868835, + 0.0033585079945623875, + 0.004302593879401684, + 0.007837377488613129, + 0.0017984320875257254, + 0.0013800155138596892, + 0.0015799248358234763, + 0.001914653810672462, + 0.01070697046816349, + 0.0012755722273141146, + 0.0010507230181246996, + 0.013179393485188484, + 0.0034935232251882553, + 0.005375640466809273, + 0.007415072526782751, + 0.0028368227649480104, + 0.010245364159345627, + 0.004586904775351286, + 0.0027204041834920645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.006585561670362949 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260791163891554 + ], + "opposite_logprob_tokens": [ + -5.255260944366455 + ], + "logprob_reward": -0.005260791163891554, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260791163891554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12118265777826309, + 0.03340334817767143, + 0.040394168347120285, + 0.10302335768938065, + 0.16919665038585663, + 0.20717325806617737, + 0.0325656495988369, + 0.12190593779087067, + 0.10664084553718567, + 0.06451410055160522 + ], + "credit_map": [ + 0.01181024219840765, + 0.014244014397263527, + 0.007330436259508133, + 0.016428863629698753, + 0.020495858043432236, + 0.028902003541588783, + 0.01528187096118927, + 0.006689365953207016, + 0.0033517146948724985, + 0.005187456030398607, + 0.00275762309320271, + 0.004389740526676178, + 0.0062861195765435696, + 0.0038493825122714043, + 0.005635326262563467, + 0.001945985946804285, + 0.0038018152117729187, + 0.005056887865066528, + 0.0025297366082668304, + 0.006881246343255043, + 0.00799570046365261, + 0.0083170086145401, + 0.0038022189401090145, + 0.0020095533691346645, + 0.00756010040640831, + 0.016726236790418625, + 0.005425788927823305, + 0.02557639218866825, + 0.016408061608672142, + 0.009888443164527416, + 0.016591060906648636, + 0.004847269505262375, + 0.016064735129475594, + 0.021151227876544, + 0.021444756537675858, + 0.02054002694785595, + 0.022701507434248924, + 0.036244459450244904, + 0.017376674339175224, + 0.013673263601958752, + 0.013734832406044006, + 0.02275647222995758, + 0.010540424846112728, + 0.04469575732946396, + 0.04088059067726135, + 0.01712079346179962, + 0.04796654358506203, + 0.009477837942540646, + 0.002967319218441844, + 0.003204854903742671, + 0.003942802548408508, + 0.00834872480481863, + 0.004084839019924402, + 0.004770190920680761, + 0.0037334985099732876, + 0.0015134193236008286, + 0.03116041235625744, + 0.007162152323871851, + 0.006519665941596031, + 0.01004968024790287, + 0.007373474538326263, + 0.04981202259659767, + 0.004998811054974794, + 0.004829720593988895, + 0.03984711319208145, + 0.005372209474444389, + 0.0044968402944505215, + 0.0036755993496626616, + 0.003978246357291937, + 0.04070675000548363, + 0.004917527548968792, + 0.0036465483717620373, + 0.007708575110882521, + 0.005805664695799351, + 0.0057876817882061005, + 0.014243205077946186, + 0.006235748063772917, + 0.007240196689963341, + 0.013181318528950214, + 0.0043117147870361805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.005260791163891554 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018198659643530846 + ], + "opposite_logprob_tokens": [ + -4.018198490142822 + ], + "logprob_reward": -0.018198659643530846, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018198659643530846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0635857880115509, + 0.05274828523397446, + 0.03000369295477867, + 0.01522267796099186, + 0.0961630791425705, + 0.5045022368431091, + 0.08054708689451218, + 0.05795954167842865, + 0.06977038085460663, + 0.029497254639863968 + ], + "credit_map": [ + 0.005486755166202784, + 0.007550393231213093, + 0.0056034573353827, + 0.00996953435242176, + 0.012517568655312061, + 0.0066518415696918964, + 0.012856251560151577, + 0.0029499917291104794, + 0.0026592868380248547, + 0.004975034389644861, + 0.0030276488978415728, + 0.012412055395543575, + 0.015835629776120186, + 0.00555027648806572, + 0.006782846990972757, + 0.0015055133262649179, + 0.0020711026154458523, + 0.00410416116937995, + 0.0014484396670013666, + 0.004234710708260536, + 0.006103844847530127, + 0.004787478130310774, + 0.0052632433362305164, + 0.0019907141104340553, + 0.0009000870632007718, + 0.002343856729567051, + 0.0006284770206548274, + 0.003583932528272271, + 0.003951702732592821, + 0.0015003717271611094, + 0.001601151074282825, + 0.0007130996673367918, + 0.008846985176205635, + 0.013532881624996662, + 0.01067393273115158, + 0.00896365288645029, + 0.017967749387025833, + 0.022512393072247505, + 0.006714927032589912, + 0.006950560957193375, + 0.026576096192002296, + 0.060449372977018356, + 0.02329813502728939, + 0.1167484000325203, + 0.1002136766910553, + 0.03135256469249725, + 0.12790127098560333, + 0.017962737008929253, + 0.02893504500389099, + 0.004964279942214489, + 0.006024910137057304, + 0.011038661003112793, + 0.004238447640091181, + 0.017495255917310715, + 0.005108903627842665, + 0.002741585485637188, + 0.013292074203491211, + 0.0034167668782174587, + 0.00381892966106534, + 0.004637644626200199, + 0.0036963168531656265, + 0.024364715442061424, + 0.002327987691387534, + 0.0024051074869930744, + 0.025562796741724014, + 0.0034491096157580614, + 0.0024763585533946753, + 0.0020683249458670616, + 0.003181258449330926, + 0.028273697942495346, + 0.002186889061704278, + 0.0025719464756548405, + 0.003945818170905113, + 0.0031865776982158422, + 0.0017446315614506602, + 0.004690566565841436, + 0.0034034650307148695, + 0.0033690594136714935, + 0.006761468015611172, + 0.002395669696852565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.018198659643530846 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741398945450783 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741398945450783, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741398945450783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07716092467308044, + 0.014026603661477566, + 0.025080200284719467, + 0.03403723984956741, + 0.09107893705368042, + 0.5103608965873718, + 0.07972648739814758, + 0.07850869745016098, + 0.029091253876686096, + 0.060928720980882645 + ], + "credit_map": [ + 0.008363031782209873, + 0.01131722517311573, + 0.004906302783638239, + 0.010763835161924362, + 0.011947479099035263, + 0.010648056864738464, + 0.014038396067917347, + 0.005176595877856016, + 0.001567981205880642, + 0.0015285441186279058, + 0.00116433494258672, + 0.0024228934198617935, + 0.0027206598315387964, + 0.0018147578230127692, + 0.002117146970704198, + 0.0006902858149260283, + 0.002150762127712369, + 0.002999431686475873, + 0.001377276610583067, + 0.004370346199721098, + 0.004964888095855713, + 0.0055021923035383224, + 0.0022243403363972902, + 0.0014909638557583094, + 0.001851961831562221, + 0.004005088936537504, + 0.0018119517480954528, + 0.007886943407356739, + 0.009221939370036125, + 0.0020130935590714216, + 0.006011858116835356, + 0.0012344017159193754, + 0.008227388374507427, + 0.013096191920340061, + 0.009914398193359375, + 0.007422313094139099, + 0.017786679789423943, + 0.021564437076449394, + 0.0067449831403791904, + 0.00632254546508193, + 0.027601653710007668, + 0.059982091188430786, + 0.024289457127451897, + 0.11810725182294846, + 0.09985541552305222, + 0.032787222415208817, + 0.12920533120632172, + 0.018532516434788704, + 0.004729572217911482, + 0.011047737672924995, + 0.004066880792379379, + 0.02559739351272583, + 0.009766235016286373, + 0.01183411292731762, + 0.010329234413802624, + 0.002355316886678338, + 0.015624329447746277, + 0.005285119172185659, + 0.006042290944606066, + 0.007989580743014812, + 0.004737363196909428, + 0.031624142080545425, + 0.003667289623990655, + 0.0035385857336223125, + 0.008737790398299694, + 0.0027259483467787504, + 0.002218791050836444, + 0.003131355857476592, + 0.0032026474364101887, + 0.004535338841378689, + 0.003562877420336008, + 0.0009765029535628855, + 0.007995671592652798, + 0.006356187164783478, + 0.004250894766300917, + 0.013728306628763676, + 0.005106165539473295, + 0.008351076394319534, + 0.01120150089263916, + 0.003938915673643351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.006741398945450783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static bird's-eye view shot vertically from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a pristine white round plate with a slightly irregular crayon outline around its edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do tiny, dry, light beige, irregular spherical food particles scatter in the center of the plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles transform into objects like a lemon-yellow fish with sky-blue fins and five-pointed stars shimmering with a soft pink glow upon contact with water drops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire image rendered with a texture of crayon hand-painted brushstrokes, moderate color saturation, warm tones, and no harsh shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731097586452961 + ], + "opposite_logprob_tokens": [ + -5.006731033325195 + ], + "logprob_reward": -0.006731097586452961, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731097586452961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04784652218222618, + 0.032487571239471436, + 0.023657485842704773, + 0.0380685031414032, + 0.21752198040485382, + 0.24806992709636688, + 0.07799893617630005, + 0.2354421466588974, + 0.02965216338634491, + 0.049254775047302246 + ], + "credit_map": [ + 0.003978061489760876, + 0.008353333920240402, + 0.0037886665668338537, + 0.007527749054133892, + 0.007749006152153015, + 0.004104007501155138, + 0.009432610124349594, + 0.0029130924958735704, + 0.00228732917457819, + 0.0037656803615391254, + 0.0015585829969495535, + 0.007454870268702507, + 0.009592597372829914, + 0.0023526980075985193, + 0.004351286217570305, + 0.001124528469517827, + 0.0017510843463242054, + 0.00315927155315876, + 0.001789469737559557, + 0.0037730105686932802, + 0.006621865089982748, + 0.002037572441622615, + 0.0036971219815313816, + 0.0008280891343019903, + 0.0019766578916460276, + 0.004600520245730877, + 0.001936233718879521, + 0.0097350487485528, + 0.009383038617670536, + 0.0017131848726421595, + 0.007402866613119841, + 0.0013209523167461157, + 0.02004833146929741, + 0.030917730182409286, + 0.0243059191852808, + 0.01819482631981373, + 0.04240703955292702, + 0.05157787725329399, + 0.014756964519619942, + 0.015313278883695602, + 0.013128360733389854, + 0.030153611674904823, + 0.011953264474868774, + 0.05629849061369896, + 0.0496961809694767, + 0.015139782801270485, + 0.06298477202653885, + 0.008715472184121609, + 0.005019292701035738, + 0.00887057650834322, + 0.003664646064862609, + 0.010387420654296875, + 0.011861467733979225, + 0.0067031350918114185, + 0.027425222098827362, + 0.0040671732276678085, + 0.04755531996488571, + 0.01401958055794239, + 0.019125552847981453, + 0.02309434488415718, + 0.012820407748222351, + 0.09934542328119278, + 0.008985526859760284, + 0.010495987720787525, + 0.007103567942976952, + 0.0021129415836185217, + 0.0015630940906703472, + 0.0026751316618174314, + 0.0025423183105885983, + 0.010281886905431747, + 0.0022473004646599293, + 0.001125923590734601, + 0.017420165240764618, + 0.0030301299411803484, + 0.006159466225653887, + 0.003578219562768936, + 0.0024631076958030462, + 0.011646869592368603, + 0.0023417058400809765, + 0.002615112578496337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.006731097586452961 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_0/index_tensor([2623, 2623])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_0/index_tensor([2623, 2623])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f9cb33edddb3d4626bcf86b440a7bd94890c3911 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_0/index_tensor([2623, 2623])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up shot from a slight overhead angle focuses precisely on a palm-sized, uncut, angular piece of clear quartz crystal, delicately held by a pair of slender, fair hands. A bright, focused beam of light streams in from outside the frame, penetrating the crystal's crystalline interior. Its naturally formed ice cracks and tiny bubbles are instantly illuminated by the light, which is then refracted and fragmented by countless sharp facets, casting a brilliant, rainbow-like spectrum of light that softly falls on the fingertips and below the crystal. The hands hold the crystal with steady composure, their nails neatly manicured, and the texture of the skin is clearly visible in the soft light. The surroundings are enveloped in a shallow depth of field, completely blurred by the soft focus. Only faintly discernible scatterings of warm bokeh light create a subtle contrast with the cool spectrum of light cast by the crystal. The entire image, within the static composition, is imbued with a dynamic play of light and shadow, creating a mysterious atmosphere of tranquility, concentration, and a sense of wondrous discovery.", + "seed": 17235, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_0/index_2623_seed_17235_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15025158124044538, + "vlm_logprob": -0.15025158124044538, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.15025158124044538, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4420447051525116 + ], + "opposite_logprob_tokens": [ + -1.067044734954834 + ], + "logprob_reward": -0.4420447051525116, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4420447051525116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26315176486968994, + 0.059853993356227875, + 0.09619132429361343, + 0.2551835775375366, + 0.038839191198349, + 0.1384226232767105, + 0.04119059443473816, + 0.025843946263194084, + 0.021845530718564987, + 0.05947748199105263 + ], + "credit_map": [ + 0.02518070489168167, + 0.0201759971678257, + 0.0384976789355278, + 0.029527055099606514, + 0.022952217608690262, + 0.044278547167778015, + 0.03863583505153656, + 0.043903741985559464, + 0.004838720895349979, + 0.004525964148342609, + 0.005415074061602354, + 0.007305535487830639, + 0.0046256715431809425, + 0.01155475340783596, + 0.006913323886692524, + 0.014674951322376728, + 0.012011963874101639, + 0.006476207170635462, + 0.008714262396097183, + 0.007677496410906315, + 0.01092921756207943, + 0.014316383749246597, + 0.013745235279202461, + 0.02232055366039276, + 0.025878241285681725, + 0.018699156120419502, + 0.018434200435876846, + 0.0253224428743124, + 0.027530750259757042, + 0.040210187435150146, + 0.040144894272089005, + 0.05896371230483055, + 0.004258188419044018, + 0.0038095528725534678, + 0.0030967481434345245, + 0.0046511744149029255, + 0.003946736454963684, + 0.006392308045178652, + 0.004607770126312971, + 0.008076709695160389, + 0.014893852174282074, + 0.010295956395566463, + 0.01942894607782364, + 0.01765602082014084, + 0.014919374138116837, + 0.021932117640972137, + 0.022275499999523163, + 0.017020845785737038, + 0.004203744698315859, + 0.003414608770981431, + 0.004829762037843466, + 0.004310324788093567, + 0.005541116464883089, + 0.005864239763468504, + 0.006748249288648367, + 0.006278547924011946, + 0.0026124659925699234, + 0.00221996009349823, + 0.004917207174003124, + 0.0024695589672774076, + 0.002289197640493512, + 0.0036389371380209923, + 0.0033849149476736784, + 0.004311701748520136, + 0.0014552323846146464, + 0.0010479140328243375, + 0.0011820420622825623, + 0.001841263147071004, + 0.002599941333755851, + 0.005294449627399445, + 0.00257843523286283, + 0.005846254527568817, + 0.0019839887972921133, + 0.0022461283951997757, + 0.0028998716734349728, + 0.0032506296411156654, + 0.0031035076826810837, + 0.021796945482492447, + 0.0026460622902959585, + 0.02155034802854061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.4420447051525116 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007659234572201967 + ], + "opposite_logprob_tokens": [ + -4.882659435272217 + ], + "logprob_reward": -0.007659234572201967, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007659234572201967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24600258469581604, + 0.087879978120327, + 0.04175170511007309, + 0.04407436028122902, + 0.05743555724620819, + 0.06332522630691528, + 0.045784905552864075, + 0.06146756932139397, + 0.05052283778786659, + 0.3017553687095642 + ], + "credit_map": [ + 0.025698117911815643, + 0.020175397396087646, + 0.03669507056474686, + 0.031306494027376175, + 0.025870976969599724, + 0.03532356768846512, + 0.03026113100349903, + 0.04067181795835495, + 0.007124024908989668, + 0.0075083631090819836, + 0.006208875682204962, + 0.010896227322518826, + 0.005380389746278524, + 0.014977737329900265, + 0.014463288709521294, + 0.021321071311831474, + 0.0036698358599096537, + 0.004075523000210524, + 0.003125885734334588, + 0.004179620649665594, + 0.0030278407502919436, + 0.008503210730850697, + 0.0032765951473265886, + 0.011893192306160927, + 0.004818553104996681, + 0.0049278452061116695, + 0.0038425566162914038, + 0.004983578342944384, + 0.0046638501808047295, + 0.006540438625961542, + 0.006552751641720533, + 0.007744784001260996, + 0.007091273088008165, + 0.00513859186321497, + 0.0061932834796607494, + 0.006178231909871101, + 0.005179491825401783, + 0.009053894318640232, + 0.009209015406668186, + 0.00939177442342043, + 0.010831975378096104, + 0.005482018925249577, + 0.0071798465214669704, + 0.008922947570681572, + 0.005821825470775366, + 0.009151521138846874, + 0.007863651029765606, + 0.008071428164839745, + 0.003299660747870803, + 0.003900277428328991, + 0.003308353014290333, + 0.005601737182587385, + 0.004761779215186834, + 0.007924848236143589, + 0.006910459604114294, + 0.010077785700559616, + 0.0075977761298418045, + 0.004561459645628929, + 0.016151586547493935, + 0.005974597297608852, + 0.004211666528135538, + 0.008545377291738987, + 0.004498486872762442, + 0.009926613420248032, + 0.009172477759420872, + 0.0025097685866057873, + 0.005746747367084026, + 0.003409910248592496, + 0.0024748891592025757, + 0.010197029449045658, + 0.004739837720990181, + 0.012272173538804054, + 0.008720713667571545, + 0.00672849640250206, + 0.008806388825178146, + 0.010347500443458557, + 0.010068180039525032, + 0.12188073992729187, + 0.012959081679582596, + 0.12224423140287399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.007659234572201967 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.281402051448822 + ], + "opposite_logprob_tokens": [ + -1.4064021110534668 + ], + "logprob_reward": -0.281402051448822, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.281402051448822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10552110522985458, + 0.07155688107013702, + 0.04509485512971878, + 0.06980478018522263, + 0.20294253528118134, + 0.28341132402420044, + 0.09605405479669571, + 0.07732166349887848, + 0.031120501458644867, + 0.017172224819660187 + ], + "credit_map": [ + 0.009263328276574612, + 0.007840692065656185, + 0.013662809506058693, + 0.012604350224137306, + 0.009254619479179382, + 0.01737053319811821, + 0.016909269616007805, + 0.01861550286412239, + 0.005900049582123756, + 0.004497510846704245, + 0.005826523993164301, + 0.00913646537810564, + 0.007617359980940819, + 0.014781750738620758, + 0.009761989116668701, + 0.014035244472324848, + 0.0032122130505740643, + 0.002421265933662653, + 0.0044532413594424725, + 0.00415431335568428, + 0.005387116689234972, + 0.010648377239704132, + 0.005839343648403883, + 0.008978987112641335, + 0.006472816225141287, + 0.007450588047504425, + 0.004720036406069994, + 0.00805224571377039, + 0.009889925830066204, + 0.009665747173130512, + 0.011780621483922005, + 0.011772805824875832, + 0.016649024561047554, + 0.015413854271173477, + 0.02492191083729267, + 0.017699647694826126, + 0.02161291427910328, + 0.03299221768975258, + 0.028485577553510666, + 0.04516741633415222, + 0.020496947690844536, + 0.021991988644003868, + 0.030558934435248375, + 0.03999888151884079, + 0.03450249880552292, + 0.0477587915956974, + 0.049439530819654465, + 0.038663774728775024, + 0.011210360564291477, + 0.008630934171378613, + 0.0076169464737176895, + 0.010099130682647228, + 0.013849074020981789, + 0.014531918801367283, + 0.01388245914131403, + 0.016233237460255623, + 0.010845706798136234, + 0.0066250464878976345, + 0.01938859187066555, + 0.00773684261366725, + 0.005380533169955015, + 0.010624995455145836, + 0.004558761138468981, + 0.012161189690232277, + 0.0020491313189268112, + 0.0020607158076018095, + 0.002553348196670413, + 0.002249999437481165, + 0.004424266517162323, + 0.007106523495167494, + 0.002949933987110853, + 0.007726585492491722, + 0.0012756016803905368, + 0.0012983412016183138, + 0.0016415981808677316, + 0.002329522743821144, + 0.002998323878273368, + 0.002737395465373993, + 0.0028512657154351473, + 0.0020401780493557453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.281402051448822 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954147316515446 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.005954147316515446, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954147316515446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11068344861268997, + 0.07327315956354141, + 0.02978469245135784, + 0.046016741544008255, + 0.06040589138865471, + 0.1100790947675705, + 0.06482502818107605, + 0.02216201275587082, + 0.06548323482275009, + 0.4172866642475128 + ], + "credit_map": [ + 0.0089168231934309, + 0.01230114046484232, + 0.01070056390017271, + 0.012390026822686195, + 0.012335014529526234, + 0.016763420775532722, + 0.0216966699808836, + 0.015579808503389359, + 0.004260758403688669, + 0.005025479476898909, + 0.006431452464312315, + 0.007893645204603672, + 0.008064397610723972, + 0.014181689359247684, + 0.010096273384988308, + 0.017319470643997192, + 0.0029038935899734497, + 0.0018819165416061878, + 0.003368839155882597, + 0.0032132421620190144, + 0.002325358334928751, + 0.0055242739617824554, + 0.003675385843962431, + 0.006891787052154541, + 0.00528393778949976, + 0.004698564764112234, + 0.0035570794716477394, + 0.0050320313312113285, + 0.004638960584998131, + 0.006733911577612162, + 0.005631661508232355, + 0.010440601035952568, + 0.009389117360115051, + 0.006565764546394348, + 0.008971483446657658, + 0.005988887045532465, + 0.00534756388515234, + 0.007775302045047283, + 0.008186143822968006, + 0.008181638084352016, + 0.011761846020817757, + 0.008781665936112404, + 0.01379149779677391, + 0.014182169921696186, + 0.012626312673091888, + 0.017575295642018318, + 0.01736212708055973, + 0.01399819366633892, + 0.004299988970160484, + 0.005535445176064968, + 0.004865711554884911, + 0.008318700827658176, + 0.009622746147215366, + 0.009767943061888218, + 0.010357451625168324, + 0.012057050131261349, + 0.0028195392806082964, + 0.0018449238268658519, + 0.004788270220160484, + 0.0028186666313558817, + 0.0022955643944442272, + 0.002758270362392068, + 0.0020594941452145576, + 0.002777284709736705, + 0.005307076498866081, + 0.002546840813010931, + 0.004098616074770689, + 0.004911057185381651, + 0.004843938630074263, + 0.019179778173565865, + 0.006001021713018417, + 0.01859492063522339, + 0.009790941141545773, + 0.008679194375872612, + 0.008387012407183647, + 0.012648150324821472, + 0.019081473350524902, + 0.17333738505840302, + 0.00969396997243166, + 0.17566856741905212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.005954147316515446 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197767712175846 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014197767712175846, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197767712175846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1893712282180786, + 0.07773414254188538, + 0.04874488711357117, + 0.08519507944583893, + 0.10929416865110397, + 0.19495058059692383, + 0.053955525159835815, + 0.04847148060798645, + 0.03900417312979698, + 0.15327872335910797 + ], + "credit_map": [ + 0.018849099054932594, + 0.017230750992894173, + 0.03840000182390213, + 0.02105771377682686, + 0.015310913324356079, + 0.025719759985804558, + 0.022213192656636238, + 0.03058980219066143, + 0.006447605788707733, + 0.004953091498464346, + 0.006329449359327555, + 0.010125773027539253, + 0.006795939989387989, + 0.013684112578630447, + 0.01469549909234047, + 0.014702665619552135, + 0.007564022671431303, + 0.004648926202207804, + 0.004273764789104462, + 0.0034781608264893293, + 0.004277828615158796, + 0.007654928136616945, + 0.0056610992178320885, + 0.01118615735322237, + 0.014235543087124825, + 0.008186882361769676, + 0.005070853512734175, + 0.00834247749298811, + 0.006448112428188324, + 0.01282363198697567, + 0.009724752977490425, + 0.020362818613648415, + 0.00851956196129322, + 0.01042590569704771, + 0.012258391827344894, + 0.008965922519564629, + 0.008552877232432365, + 0.01934441551566124, + 0.009722315706312656, + 0.03150477632880211, + 0.027217457070946693, + 0.015223332680761814, + 0.04948916658759117, + 0.022294005379080772, + 0.014009581878781319, + 0.02687084674835205, + 0.01799112744629383, + 0.021855060011148453, + 0.005735302343964577, + 0.0034849357325583696, + 0.006791930645704269, + 0.005822553299367428, + 0.006752708461135626, + 0.008541231974959373, + 0.008789297193288803, + 0.008037564344704151, + 0.0070068128407001495, + 0.004171493463218212, + 0.011196334846317768, + 0.004183030687272549, + 0.0030929793138056993, + 0.0077234008349478245, + 0.0033546241465955973, + 0.007742804475128651, + 0.0028112372383475304, + 0.0011368142440915108, + 0.0030363991390913725, + 0.002193455584347248, + 0.0023594293743371964, + 0.012380087748169899, + 0.0024249181151390076, + 0.012661831453442574, + 0.004493291024118662, + 0.0035806382074952126, + 0.0041364324279129505, + 0.005277142859995365, + 0.006971312221139669, + 0.06043769419193268, + 0.007652196101844311, + 0.060730013996362686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.014197767712175846 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a slight overhead angle focuses precisely on a palm-sized, uncut, angular piece of clear quartz crystal, delicately held by a pair of slender, fair hands. A bright, focused beam of light streams in from outside the frame, penetrating the crystal's crystalline interior. Its naturally formed ice cracks and tiny bubbles are instantly illuminated by the light, which is then refracted and fragmented by countless sharp facets, casting a brilliant, rainbow-like spectrum of light that softly falls on the fingertips and below the crystal. The hands hold the crystal with steady composure, their nails neatly manicured, and the texture of the skin is clearly visible in the soft light. The surroundings are enveloped in a shallow depth of field, completely blurred by the soft focus. Only faintly discernible scatterings of warm bokeh light create a subtle contrast with the cool spectrum of light cast by the crystal. The entire image, within the static composition, is imbued with a dynamic play of light and shadow, creating a mysterious atmosphere of tranquility, concentration, and a sense of wondrous discovery.", + "seed": 17236, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_0/index_2623_seed_17236_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10699779884889722, + "vlm_logprob": -0.10699779884889722, + "vlm_yes_no_margin": 3.625000011920929, + "vlm_token_credit": -0.10699779884889722, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08220026642084122 + ], + "opposite_logprob_tokens": [ + -2.832200288772583 + ], + "logprob_reward": -0.08220026642084122, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08220026642084122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.749053955078125, + 0.028606809675693512, + 0.018083909526467323, + 0.053270626813173294, + 0.0370568111538887, + 0.010658588260412216, + 0.03677625581622124, + 0.020970419049263, + 0.022675488144159317, + 0.022847115993499756 + ], + "credit_map": [ + 0.06740730255842209, + 0.03347206115722656, + 0.3003811240196228, + 0.024902377277612686, + 0.024064350873231888, + 0.04946011304855347, + 0.015839330852031708, + 0.23352734744548798, + 0.0041778963059186935, + 0.002351716160774231, + 0.0035853504668921232, + 0.002956867218017578, + 0.004360430873930454, + 0.003863074816763401, + 0.0023767000529915094, + 0.004934776108711958, + 0.002350491937249899, + 0.0017185454489663243, + 0.0023114890791475773, + 0.0018627169774845243, + 0.00327180209569633, + 0.0020383934024721384, + 0.001929228426888585, + 0.0026012430898845196, + 0.0078964838758111, + 0.004900296218693256, + 0.008186576887965202, + 0.003938120324164629, + 0.007869087159633636, + 0.006064246874302626, + 0.007004562299698591, + 0.007411260157823563, + 0.004064998123794794, + 0.0033680368214845657, + 0.004837909247726202, + 0.0029736021533608437, + 0.007987874560058117, + 0.003392481477931142, + 0.00563293369486928, + 0.004798976704478264, + 0.0010832458501681685, + 0.0009455775725655258, + 0.0012608051765710115, + 0.0008647168870083988, + 0.002837418345734477, + 0.0009713824256323278, + 0.0012557687005028129, + 0.0014396747574210167, + 0.0029874604661017656, + 0.002350992988795042, + 0.00365717732347548, + 0.0032837367616593838, + 0.00570251839235425, + 0.002812239807099104, + 0.010755039751529694, + 0.0052270954474806786, + 0.0023674755357205868, + 0.001583786099217832, + 0.001742773107253015, + 0.001619827002286911, + 0.004498232156038284, + 0.0024247062392532825, + 0.003636573674157262, + 0.003097047097980976, + 0.0018289190484210849, + 0.0020641321316361427, + 0.002467868849635124, + 0.002733865985646844, + 0.0049115801230072975, + 0.0016872695414349437, + 0.004013240337371826, + 0.0029686149209737778, + 0.002163512632250786, + 0.0017282760236412287, + 0.0035909959115087986, + 0.0023378736805170774, + 0.00469090836122632, + 0.0016607032157480717, + 0.003877338720485568, + 0.0027975093107670546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.08220026642084122 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012595508247613907 + ], + "opposite_logprob_tokens": [ + -4.3875956535339355 + ], + "logprob_reward": -0.012595508247613907, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012595508247613907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3949123024940491, + 0.10298524796962738, + 0.052652109414339066, + 0.04733346775174141, + 0.048362962901592255, + 0.05890539661049843, + 0.04981726035475731, + 0.054860975593328476, + 0.04268631711602211, + 0.1474839746952057 + ], + "credit_map": [ + 0.03859313577413559, + 0.02590637467801571, + 0.12374661862850189, + 0.030705703422427177, + 0.028911029919981956, + 0.027819084003567696, + 0.026615453884005547, + 0.09261487424373627, + 0.011791402474045753, + 0.005888713989406824, + 0.021136699244379997, + 0.008985546417534351, + 0.00761406822130084, + 0.014271100051701069, + 0.006439002696424723, + 0.026858719065785408, + 0.007069878745824099, + 0.004765473771840334, + 0.009798579849302769, + 0.004543868824839592, + 0.00562188308686018, + 0.006894703023135662, + 0.006119825411587954, + 0.007837898097932339, + 0.0077708992175757885, + 0.0038759191520512104, + 0.006483203265815973, + 0.0038372224662452936, + 0.006522559095174074, + 0.005774169694632292, + 0.006745050195604563, + 0.006324445828795433, + 0.006188805680721998, + 0.00502367876470089, + 0.010146471671760082, + 0.004250367172062397, + 0.005516883917152882, + 0.004710306413471699, + 0.006201855838298798, + 0.0063245948404073715, + 0.010656697675585747, + 0.0071086701937019825, + 0.007311177905648947, + 0.00643448019400239, + 0.007170931436121464, + 0.005237498786300421, + 0.007764528505504131, + 0.007221415638923645, + 0.011685317382216454, + 0.00456391042098403, + 0.005414863117039204, + 0.003319031558930874, + 0.004719937220215797, + 0.007174069061875343, + 0.006106276530772448, + 0.0068338532000780106, + 0.007476486265659332, + 0.008965179324150085, + 0.005914691835641861, + 0.004924899898469448, + 0.007209488656371832, + 0.006130264140665531, + 0.007250484079122543, + 0.006989477202296257, + 0.0057345167733728886, + 0.006592788267880678, + 0.0061525609344244, + 0.0031320261768996716, + 0.0060053239576518536, + 0.00482857134193182, + 0.0046708169393241405, + 0.005569714121520519, + 0.020133623853325844, + 0.012376690283417702, + 0.03702859580516815, + 0.014530402608215809, + 0.01840911991894245, + 0.01660105399787426, + 0.013215567916631699, + 0.015188916586339474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.012595508247613907 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289814829826355 + ], + "opposite_logprob_tokens": [ + -1.0539815425872803 + ], + "logprob_reward": -0.4289814829826355, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.4289814829826355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5666860342025757, + 0.04164142906665802, + 0.07133599370718002, + 0.09745404124259949, + 0.036555204540491104, + 0.02736426331102848, + 0.06446580588817596, + 0.020467612892389297, + 0.01952149160206318, + 0.05450809374451637 + ], + "credit_map": [ + 0.05181516334414482, + 0.027525313198566437, + 0.2205391526222229, + 0.02383584715425968, + 0.02395853027701378, + 0.03647476062178612, + 0.01550650130957365, + 0.16703076660633087, + 0.00646476075053215, + 0.003622977063059807, + 0.00781333725899458, + 0.0037123640067875385, + 0.0044041951186954975, + 0.005401110276579857, + 0.004604235757142305, + 0.005618450231850147, + 0.013292043469846249, + 0.007406917400658131, + 0.01001797802746296, + 0.004929989110678434, + 0.008546173572540283, + 0.010453294962644577, + 0.006178098730742931, + 0.01051149982959032, + 0.016579067334532738, + 0.009617388248443604, + 0.015030033886432648, + 0.007132955361157656, + 0.01302643958479166, + 0.009279828518629074, + 0.012849405407905579, + 0.013938923366367817, + 0.006077989935874939, + 0.003089497797191143, + 0.004238009452819824, + 0.002771294442936778, + 0.006558447610586882, + 0.00410440843552351, + 0.005230647977441549, + 0.004484908189624548, + 0.005038195289671421, + 0.0021417627576738596, + 0.0027184919454157352, + 0.00227112858556211, + 0.004972873255610466, + 0.0028505963273346424, + 0.0038804602809250355, + 0.0034907555673271418, + 0.005913203116506338, + 0.005587057210505009, + 0.008247868157923222, + 0.004881186410784721, + 0.01064938586205244, + 0.004454524256289005, + 0.016215406358242035, + 0.008517177775502205, + 0.0018740675877779722, + 0.0020076509099453688, + 0.0028338476549834013, + 0.0020015286281704903, + 0.003729318268597126, + 0.001646466669626534, + 0.0036979159340262413, + 0.002676816424354911, + 0.002287296811118722, + 0.0029472103342413902, + 0.0035214831586927176, + 0.0018917567795142531, + 0.002228358993306756, + 0.0019745195750147104, + 0.002056960016489029, + 0.0026139060501009226, + 0.008361287415027618, + 0.0048065087758004665, + 0.008176392875611782, + 0.00374598428606987, + 0.006056339014321566, + 0.006100322585552931, + 0.010484985075891018, + 0.0067762755788862705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.4289814829826355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954503081738949 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.005954503081738949, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954503081738949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7708846926689148, + 0.024589024484157562, + 0.020475957542657852, + 0.014541535638272762, + 0.035824861377477646, + 0.019972411915659904, + 0.03611360490322113, + 0.039843324571847916, + 0.0169422198086977, + 0.020812340080738068 + ], + "credit_map": [ + 0.07217741012573242, + 0.034737393260002136, + 0.3104134798049927, + 0.027201471850275993, + 0.020439255982637405, + 0.05044268071651459, + 0.013570528477430344, + 0.24190251529216766, + 0.00212094746530056, + 0.0013195659266784787, + 0.005198437720537186, + 0.0021057063713669777, + 0.002250605495646596, + 0.003021342447027564, + 0.002471580635756254, + 0.00610083993524313, + 0.004934297874569893, + 0.0018795236246660352, + 0.0026545885484665632, + 0.0014678691513836384, + 0.0019759247079491615, + 0.0026055846828967333, + 0.002044356893748045, + 0.002913811942562461, + 0.00149954529479146, + 0.0012570988619700074, + 0.002158006187528372, + 0.001325829653069377, + 0.0027444486040621996, + 0.0012614624574780464, + 0.0025474748108536005, + 0.0017476691864430904, + 0.003588775172829628, + 0.003931140992790461, + 0.005210306961089373, + 0.0030641581397503614, + 0.006869552657008171, + 0.003661218797788024, + 0.004397082142531872, + 0.005102624651044607, + 0.0021976588759571314, + 0.0013261467684060335, + 0.001974930753931403, + 0.0016541560180485249, + 0.004925334360450506, + 0.0013238497776910663, + 0.003484598593786359, + 0.003085737582296133, + 0.0025654828641563654, + 0.0024487089831382036, + 0.0036901612766087055, + 0.002978730481117964, + 0.011302342638373375, + 0.002362198429182172, + 0.005298498552292585, + 0.0054674833081662655, + 0.0033801267854869366, + 0.0029533219058066607, + 0.003563588485121727, + 0.002987307496368885, + 0.011735601350665092, + 0.002989632310345769, + 0.005958592053502798, + 0.0062751551158726215, + 0.0021896606776863337, + 0.0016866116784512997, + 0.0017148035112768412, + 0.002125318394973874, + 0.0022558069322258234, + 0.0013034824514761567, + 0.003294671420007944, + 0.002371865790337324, + 0.0033705050591379404, + 0.0015580194303765893, + 0.0036786405835300684, + 0.002622379921376705, + 0.0025444328784942627, + 0.0014034881023690104, + 0.0032471863087266684, + 0.00238768826238811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.005954503081738949 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257233511656523 + ], + "opposite_logprob_tokens": [ + -5.2552571296691895 + ], + "logprob_reward": -0.005257233511656523, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257233511656523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1172567829489708, + 0.019944369792938232, + 0.06444743275642395, + 0.03330330178141594, + 0.025847140699625015, + 0.03619948774576187, + 0.036634478718042374, + 0.03922772780060768, + 0.043046943843364716, + 0.5840923190116882 + ], + "credit_map": [ + 0.012834065593779087, + 0.006573641672730446, + 0.03264137730002403, + 0.008677561767399311, + 0.008489018306136131, + 0.011679441668093204, + 0.007513160817325115, + 0.02884851023554802, + 0.0025481623597443104, + 0.0017794116865843534, + 0.0028375994879752398, + 0.002253370825201273, + 0.0019860868342220783, + 0.002375757321715355, + 0.002312774071469903, + 0.003851208370178938, + 0.012484799139201641, + 0.006416748743504286, + 0.011790378950536251, + 0.006021119188517332, + 0.00434077437967062, + 0.00860676821321249, + 0.005801247898489237, + 0.008985595777630806, + 0.008708271197974682, + 0.0029017445631325245, + 0.005537395365536213, + 0.002198629081249237, + 0.002160732401534915, + 0.003857234725728631, + 0.0029723276384174824, + 0.004966966342180967, + 0.0035650262143462896, + 0.0024232978466898203, + 0.003723782952874899, + 0.0017643377650529146, + 0.004797901958227158, + 0.003045212011784315, + 0.0031205276027321815, + 0.0034070538822561502, + 0.008844505064189434, + 0.003679631743580103, + 0.0034760565031319857, + 0.0020129354670643806, + 0.0050407215021550655, + 0.004160991869866848, + 0.003909582272171974, + 0.005075066350400448, + 0.0053002675995230675, + 0.0026434350293129683, + 0.0034466327633708715, + 0.0027417801320552826, + 0.004564226604998112, + 0.0031173310708254576, + 0.009719760157167912, + 0.005101043730974197, + 0.009713361971080303, + 0.0037770522758364677, + 0.0045253741554915905, + 0.0024389615282416344, + 0.004365453030914068, + 0.005151813849806786, + 0.0036502762231975794, + 0.005605432670563459, + 0.00582992285490036, + 0.005195733159780502, + 0.010767164640128613, + 0.004130870569497347, + 0.0032763048075139523, + 0.004946088884025812, + 0.0031829990912228823, + 0.005717862863093615, + 0.05750049278140068, + 0.06691058725118637, + 0.1269283890724182, + 0.04677579924464226, + 0.05815204232931137, + 0.05564460530877113, + 0.10011935979127884, + 0.07206108421087265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.005257233511656523 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a slight overhead angle focuses precisely on a palm-sized, uncut, angular piece of clear quartz crystal, delicately held by a pair of slender, fair hands. A bright, focused beam of light streams in from outside the frame, penetrating the crystal's crystalline interior. Its naturally formed ice cracks and tiny bubbles are instantly illuminated by the light, which is then refracted and fragmented by countless sharp facets, casting a brilliant, rainbow-like spectrum of light that softly falls on the fingertips and below the crystal. The hands hold the crystal with steady composure, their nails neatly manicured, and the texture of the skin is clearly visible in the soft light. The surroundings are enveloped in a shallow depth of field, completely blurred by the soft focus. Only faintly discernible scatterings of warm bokeh light create a subtle contrast with the cool spectrum of light cast by the crystal. The entire image, within the static composition, is imbued with a dynamic play of light and shadow, creating a mysterious atmosphere of tranquility, concentration, and a sense of wondrous discovery.", + "seed": 17237, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_0/index_2623_seed_17237_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04608329525217414, + "vlm_logprob": -0.04608329525217414, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.04608329525217414, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0423564612865448 + ], + "opposite_logprob_tokens": [ + -3.667356491088867 + ], + "logprob_reward": -0.0423564612865448, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0423564612865448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0999380499124527, + 0.25960591435432434, + 0.030107690021395683, + 0.0251015592366457, + 0.07845824211835861, + 0.08366259187459946, + 0.2668766677379608, + 0.043122511357069016, + 0.033620547503232956, + 0.0795062780380249 + ], + "credit_map": [ + 0.01023964025080204, + 0.012330501340329647, + 0.012767725624144077, + 0.020058829337358475, + 0.008800539188086987, + 0.014028901234269142, + 0.013593494892120361, + 0.00811841432005167, + 0.05172218382358551, + 0.023197751492261887, + 0.04086131229996681, + 0.04211676865816116, + 0.019468283280730247, + 0.03215395659208298, + 0.026454810053110123, + 0.023630864918231964, + 0.004656423814594746, + 0.003259076038375497, + 0.004071390256285667, + 0.005373802036046982, + 0.002674050396308303, + 0.0031812877859920263, + 0.004193954635411501, + 0.0026977062225341797, + 0.0030731409788131714, + 0.003236280055716634, + 0.0032431476283818483, + 0.005391099490225315, + 0.0021918397396802902, + 0.002713835332542658, + 0.0031530188862234354, + 0.0020991964265704155, + 0.015307122841477394, + 0.0095317792147398, + 0.010194248519837856, + 0.015361452475190163, + 0.004520769231021404, + 0.009425181895494461, + 0.008021755144000053, + 0.006095935124903917, + 0.014378544874489307, + 0.00849829986691475, + 0.01023026928305626, + 0.011494637466967106, + 0.006620454136282206, + 0.014380741864442825, + 0.009819990023970604, + 0.008239648304879665, + 0.04725167527794838, + 0.024407660588622093, + 0.03049863688647747, + 0.04457125440239906, + 0.020392434671521187, + 0.041074708104133606, + 0.033956483006477356, + 0.024723809212446213, + 0.004230794496834278, + 0.004841037560254335, + 0.005089797545224428, + 0.00784018449485302, + 0.004991438239812851, + 0.0058224815875291824, + 0.006440126337110996, + 0.003866651328280568, + 0.004937016870826483, + 0.0034931411501020193, + 0.004221722949296236, + 0.006022818386554718, + 0.003193739103153348, + 0.003947171848267317, + 0.0048102447763085365, + 0.0029946924187242985, + 0.007974853739142418, + 0.008522042073309422, + 0.008777990937232971, + 0.01219929102808237, + 0.006788153201341629, + 0.016988040879368782, + 0.010696723125874996, + 0.007559183053672314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.0423564612865448 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337299108505249 + ], + "opposite_logprob_tokens": [ + -3.4087300300598145 + ], + "logprob_reward": -0.0337299108505249, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337299108505249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11097016930580139, + 0.1165333017706871, + 0.05497921630740166, + 0.022908871993422508, + 0.07349739223718643, + 0.05897555500268936, + 0.037031110376119614, + 0.06975264102220535, + 0.04197319597005844, + 0.41337859630584717 + ], + "credit_map": [ + 0.013456575572490692, + 0.012277868576347828, + 0.014677315950393677, + 0.0194924995303154, + 0.008358461782336235, + 0.017334915697574615, + 0.016143586486577988, + 0.009228946641087532, + 0.01852750964462757, + 0.012282174080610275, + 0.01300270389765501, + 0.013645846396684647, + 0.008173190988600254, + 0.022403810173273087, + 0.01661599986255169, + 0.011882058344781399, + 0.006950817070901394, + 0.006202456075698137, + 0.005899591837078333, + 0.008135892450809479, + 0.004145655315369368, + 0.011882320046424866, + 0.006299799773842096, + 0.005462689325213432, + 0.0031302294228225946, + 0.0023445484694093466, + 0.0031485927756875753, + 0.0036145492922514677, + 0.0018811463378369808, + 0.0037855564150959253, + 0.0028051133267581463, + 0.0021991354878991842, + 0.006641869433224201, + 0.008456568233668804, + 0.007562067359685898, + 0.011791782453656197, + 0.007421927526593208, + 0.010464741848409176, + 0.014804751612246037, + 0.006353677716106176, + 0.011102385818958282, + 0.004608465824276209, + 0.006190425716340542, + 0.0073126633651554585, + 0.004346048459410667, + 0.012181810103356838, + 0.0066828737035393715, + 0.006550880149006844, + 0.0038245897740125656, + 0.0037496830336749554, + 0.004325300920754671, + 0.004401880782097578, + 0.0035316981375217438, + 0.007910938002169132, + 0.005434213671833277, + 0.0038528069853782654, + 0.00812462903559208, + 0.008870587684214115, + 0.008672227151691914, + 0.01178698055446148, + 0.006909784860908985, + 0.012427978217601776, + 0.006451187655329704, + 0.006509270053356886, + 0.0058965603820979595, + 0.003629551501944661, + 0.007115728221833706, + 0.005660209804773331, + 0.005484294146299362, + 0.0054961699061095715, + 0.005101636052131653, + 0.0035890419967472553, + 0.05927078425884247, + 0.03294330835342407, + 0.04629593342542648, + 0.04421401023864746, + 0.02832719497382641, + 0.08797292411327362, + 0.05417769402265549, + 0.060176730155944824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.0337299108505249 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283083379268646 + ], + "opposite_logprob_tokens": [ + -2.0178308486938477 + ], + "logprob_reward": -0.14283083379268646, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283083379268646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14682534337043762, + 0.15890827775001526, + 0.06686917692422867, + 0.0588652566075325, + 0.04752659797668457, + 0.0632447749376297, + 0.052669767290353775, + 0.21542955935001373, + 0.037128645926713943, + 0.1525326371192932 + ], + "credit_map": [ + 0.022616082802414894, + 0.013797976076602936, + 0.02748963050544262, + 0.02400319091975689, + 0.010522234253585339, + 0.017354691401124, + 0.019594257697463036, + 0.011447285301983356, + 0.029222581535577774, + 0.015776511281728745, + 0.023704443126916885, + 0.02609909325838089, + 0.013251484371721745, + 0.02172716334462166, + 0.014469942077994347, + 0.014657054096460342, + 0.009008639492094517, + 0.00751251308247447, + 0.010514670051634312, + 0.012309547513723373, + 0.0050420560874044895, + 0.008638990111649036, + 0.008160052821040154, + 0.00568270543590188, + 0.00902662705630064, + 0.004932009615004063, + 0.008101620711386204, + 0.011283555068075657, + 0.003363980446010828, + 0.008596289902925491, + 0.008396382443606853, + 0.005164790898561478, + 0.006085892673581839, + 0.0058476608246564865, + 0.0055360374972224236, + 0.008402392268180847, + 0.004012808203697205, + 0.006902538239955902, + 0.006432426627725363, + 0.00430684071034193, + 0.00806404184550047, + 0.008585147559642792, + 0.007496614009141922, + 0.009865348227322102, + 0.004886206705123186, + 0.009388805367052555, + 0.008084353059530258, + 0.006874254904687405, + 0.007829885929822922, + 0.005186813417822123, + 0.006817819084972143, + 0.00854240171611309, + 0.004144825506955385, + 0.008588791824877262, + 0.006068113725632429, + 0.005491121206432581, + 0.026101194322109222, + 0.022645337507128716, + 0.030296724289655685, + 0.026631491258740425, + 0.019704144448041916, + 0.04391733929514885, + 0.02584071271121502, + 0.020292609930038452, + 0.004698409698903561, + 0.00376976584084332, + 0.004475218243896961, + 0.0049329656176269054, + 0.003691337537020445, + 0.008775091730058193, + 0.0034079416655004025, + 0.003377916757017374, + 0.016932524740695953, + 0.016087567433714867, + 0.018052171915769577, + 0.02261609397828579, + 0.01089827623218298, + 0.025623708963394165, + 0.023025348782539368, + 0.019296962767839432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.14283083379268646 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017158561386168003 + ], + "opposite_logprob_tokens": [ + -6.376715660095215 + ], + "logprob_reward": -0.0017158561386168003, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017158561386168003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16307318210601807, + 0.13303904235363007, + 0.07113195955753326, + 0.046328406780958176, + 0.04892042279243469, + 0.03623161092400551, + 0.05170978978276253, + 0.1333940327167511, + 0.11379019170999527, + 0.20238138735294342 + ], + "credit_map": [ + 0.027767715975642204, + 0.013229237869381905, + 0.023260846734046936, + 0.031331244856119156, + 0.010747377760708332, + 0.024093668907880783, + 0.017034554854035378, + 0.01560852862894535, + 0.020925724878907204, + 0.015429573133587837, + 0.020560704171657562, + 0.024423999711871147, + 0.011868134140968323, + 0.013903398998081684, + 0.01449780073016882, + 0.011429707519710064, + 0.010121633298695087, + 0.008464626967906952, + 0.010865096002817154, + 0.016386019065976143, + 0.004340325016528368, + 0.007407851051539183, + 0.008173036389052868, + 0.005373374093323946, + 0.006630104500800371, + 0.004925715737044811, + 0.0065572322346270084, + 0.009321775287389755, + 0.002667095046490431, + 0.004841485060751438, + 0.007544999476522207, + 0.0038399978075176477, + 0.007409887854009867, + 0.0053050098940730095, + 0.006525760516524315, + 0.009729287587106228, + 0.0030731300357729197, + 0.0070158978924155235, + 0.006171368528157473, + 0.0036900783888995647, + 0.004813617095351219, + 0.004213112872093916, + 0.004288214258849621, + 0.004779417999088764, + 0.00306499470025301, + 0.007258431054651737, + 0.0038517937064170837, + 0.003962031099945307, + 0.005364834330976009, + 0.004845941439270973, + 0.005870952270925045, + 0.007189480122178793, + 0.006402234081178904, + 0.009707057848572731, + 0.007336712907999754, + 0.00499257817864418, + 0.02030172012746334, + 0.014882606454193592, + 0.01831498183310032, + 0.019965944811701775, + 0.00875876285135746, + 0.027428308501839638, + 0.011052865535020828, + 0.01268883515149355, + 0.016010450199246407, + 0.010761951096355915, + 0.0138686653226614, + 0.014023524709045887, + 0.00988257396966219, + 0.023506438359618187, + 0.01322172861546278, + 0.012514865025877953, + 0.026437772437930107, + 0.018651045858860016, + 0.02492617256939411, + 0.029142074286937714, + 0.012389730662107468, + 0.0342119000852108, + 0.02900976873934269, + 0.027612915262579918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.0017158561386168003 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978341419249773 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.00978341419249773, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978341419249773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08401598781347275, + 0.049701813608407974, + 0.0539926253259182, + 0.030701769515872, + 0.020428596064448357, + 0.023360975086688995, + 0.0220966599881649, + 0.16149026155471802, + 0.06616102904081345, + 0.48805031180381775 + ], + "credit_map": [ + 0.015691962093114853, + 0.007249342277646065, + 0.013359603472054005, + 0.011081763543188572, + 0.0047011240385472775, + 0.013895544223487377, + 0.010658654384315014, + 0.007377991918474436, + 0.0066432049497962, + 0.004389981273561716, + 0.0069898623041808605, + 0.006461218930780888, + 0.0036846199072897434, + 0.010620374232530594, + 0.006270525045692921, + 0.004642019048333168, + 0.006963052321225405, + 0.0051922001875936985, + 0.0068394835107028484, + 0.007509630639106035, + 0.004679440520703793, + 0.01002245582640171, + 0.0065330988727509975, + 0.0062532611191272736, + 0.003784127999097109, + 0.00281928782351315, + 0.003927748650312424, + 0.004709064494818449, + 0.0023476446513086557, + 0.005463937297463417, + 0.00429063756018877, + 0.0033593184780329466, + 0.002853458048775792, + 0.001894359476864338, + 0.002512363949790597, + 0.003934514708817005, + 0.0013920852215960622, + 0.003503549611195922, + 0.0024577705189585686, + 0.0018804939463734627, + 0.003099890425801277, + 0.002525348449125886, + 0.002764276694506407, + 0.0048147751949727535, + 0.0021974314004182816, + 0.0028774249367415905, + 0.003420037915930152, + 0.0016617903020232916, + 0.0028152833692729473, + 0.002589994575828314, + 0.003464898094534874, + 0.00381023483350873, + 0.0017327620880678296, + 0.0032858666963875294, + 0.0024795092176645994, + 0.001918110647238791, + 0.014326724223792553, + 0.01864415965974331, + 0.016590958461165428, + 0.02158544398844242, + 0.015692848712205887, + 0.03478160500526428, + 0.022019822150468826, + 0.017848707735538483, + 0.009710688143968582, + 0.006457829847931862, + 0.011593788862228394, + 0.008003203198313713, + 0.006005414761602879, + 0.009891963563859463, + 0.008611521683633327, + 0.005886616185307503, + 0.05694018304347992, + 0.05475473031401634, + 0.06341640651226044, + 0.07531772553920746, + 0.03488098829984665, + 0.07692812383174896, + 0.07083223015069962, + 0.054979920387268066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.00978341419249773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a slight overhead angle focuses precisely on a palm-sized, uncut, angular piece of clear quartz crystal, delicately held by a pair of slender, fair hands. A bright, focused beam of light streams in from outside the frame, penetrating the crystal's crystalline interior. Its naturally formed ice cracks and tiny bubbles are instantly illuminated by the light, which is then refracted and fragmented by countless sharp facets, casting a brilliant, rainbow-like spectrum of light that softly falls on the fingertips and below the crystal. The hands hold the crystal with steady composure, their nails neatly manicured, and the texture of the skin is clearly visible in the soft light. The surroundings are enveloped in a shallow depth of field, completely blurred by the soft focus. Only faintly discernible scatterings of warm bokeh light create a subtle contrast with the cool spectrum of light cast by the crystal. The entire image, within the static composition, is imbued with a dynamic play of light and shadow, creating a mysterious atmosphere of tranquility, concentration, and a sense of wondrous discovery.", + "seed": 17238, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_0/index_2623_seed_17238_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10336377283092588, + "vlm_logprob": -0.10336377283092588, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.10336377283092588, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020201295614242554 + ], + "opposite_logprob_tokens": [ + -5.145201206207275 + ], + "logprob_reward": -0.020201295614242554, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.020201295614242554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09893441200256348, + 0.06362900882959366, + 0.14842940866947174, + 0.2558898329734802, + 0.09857575595378876, + 0.05775592476129532, + 0.1062011569738388, + 0.05761461332440376, + 0.08037157356739044, + 0.03259826451539993 + ], + "credit_map": [ + 0.011234861798584461, + 0.016187820583581924, + 0.013171245343983173, + 0.018023353070020676, + 0.009256094694137573, + 0.010020583868026733, + 0.0071072326973080635, + 0.013933215290307999, + 0.009311496280133724, + 0.010038211941719055, + 0.007709217257797718, + 0.006274471525102854, + 0.004826353397220373, + 0.007571962662041187, + 0.003973096609115601, + 0.013924198225140572, + 0.012017753906548023, + 0.014285369776189327, + 0.0209294855594635, + 0.014442925341427326, + 0.0145645122975111, + 0.03221859782934189, + 0.014561733230948448, + 0.025409040972590446, + 0.0286527331918478, + 0.031209351494908333, + 0.03154616802930832, + 0.042141158133745193, + 0.026171209290623665, + 0.0376446396112442, + 0.02345832996070385, + 0.035066209733486176, + 0.009404693730175495, + 0.017071150243282318, + 0.012795842252671719, + 0.014882778748869896, + 0.008513198234140873, + 0.01193813793361187, + 0.008133557625114918, + 0.01583639532327652, + 0.004792461171746254, + 0.004928027745336294, + 0.009425289928913116, + 0.004002004861831665, + 0.004107080399990082, + 0.009364712052047253, + 0.0036627391818910837, + 0.01747361198067665, + 0.017268911004066467, + 0.010345564223825932, + 0.014958770014345646, + 0.008663788437843323, + 0.009507023729383945, + 0.01662949100136757, + 0.013439667411148548, + 0.015387938357889652, + 0.006772763095796108, + 0.0057235113345086575, + 0.007859992794692516, + 0.0057286168448626995, + 0.005494632758200169, + 0.00849182903766632, + 0.007444743067026138, + 0.010098529048264027, + 0.008499067276716232, + 0.008155714720487595, + 0.01222979836165905, + 0.00531095452606678, + 0.007836992852389812, + 0.01413387805223465, + 0.01292618177831173, + 0.011278982274234295, + 0.003221869934350252, + 0.0034803610760718584, + 0.0042764730751514435, + 0.0027746951673179865, + 0.003248505527153611, + 0.005475662183016539, + 0.004577352199703455, + 0.0055433474481105804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.020201295614242554 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016105085611343384 + ], + "opposite_logprob_tokens": [ + -4.1411051750183105 + ], + "logprob_reward": -0.016105085611343384, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016105085611343384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12536649405956268, + 0.09886825084686279, + 0.33129188418388367, + 0.06720616668462753, + 0.021970519796013832, + 0.06254256516695023, + 0.11069174110889435, + 0.07716362178325653, + 0.06863676011562347, + 0.036262065172195435 + ], + "credit_map": [ + 0.006897286046296358, + 0.01118168793618679, + 0.01862828992307186, + 0.013215637765824795, + 0.007588450331240892, + 0.03274823725223541, + 0.005970886442810297, + 0.029136009514331818, + 0.01083123404532671, + 0.012180090881884098, + 0.015362238511443138, + 0.008424402214586735, + 0.003930947277694941, + 0.0236541498452425, + 0.005498338956385851, + 0.01898685097694397, + 0.025783570483326912, + 0.04285135865211487, + 0.03810008987784386, + 0.0505678690969944, + 0.03994680568575859, + 0.03652438893914223, + 0.033802542835474014, + 0.06371524184942245, + 0.006538699846714735, + 0.0063085006549954414, + 0.008792228996753693, + 0.009781534783542156, + 0.006022052839398384, + 0.008942141197621822, + 0.006639257539063692, + 0.014181751757860184, + 0.0019269195618107915, + 0.0033574863336980343, + 0.002682530088350177, + 0.0030984499026089907, + 0.0018995563732460141, + 0.003597893053665757, + 0.001908913953229785, + 0.0034987707622349262, + 0.005385413765907288, + 0.007331832777708769, + 0.007991977035999298, + 0.007807039190083742, + 0.005170764401555061, + 0.010704170912504196, + 0.005501471925526857, + 0.012649893760681152, + 0.016509847715497017, + 0.009547964669764042, + 0.015398083254694939, + 0.008602898567914963, + 0.009202241897583008, + 0.019747203215956688, + 0.013217160478234291, + 0.018466340377926826, + 0.009099452756345272, + 0.008857619017362595, + 0.009731967002153397, + 0.010116292163729668, + 0.010339486412703991, + 0.009765883907675743, + 0.007774696219712496, + 0.01147821918129921, + 0.008502624928951263, + 0.008586964569985867, + 0.00949111208319664, + 0.005078973714262247, + 0.007826797664165497, + 0.011586648412048817, + 0.00807610061019659, + 0.009487542323768139, + 0.004466210957616568, + 0.0057840170338749886, + 0.007293325383216143, + 0.002784733660519123, + 0.0032838473562151194, + 0.0047098309732973576, + 0.003118760185316205, + 0.00482134148478508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.016105085611343384 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743555784225464 + ], + "opposite_logprob_tokens": [ + -0.9743555784225464 + ], + "logprob_reward": -0.4743555784225464, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743555784225464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10198967903852463, + 0.12368322908878326, + 0.24215178191661835, + 0.21283058822155, + 0.06306153535842896, + 0.08882748335599899, + 0.059741247445344925, + 0.03506581112742424, + 0.026158779859542847, + 0.0464899018406868 + ], + "credit_map": [ + 0.008676464669406414, + 0.008331844583153725, + 0.01724625937640667, + 0.009630408138036728, + 0.0066940076649188995, + 0.02966109849512577, + 0.005082971882075071, + 0.01666661724448204, + 0.017748568207025528, + 0.009899739176034927, + 0.018490854650735855, + 0.008110369555652142, + 0.007586283143609762, + 0.02055458538234234, + 0.006236669607460499, + 0.035056158900260925, + 0.02406030334532261, + 0.027545776218175888, + 0.030231349170207977, + 0.03886030986905098, + 0.01888643205165863, + 0.03731939196586609, + 0.016178449615836143, + 0.04906976968050003, + 0.03088187798857689, + 0.01975565403699875, + 0.03126909211277962, + 0.018140487372875214, + 0.01335925329476595, + 0.04213379696011543, + 0.013641244731843472, + 0.04364917427301407, + 0.005509601440280676, + 0.009115321561694145, + 0.007068488746881485, + 0.009044277481734753, + 0.004605161026120186, + 0.00841654371470213, + 0.005382440984249115, + 0.013919703662395477, + 0.0075226654298603535, + 0.0075583565048873425, + 0.012550384737551212, + 0.009151162579655647, + 0.005493451841175556, + 0.01522170938551426, + 0.006025123409926891, + 0.025304626673460007, + 0.007115636020898819, + 0.004977609496563673, + 0.009778444655239582, + 0.00452402513474226, + 0.004337981343269348, + 0.011317639611661434, + 0.0066319615580141544, + 0.011057949624955654, + 0.004656224045902491, + 0.0037710105534642935, + 0.006122879218310118, + 0.0029499458614736795, + 0.003435009391978383, + 0.005755303427577019, + 0.003118334338068962, + 0.005257103126496077, + 0.0029872716404497623, + 0.002178790047764778, + 0.0038047153502702713, + 0.001909154001623392, + 0.0026614924427121878, + 0.005952093284577131, + 0.002125089056789875, + 0.004540173336863518, + 0.007163504604250193, + 0.0076566413044929504, + 0.005949939135462046, + 0.00505361333489418, + 0.005025159567594528, + 0.004658287391066551, + 0.004834963474422693, + 0.0061477976851165295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.4743555784225464 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464171078056097 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.00464171078056097, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464171078056097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12498866766691208, + 0.04209355637431145, + 0.06251101940870285, + 0.09973369538784027, + 0.07801961898803711, + 0.17083381116390228, + 0.15841171145439148, + 0.09025907516479492, + 0.08858342468738556, + 0.08456548303365707 + ], + "credit_map": [ + 0.006718428805470467, + 0.012527111917734146, + 0.018665820360183716, + 0.010913624428212643, + 0.007448081858456135, + 0.030780991539359093, + 0.006773442029953003, + 0.031161164864897728, + 0.0056180101819336414, + 0.006439738906919956, + 0.004130137152969837, + 0.005619172938168049, + 0.003936328459531069, + 0.005686061456799507, + 0.00391360092908144, + 0.006750506814569235, + 0.007799658924341202, + 0.00693921372294426, + 0.0067905536852777, + 0.008673450909554958, + 0.004802200477570295, + 0.008602460846304893, + 0.00394816230982542, + 0.014955317601561546, + 0.011570833623409271, + 0.009232884272933006, + 0.013017721474170685, + 0.013109630905091763, + 0.0065540228970348835, + 0.017783259972929955, + 0.007507153321057558, + 0.020958187058568, + 0.0063599152490496635, + 0.013354930095374584, + 0.008372467011213303, + 0.014439939521253109, + 0.009000389836728573, + 0.00788502674549818, + 0.006424609571695328, + 0.012182344682514668, + 0.012231181375682354, + 0.020755354315042496, + 0.02238444983959198, + 0.023853903636336327, + 0.013402041979134083, + 0.030796384438872337, + 0.015220512636005878, + 0.032189980149269104, + 0.02038346417248249, + 0.017251642420887947, + 0.027302097529172897, + 0.013074248097836971, + 0.015462144277989864, + 0.02794494666159153, + 0.014101811684668064, + 0.02289135940372944, + 0.012367037124931812, + 0.010115982964634895, + 0.012791452929377556, + 0.010566902346909046, + 0.009412790648639202, + 0.012271950021386147, + 0.009684777818620205, + 0.013048185966908932, + 0.011236107908189297, + 0.011152079328894615, + 0.011961071752011776, + 0.006916711572557688, + 0.00949473213404417, + 0.01420908235013485, + 0.011966145597398281, + 0.01164749450981617, + 0.012950848788022995, + 0.0103798508644104, + 0.01227722316980362, + 0.006969295907765627, + 0.009007252752780914, + 0.013958926312625408, + 0.007516308221966028, + 0.011505781672894955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.00464171078056097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a palm-sized, uncut, angular piece of clear quartz crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands holding the crystal slender, fair, and featuring neatly manicured nails?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bright, focused beam of light visible streaming from outside the frame into the crystal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light refract through the crystal to cast a rainbow-like spectrum on the fingertips and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with shallow depth of field and warm bokeh light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015151937259361148 + ], + "opposite_logprob_tokens": [ + -6.5015153884887695 + ], + "logprob_reward": -0.0015151937259361148, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015151937259361148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19928884506225586, + 0.06935692578554153, + 0.0504581592977047, + 0.15316273272037506, + 0.10291584581136703, + 0.07373523712158203, + 0.16278652846813202, + 0.07494490593671799, + 0.08003245294094086, + 0.033318374305963516 + ], + "credit_map": [ + 0.016932008787989616, + 0.013033214025199413, + 0.032082751393318176, + 0.011749877594411373, + 0.010337007232010365, + 0.04597637802362442, + 0.008131123147904873, + 0.06104648858308792, + 0.006051875650882721, + 0.00660267798230052, + 0.012034098617732525, + 0.004979201592504978, + 0.0033799009397625923, + 0.011345827020704746, + 0.003154320875182748, + 0.021809032186865807, + 0.003914446569979191, + 0.005986671429127455, + 0.0057042683474719524, + 0.008908011950552464, + 0.004588288255035877, + 0.006277496926486492, + 0.005357964430004358, + 0.00972101092338562, + 0.014216789044439793, + 0.012110288254916668, + 0.021929895505309105, + 0.014220649376511574, + 0.009476210922002792, + 0.027685679495334625, + 0.008748677559196949, + 0.04477454349398613, + 0.009690197184681892, + 0.011263621971011162, + 0.014119025319814682, + 0.010355084203183651, + 0.006506603676825762, + 0.020439594984054565, + 0.007004269398748875, + 0.02353745512664318, + 0.005401311442255974, + 0.005913243629038334, + 0.009968153201043606, + 0.00712299020960927, + 0.0047900923527777195, + 0.013142846524715424, + 0.005197149235755205, + 0.02219945192337036, + 0.02270413376390934, + 0.017726602032780647, + 0.02230701968073845, + 0.013618177734315395, + 0.01603798381984234, + 0.027628544718027115, + 0.01749066449701786, + 0.02527341991662979, + 0.010465665720403194, + 0.0071824477054178715, + 0.011082589626312256, + 0.006993042770773172, + 0.005860466975718737, + 0.012702589854598045, + 0.008177529089152813, + 0.012480577453970909, + 0.00980257149785757, + 0.008370010182261467, + 0.011177271604537964, + 0.0065730269998312, + 0.009212082251906395, + 0.012194335460662842, + 0.013940343633294106, + 0.008762811310589314, + 0.005732792429625988, + 0.0037808932829648256, + 0.004664885811507702, + 0.002191386418417096, + 0.002995023736730218, + 0.005487225018441677, + 0.0033681984059512615, + 0.005097972694784403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.0015151937259361148 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_1/index_tensor([783, 783])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_1/index_tensor([783, 783])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e6439ba8ab449f7689584e5fd4b9dde1fc1251c5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_1/index_tensor([783, 783])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium shot, held at eye level, features a plump, cartoonish blue bird, rendered in a classic 2D animation style with crisp lines and vibrant colors. The bird, with its bright sky-blue feathers and soft, creamy-white belly, stands cheerfully in the center of the frame. Its large, round, black eyes sparkle with friendly warmth, and its small, triangular yellow beak opens in a broad, joyful smile. The bird enthusiastically puffs out its chest and spreads its wings to their full extent, not in a gesture of flight, but rather as a grand gesture of welcome, like a person extending their arms in a hug. It stands amidst a lush green meadow dotted with simple white and yellow flowers. The background features a perfectly clear blue sky with a few stylized, fluffy white clouds, all bathed in bright, cheerful sunlight, creating a warm, high-key lighting effect with minimal, soft shadows. The entire scene exudes an atmosphere of pure, innocent joy and open friendliness.", + "seed": 32678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_1/index_783_seed_32678_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009850105387158692, + "vlm_logprob": -0.009850105387158692, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.009850105387158692, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02411591447889805 + ], + "opposite_logprob_tokens": [ + -3.774115800857544 + ], + "logprob_reward": -0.02411591447889805, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02411591447889805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1064920574426651, + 0.06233152002096176, + 0.1021537333726883, + 0.0637088492512703, + 0.054488249123096466, + 0.16574527323246002, + 0.1916237771511078, + 0.21278177201747894, + 0.01947268284857273, + 0.0212020892649889 + ], + "credit_map": [ + 0.013701128773391247, + 0.011730887927114964, + 0.012634659186005592, + 0.014029123820364475, + 0.01845233142375946, + 0.011498472653329372, + 0.012651734054088593, + 0.011793715879321098, + 0.007726529147475958, + 0.008994756266474724, + 0.004791966639459133, + 0.0047621228732168674, + 0.015483612194657326, + 0.008680175989866257, + 0.007308935280889273, + 0.004583416040986776, + 0.015935400500893593, + 0.017012769356369972, + 0.0071841515600681305, + 0.00805988535284996, + 0.01530341524630785, + 0.015014086849987507, + 0.01783161610364914, + 0.005812416318804026, + 0.011444365605711937, + 0.009587353095412254, + 0.004830265883356333, + 0.005571647547185421, + 0.008010569959878922, + 0.010201920755207539, + 0.008579700253903866, + 0.0054830266162753105, + 0.009011785499751568, + 0.006281469948589802, + 0.004644442815333605, + 0.0045637208968400955, + 0.007860030047595501, + 0.008795236237347126, + 0.007432510145008564, + 0.0058990526013076305, + 0.018497902899980545, + 0.017834387719631195, + 0.018629498779773712, + 0.018505357205867767, + 0.031211767345666885, + 0.019505556672811508, + 0.02840081974864006, + 0.013159982860088348, + 0.019707245752215385, + 0.018664907664060593, + 0.016107937321066856, + 0.018654413521289825, + 0.02870633266866207, + 0.027565691620111465, + 0.049153875559568405, + 0.013063373975455761, + 0.025304550305008888, + 0.023028692230582237, + 0.020718660205602646, + 0.030418645590543747, + 0.02112749218940735, + 0.03366902470588684, + 0.03330124542117119, + 0.02521345764398575, + 0.0028196831699460745, + 0.002750190906226635, + 0.0017737572779878974, + 0.0021892318036407232, + 0.0024260191712528467, + 0.0026292544789612293, + 0.0031194875482469797, + 0.0017650594236329198, + 0.0038357728626579046, + 0.0035796817392110825, + 0.0023854230530560017, + 0.0021084623876959085, + 0.00301568116992712, + 0.002910517854616046, + 0.0020036923233419657, + 0.0013628582237288356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.02411591447889805 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001944795367307961 + ], + "opposite_logprob_tokens": [ + -6.2519450187683105 + ], + "logprob_reward": -0.001944795367307961, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001944795367307961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15210598707199097, + 0.08680476248264313, + 0.10246988385915756, + 0.04737979173660278, + 0.06967490911483765, + 0.16034731268882751, + 0.14086389541625977, + 0.11420828104019165, + 0.03794946148991585, + 0.08819571882486343 + ], + "credit_map": [ + 0.0184786356985569, + 0.022989638149738312, + 0.01635158620774746, + 0.016606969758868217, + 0.02390376105904579, + 0.019178936257958412, + 0.023118354380130768, + 0.01147810835391283, + 0.010006421245634556, + 0.011332323774695396, + 0.00886995904147625, + 0.006578707601875067, + 0.017527449876070023, + 0.013086595572531223, + 0.012764719314873219, + 0.0066385772079229355, + 0.021244056522846222, + 0.014250322245061398, + 0.00808578822761774, + 0.007413998246192932, + 0.011784208007156849, + 0.019262051209807396, + 0.014338319189846516, + 0.006091130431741476, + 0.0044051106087863445, + 0.006244205404073, + 0.002773357555270195, + 0.004766132216900587, + 0.00879514031112194, + 0.00684091541916132, + 0.009494490921497345, + 0.004060436971485615, + 0.011364472098648548, + 0.0075873215682804585, + 0.0061529758386313915, + 0.0044868262484669685, + 0.011441541835665703, + 0.011190931312739849, + 0.009862384758889675, + 0.00758845079690218, + 0.017392437905073166, + 0.01787416636943817, + 0.013494335114955902, + 0.020044011995196342, + 0.02959410287439823, + 0.02071511559188366, + 0.03124312125146389, + 0.009990002028644085, + 0.022045373916625977, + 0.01736433245241642, + 0.010485704988241196, + 0.011253592558205128, + 0.017688244581222534, + 0.024502143263816833, + 0.028214583173394203, + 0.009309899061918259, + 0.01235076505690813, + 0.013037454336881638, + 0.013042849488556385, + 0.014737318269908428, + 0.014060488902032375, + 0.015666715800762177, + 0.017241723835468292, + 0.01407095231115818, + 0.00546667817980051, + 0.004920864012092352, + 0.003685558680444956, + 0.003456602105870843, + 0.004880833439528942, + 0.007811765652149916, + 0.0034054501447826624, + 0.004321706481277943, + 0.021618574857711792, + 0.01832941174507141, + 0.006041720509529114, + 0.009573016315698624, + 0.009060779586434364, + 0.011870610527694225, + 0.007140615489333868, + 0.004560976754873991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.001944795367307961 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028160461224615574 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.0028160461224615574, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028160461224615574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19822144508361816, + 0.06417115777730942, + 0.1045207530260086, + 0.05712776258587837, + 0.05242761969566345, + 0.11193489283323288, + 0.14443425834178925, + 0.146171435713768, + 0.045748356729745865, + 0.07524226605892181 + ], + "credit_map": [ + 0.03133349493145943, + 0.02958512119948864, + 0.0263240747153759, + 0.021348215639591217, + 0.02705053798854351, + 0.025723138824105263, + 0.018595291301608086, + 0.018261583521962166, + 0.00870103482156992, + 0.008031957782804966, + 0.005870735738426447, + 0.008903440088033676, + 0.010557480156421661, + 0.008389994502067566, + 0.008745355531573296, + 0.004971162416040897, + 0.01862671785056591, + 0.017970528453588486, + 0.011230194941163063, + 0.007033265195786953, + 0.012045301496982574, + 0.016832517459988594, + 0.013479096814990044, + 0.007303131278604269, + 0.007683856412768364, + 0.008954115211963654, + 0.004330034367740154, + 0.005303090438246727, + 0.008720975369215012, + 0.007051101420074701, + 0.008571895770728588, + 0.006512693129479885, + 0.007164038717746735, + 0.007961266674101353, + 0.0033483768347650766, + 0.0036279200576245785, + 0.01086377166211605, + 0.007992876693606377, + 0.00794151984155178, + 0.0035278492141515017, + 0.009375452995300293, + 0.009793479926884174, + 0.012956839054822922, + 0.013327608816325665, + 0.019115988165140152, + 0.013406714424490929, + 0.024106280878186226, + 0.009852524846792221, + 0.015660705044865608, + 0.01687086559832096, + 0.01126108504831791, + 0.014299370348453522, + 0.024124206975102425, + 0.017729243263602257, + 0.03524244576692581, + 0.009246342815458775, + 0.014229920692741871, + 0.014760860241949558, + 0.01451821718364954, + 0.02455597184598446, + 0.017660992220044136, + 0.01754087395966053, + 0.02817237377166748, + 0.014732226729393005, + 0.006521656643599272, + 0.006546970456838608, + 0.0033138685394078493, + 0.004713438451290131, + 0.007892812602221966, + 0.006597052328288555, + 0.005162399727851152, + 0.005000153556466103, + 0.017650125548243523, + 0.013761845417320728, + 0.00645341444760561, + 0.008324994705617428, + 0.005270221270620823, + 0.016271306201815605, + 0.0034207915887236595, + 0.004089568741619587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0028160461224615574 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817641779780388 + ], + "opposite_logprob_tokens": [ + -4.018176555633545 + ], + "logprob_reward": -0.01817641779780388, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817641779780388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15620742738246918, + 0.09887316823005676, + 0.1490548551082611, + 0.1022416278719902, + 0.0312026496976614, + 0.08057894557714462, + 0.21550054848194122, + 0.07327555865049362, + 0.038586679846048355, + 0.05447857454419136 + ], + "credit_map": [ + 0.02070283517241478, + 0.01798226684331894, + 0.019624853506684303, + 0.013187323696911335, + 0.0235082246363163, + 0.016344865784049034, + 0.0264753270894289, + 0.018381722271442413, + 0.01130578201264143, + 0.01109330728650093, + 0.016189904883503914, + 0.009406187571585178, + 0.016321806237101555, + 0.012968414463102818, + 0.010111285373568535, + 0.011476471088826656, + 0.026592567563056946, + 0.018244052305817604, + 0.012695561163127422, + 0.01645212434232235, + 0.020008739084005356, + 0.023791061714291573, + 0.014261787757277489, + 0.0170089490711689, + 0.016680147498846054, + 0.019554441794753075, + 0.007689806167036295, + 0.010502731427550316, + 0.014396446757018566, + 0.013984703458845615, + 0.00928991474211216, + 0.010143430903553963, + 0.006017015315592289, + 0.006127652712166309, + 0.0020102837588638067, + 0.0025788615457713604, + 0.0037961944472044706, + 0.00528790382668376, + 0.0032969212625175714, + 0.0020878135692328215, + 0.007718508131802082, + 0.008567600511014462, + 0.008114641532301903, + 0.012226574122905731, + 0.013538974337279797, + 0.009555390104651451, + 0.01401300448924303, + 0.0068442365154623985, + 0.023745937272906303, + 0.015981752425432205, + 0.02533789351582527, + 0.024808259680867195, + 0.03326556086540222, + 0.021468698978424072, + 0.049593519419431686, + 0.021298909559845924, + 0.006676968652755022, + 0.006750892382115126, + 0.005548280198127031, + 0.012244760058820248, + 0.00942754466086626, + 0.006946518085896969, + 0.0162437055259943, + 0.00943688303232193, + 0.00518658384680748, + 0.008664043620228767, + 0.0032202547881752253, + 0.0032860771752893925, + 0.0026799170300364494, + 0.010442522354424, + 0.0026500073727220297, + 0.002457269700244069, + 0.013773493468761444, + 0.009908151812851429, + 0.003034912282600999, + 0.006254233419895172, + 0.005565210711210966, + 0.008979992009699345, + 0.003712895093485713, + 0.0032496810890734196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.01817641779780388 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002197353169322014 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.002197353169322014, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002197353169322014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14453011751174927, + 0.08187682926654816, + 0.09320998936891556, + 0.029317794367671013, + 0.0366453193128109, + 0.11075560003519058, + 0.17848041653633118, + 0.10855462402105331, + 0.13618935644626617, + 0.0804399773478508 + ], + "credit_map": [ + 0.013707493431866169, + 0.016450222581624985, + 0.009876050055027008, + 0.016622187569737434, + 0.03239821270108223, + 0.018797045573592186, + 0.02714746631681919, + 0.009531435556709766, + 0.009179259650409222, + 0.011317222379148006, + 0.007505588699132204, + 0.006719400640577078, + 0.01838044263422489, + 0.011560478247702122, + 0.011629030108451843, + 0.005585408769547939, + 0.013741726987063885, + 0.017036186531186104, + 0.005536330863833427, + 0.007312290836125612, + 0.01408564392477274, + 0.013706333935260773, + 0.01548124197870493, + 0.006310228258371353, + 0.004923080559819937, + 0.005169564392417669, + 0.0027581031899899244, + 0.0026097935624420643, + 0.004729080479592085, + 0.004096965305507183, + 0.003151260083541274, + 0.001879946910776198, + 0.005998122971504927, + 0.005906327627599239, + 0.0033190620597451925, + 0.002599150873720646, + 0.00638005044311285, + 0.0052194842137396336, + 0.0041752224788069725, + 0.0030478988774120808, + 0.012836926616728306, + 0.01208056602627039, + 0.010868352837860584, + 0.012785233557224274, + 0.021185753867030144, + 0.015067710541188717, + 0.01806553825736046, + 0.007865517400205135, + 0.02996876835823059, + 0.024415621533989906, + 0.0138764139264822, + 0.013057347387075424, + 0.022913310676813126, + 0.03825807198882103, + 0.02504781074821949, + 0.01094305980950594, + 0.01124673429876566, + 0.012253323569893837, + 0.011669478379189968, + 0.015912992879748344, + 0.013645399361848831, + 0.014880361035466194, + 0.017364513128995895, + 0.011581819504499435, + 0.02257266454398632, + 0.029040062800049782, + 0.009571533650159836, + 0.008625788614153862, + 0.01041707769036293, + 0.0409606471657753, + 0.008190713822841644, + 0.0068108621053397655, + 0.02008792571723461, + 0.016291920095682144, + 0.004531291779130697, + 0.0087564243003726, + 0.008160310797393322, + 0.013981206342577934, + 0.004447450861334801, + 0.004183449316769838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.002197353169322014 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, held at eye level, features a plump, cartoonish blue bird, rendered in a classic 2D animation style with crisp lines and vibrant colors. The bird, with its bright sky-blue feathers and soft, creamy-white belly, stands cheerfully in the center of the frame. Its large, round, black eyes sparkle with friendly warmth, and its small, triangular yellow beak opens in a broad, joyful smile. The bird enthusiastically puffs out its chest and spreads its wings to their full extent, not in a gesture of flight, but rather as a grand gesture of welcome, like a person extending their arms in a hug. It stands amidst a lush green meadow dotted with simple white and yellow flowers. The background features a perfectly clear blue sky with a few stylized, fluffy white clouds, all bathed in bright, cheerful sunlight, creating a warm, high-key lighting effect with minimal, soft shadows. The entire scene exudes an atmosphere of pure, innocent joy and open friendliness.", + "seed": 32679, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_1/index_783_seed_32679_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.005187255749478936, + "vlm_logprob": -0.005187255749478936, + "vlm_yes_no_margin": 5.75, + "vlm_token_credit": -0.005187255749478936, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025327291805297136 + ], + "opposite_logprob_tokens": [ + -6.377532958984375 + ], + "logprob_reward": -0.0025327291805297136, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0025327291805297136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18031546473503113, + 0.09547015279531479, + 0.09789364784955978, + 0.07564330101013184, + 0.3088781237602234, + 0.09577988088130951, + 0.053359683603048325, + 0.02010420523583889, + 0.027141237631440163, + 0.045414332300424576 + ], + "credit_map": [ + 0.021799355745315552, + 0.0344887375831604, + 0.021653546020388603, + 0.022305909544229507, + 0.012561806477606297, + 0.02598574571311474, + 0.026825739070773125, + 0.014694631099700928, + 0.006925890687853098, + 0.026112748309969902, + 0.005243236664682627, + 0.005623792763799429, + 0.010567877441644669, + 0.026428204029798508, + 0.008764647878706455, + 0.005803755018860102, + 0.012049635872244835, + 0.026424802839756012, + 0.005769538227468729, + 0.008188122883439064, + 0.016619345173239708, + 0.011941527016460896, + 0.011575020849704742, + 0.0053256540559232235, + 0.00816468894481659, + 0.007135912775993347, + 0.00795309990644455, + 0.01017740461975336, + 0.012422261759638786, + 0.008826390840113163, + 0.014738011173903942, + 0.006225530058145523, + 0.04589470475912094, + 0.07397551834583282, + 0.016030292958021164, + 0.016238389536738396, + 0.03345820680260658, + 0.0820421651005745, + 0.021413015201687813, + 0.019825834780931473, + 0.009158424101769924, + 0.009918836876749992, + 0.006490603089332581, + 0.008853447623550892, + 0.013717694208025932, + 0.01431206613779068, + 0.027216162532567978, + 0.006112650502473116, + 0.0065915994346141815, + 0.003948455210775137, + 0.008968528360128403, + 0.00633180933073163, + 0.005033399909734726, + 0.006398719735443592, + 0.009552773088216782, + 0.006534395273774862, + 0.002082637744024396, + 0.002150144660845399, + 0.0016024400247260928, + 0.0021485353354364634, + 0.0025194454938173294, + 0.002233925275504589, + 0.005181066691875458, + 0.0021860103588551283, + 0.003300125477835536, + 0.0031323451548814774, + 0.0044256458058953285, + 0.002533992752432823, + 0.0031639880035072565, + 0.0028913712594658136, + 0.004206348210573196, + 0.0034874205011874437, + 0.004736762028187513, + 0.004462323151528835, + 0.0025806757621467113, + 0.004223475232720375, + 0.00731590436771512, + 0.005177251994609833, + 0.014348643831908703, + 0.002569297095760703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.0025327291805297136 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013390155509114265 + ], + "opposite_logprob_tokens": [ + -6.626338958740234 + ], + "logprob_reward": -0.0013390155509114265, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013390155509114265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18699423968791962, + 0.13080455362796783, + 0.12356813251972198, + 0.10575956106185913, + 0.0896715596318245, + 0.052164144814014435, + 0.06759348511695862, + 0.09746935963630676, + 0.07253029197454453, + 0.07344464212656021 + ], + "credit_map": [ + 0.026270439848303795, + 0.03234728053212166, + 0.02185656875371933, + 0.019064249470829964, + 0.022664295509457588, + 0.022290050983428955, + 0.026872286573052406, + 0.015629077330231667, + 0.010215451009571552, + 0.03058204986155033, + 0.006335515528917313, + 0.007753163110464811, + 0.018612295389175415, + 0.03161776065826416, + 0.017760396003723145, + 0.007927929051220417, + 0.015015206299722195, + 0.016162138432264328, + 0.01129323709756136, + 0.011000491678714752, + 0.025661755353212357, + 0.019046682864427567, + 0.019446371123194695, + 0.0059422641061246395, + 0.012664630077779293, + 0.008555132895708084, + 0.01769930310547352, + 0.014388692565262318, + 0.011524691246449947, + 0.012822628021240234, + 0.016426555812358856, + 0.011677931062877178, + 0.008833172731101513, + 0.01387932151556015, + 0.005023759789764881, + 0.006796800531446934, + 0.013598273508250713, + 0.017849020659923553, + 0.01553585845977068, + 0.00815536081790924, + 0.0061000739224255085, + 0.006450160406529903, + 0.004138241056352854, + 0.004696204327046871, + 0.006726647261530161, + 0.006480662617832422, + 0.00982645433396101, + 0.007745707407593727, + 0.009292337112128735, + 0.007260455749928951, + 0.005056879948824644, + 0.006079491227865219, + 0.009052514098584652, + 0.007619209587574005, + 0.019886652007699013, + 0.0033459491096436977, + 0.012250404804944992, + 0.00921076163649559, + 0.010593422688543797, + 0.009735013358294964, + 0.011148998513817787, + 0.01247719582170248, + 0.02480495721101761, + 0.007248610258102417, + 0.011305224150419235, + 0.006447249092161655, + 0.00452331593260169, + 0.005462986882776022, + 0.008939455263316631, + 0.007304109167307615, + 0.019216390326619148, + 0.009331569075584412, + 0.009938286617398262, + 0.010546276345849037, + 0.004483519122004509, + 0.006686616223305464, + 0.011024261824786663, + 0.008930400013923645, + 0.01806981861591339, + 0.0037654698826372623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0013390155509114265 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190072951838374 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.003190072951838374, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190072951838374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21066215634346008, + 0.20558328926563263, + 0.13194507360458374, + 0.086534284055233, + 0.06719893217086792, + 0.04447077959775925, + 0.09310257434844971, + 0.07213670015335083, + 0.048389334231615067, + 0.039976853877305984 + ], + "credit_map": [ + 0.029690401628613472, + 0.03353061154484749, + 0.02258516289293766, + 0.030231885612010956, + 0.02214314416050911, + 0.025917887687683105, + 0.03200697526335716, + 0.014556103385984898, + 0.019684432074427605, + 0.051841575652360916, + 0.016724547371268272, + 0.015395522117614746, + 0.018501918762922287, + 0.046522945165634155, + 0.0208527073264122, + 0.016059638932347298, + 0.014880643226206303, + 0.027135318145155907, + 0.014289875514805317, + 0.014282395131886005, + 0.014687000773847103, + 0.01412087120115757, + 0.013613488525152206, + 0.01893547922372818, + 0.011760568246245384, + 0.010457310825586319, + 0.010176929645240307, + 0.010275784879922867, + 0.013737629167735577, + 0.009870118461549282, + 0.01083833072334528, + 0.009417612105607986, + 0.008462491445243359, + 0.01238677091896534, + 0.004585612565279007, + 0.005170583724975586, + 0.00855452660471201, + 0.011447945609688759, + 0.009104705415666103, + 0.007486293558031321, + 0.004492825362831354, + 0.0056753456592559814, + 0.0034008596558123827, + 0.004228231497108936, + 0.006430594716221094, + 0.005493375938385725, + 0.011315718293190002, + 0.003433824982494116, + 0.011365465819835663, + 0.0073778992518782616, + 0.007338767405599356, + 0.009151937440037727, + 0.010220604948699474, + 0.010260023176670074, + 0.02742645889520645, + 0.009961423464119434, + 0.008275485597550869, + 0.006930232048034668, + 0.007812071591615677, + 0.0060895937494933605, + 0.009878004901111126, + 0.012723185122013092, + 0.016150109469890594, + 0.004278015810996294, + 0.005270793568342924, + 0.005170122254639864, + 0.005707398522645235, + 0.005419465247541666, + 0.007663901429623365, + 0.0055697704665362835, + 0.010001536458730698, + 0.003586347447708249, + 0.0072200000286102295, + 0.006342127453535795, + 0.0028316862881183624, + 0.0037319534458220005, + 0.0073088109493255615, + 0.004816488828510046, + 0.00522210169583559, + 0.002503682393580675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.003190072951838374 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016056988388299942 + ], + "opposite_logprob_tokens": [ + -4.141057014465332 + ], + "logprob_reward": -0.016056988388299942, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016056988388299942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13192273676395416, + 0.16107094287872314, + 0.14658775925636292, + 0.14369995892047882, + 0.028795918449759483, + 0.0314268097281456, + 0.03399403393268585, + 0.0785399079322815, + 0.041494473814964294, + 0.2024674415588379 + ], + "credit_map": [ + 0.01825372502207756, + 0.02019999362528324, + 0.011388758197426796, + 0.01191079430282116, + 0.01694212481379509, + 0.016306955367326736, + 0.02694942243397236, + 0.00997096486389637, + 0.013483616523444653, + 0.024175280705094337, + 0.010726790875196457, + 0.011424344964325428, + 0.02909589745104313, + 0.029006026685237885, + 0.030676936730742455, + 0.01248204056173563, + 0.018065860494971275, + 0.024821804836392403, + 0.014802427031099796, + 0.01390385814011097, + 0.026952868327498436, + 0.020040025934576988, + 0.020010320469737053, + 0.00799059122800827, + 0.01602255180478096, + 0.014759520068764687, + 0.011737925931811333, + 0.016191741451621056, + 0.023493971675634384, + 0.015509173274040222, + 0.03626960515975952, + 0.009715463034808636, + 0.003273535519838333, + 0.004092284478247166, + 0.0026235305704176426, + 0.0032423443626612425, + 0.004884278867393732, + 0.0029727350920438766, + 0.004991509951651096, + 0.0027157014701515436, + 0.003401766298338771, + 0.0039291721768677235, + 0.0040129926055669785, + 0.0031199813820421696, + 0.004713895730674267, + 0.003932165913283825, + 0.004924316890537739, + 0.0033925212919712067, + 0.004772979766130447, + 0.0047429646365344524, + 0.0027914431411772966, + 0.0029802119825035334, + 0.003710969816893339, + 0.004409449640661478, + 0.0072968523018062115, + 0.003289160318672657, + 0.009732397273182869, + 0.007406984455883503, + 0.007827573455870152, + 0.0072420453652739525, + 0.010356279090046883, + 0.011062035337090492, + 0.020094426348805428, + 0.00481816753745079, + 0.0060034519992768764, + 0.004088948480784893, + 0.004340045154094696, + 0.003913777880370617, + 0.004919991362839937, + 0.004631748888641596, + 0.0066697378642857075, + 0.006926773115992546, + 0.02352777123451233, + 0.020971983671188354, + 0.011032328940927982, + 0.018356313928961754, + 0.028562719002366066, + 0.02195145934820175, + 0.06663715094327927, + 0.011427725665271282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.016056988388299942 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028174726758152246 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028174726758152246, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028174726758152246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14098268747329712, + 0.12488006055355072, + 0.0829082727432251, + 0.09207016229629517, + 0.11969545483589172, + 0.07489429414272308, + 0.06098935008049011, + 0.05464302748441696, + 0.08274943381547928, + 0.16618728637695312 + ], + "credit_map": [ + 0.02105511538684368, + 0.026027942076325417, + 0.01859787479043007, + 0.01673634722828865, + 0.012767468579113483, + 0.01753864623606205, + 0.01727340556681156, + 0.010985900647938251, + 0.011190451681613922, + 0.028064711019396782, + 0.008924715220928192, + 0.009482396766543388, + 0.015468376688659191, + 0.02763316035270691, + 0.016749659553170204, + 0.00736659299582243, + 0.007622021250426769, + 0.01339442003518343, + 0.007316044997423887, + 0.006884903181344271, + 0.018575765192508698, + 0.008820615708827972, + 0.016340846195816994, + 0.00395365571603179, + 0.011247829534113407, + 0.0097501864656806, + 0.009915296919643879, + 0.008520714938640594, + 0.011706625111401081, + 0.012181006371974945, + 0.020401278510689735, + 0.008347230963408947, + 0.018917666748166084, + 0.02219640463590622, + 0.006061087362468243, + 0.007928543724119663, + 0.013916440308094025, + 0.02259213849902153, + 0.017367104068398476, + 0.010716073215007782, + 0.008620785549283028, + 0.010774146765470505, + 0.004865758121013641, + 0.005781557410955429, + 0.009071934968233109, + 0.012903072871267796, + 0.018568960949778557, + 0.004308079835027456, + 0.00647850101813674, + 0.004139801952987909, + 0.007592902984470129, + 0.006110950838774443, + 0.006889515556395054, + 0.00621004169806838, + 0.014816820621490479, + 0.008750815875828266, + 0.006248547695577145, + 0.00581790367141366, + 0.00402700062841177, + 0.004627687856554985, + 0.008778784424066544, + 0.007197904866188765, + 0.013313482515513897, + 0.00463171536102891, + 0.010950843803584576, + 0.007296191528439522, + 0.005979582667350769, + 0.00716974725946784, + 0.011866599321365356, + 0.009129202924668789, + 0.02324233390390873, + 0.007114932406693697, + 0.019773252308368683, + 0.016056552529335022, + 0.013843115419149399, + 0.015210249461233616, + 0.022173060104250908, + 0.01853189431130886, + 0.05156981945037842, + 0.009029341861605644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.0028174726758152246 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, held at eye level, features a plump, cartoonish blue bird, rendered in a classic 2D animation style with crisp lines and vibrant colors. The bird, with its bright sky-blue feathers and soft, creamy-white belly, stands cheerfully in the center of the frame. Its large, round, black eyes sparkle with friendly warmth, and its small, triangular yellow beak opens in a broad, joyful smile. The bird enthusiastically puffs out its chest and spreads its wings to their full extent, not in a gesture of flight, but rather as a grand gesture of welcome, like a person extending their arms in a hug. It stands amidst a lush green meadow dotted with simple white and yellow flowers. The background features a perfectly clear blue sky with a few stylized, fluffy white clouds, all bathed in bright, cheerful sunlight, creating a warm, high-key lighting effect with minimal, soft shadows. The entire scene exudes an atmosphere of pure, innocent joy and open friendliness.", + "seed": 32680, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_1/index_783_seed_32680_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.006149520562030375, + "vlm_logprob": -0.006149520562030375, + "vlm_yes_no_margin": 5.575, + "vlm_token_credit": -0.006149520562030375, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028973526787012815 + ], + "opposite_logprob_tokens": [ + -6.377897262573242 + ], + "logprob_reward": -0.0028973526787012815, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0028973526787012815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2901131808757782, + 0.143132746219635, + 0.07125871628522873, + 0.05752900615334511, + 0.04064636677503586, + 0.12159238010644913, + 0.0794723704457283, + 0.07517760246992111, + 0.04858965426683426, + 0.07248803228139877 + ], + "credit_map": [ + 0.03284402936697006, + 0.02630607970058918, + 0.022181518375873566, + 0.014544004574418068, + 0.04283200949430466, + 0.03582042455673218, + 0.03285960853099823, + 0.08272548764944077, + 0.021480701863765717, + 0.02092992141842842, + 0.012637014500796795, + 0.012581493705511093, + 0.025995485484600067, + 0.020370090380311012, + 0.016680601984262466, + 0.012457420118153095, + 0.008728638291358948, + 0.009818632155656815, + 0.005841746926307678, + 0.007169949822127819, + 0.010239430703222752, + 0.011500056833028793, + 0.010981425642967224, + 0.006978824734687805, + 0.007800415623933077, + 0.009932885877788067, + 0.005145360250025988, + 0.005055941175669432, + 0.0066098920069634914, + 0.010569258593022823, + 0.005795537494122982, + 0.0066197095438838005, + 0.006921734660863876, + 0.005875609815120697, + 0.0036577151622623205, + 0.0050902036018669605, + 0.00622905045747757, + 0.004309725482016802, + 0.005829217378050089, + 0.0027331095188856125, + 0.011241037398576736, + 0.014785395003855228, + 0.013028187677264214, + 0.027578771114349365, + 0.018126463517546654, + 0.013059375807642937, + 0.01208513230085373, + 0.011688009835779667, + 0.007543761283159256, + 0.011913060210645199, + 0.008749129250645638, + 0.008186169900000095, + 0.010150441899895668, + 0.012662283144891262, + 0.014393524266779423, + 0.005873996298760176, + 0.013080763630568981, + 0.013523109257221222, + 0.006572348531335592, + 0.006408224813640118, + 0.007837331853806973, + 0.012178996577858925, + 0.011256790719926357, + 0.004320038948208094, + 0.005782762076705694, + 0.006763197481632233, + 0.0024468277115374804, + 0.0047561307437717915, + 0.009399745613336563, + 0.005911937914788723, + 0.010456099174916744, + 0.003072952851653099, + 0.01049781497567892, + 0.013376773335039616, + 0.00369960255920887, + 0.0053605311550199986, + 0.007359675597399473, + 0.018444664776325226, + 0.0052530476823449135, + 0.008495915681123734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0028973526787012815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019427727675065398 + ], + "opposite_logprob_tokens": [ + -6.2519426345825195 + ], + "logprob_reward": -0.0019427727675065398, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019427727675065398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17576679587364197, + 0.11439929157495499, + 0.0769771933555603, + 0.06084870547056198, + 0.05081384256482124, + 0.10068900138139725, + 0.11821973323822021, + 0.07485101372003555, + 0.08959691226482391, + 0.1378374993801117 + ], + "credit_map": [ + 0.015086553990840912, + 0.01562982052564621, + 0.01770302653312683, + 0.010994729585945606, + 0.025508573278784752, + 0.022693198174238205, + 0.018289431929588318, + 0.049861468374729156, + 0.012412737123668194, + 0.01695791631937027, + 0.009536991827189922, + 0.009612069465219975, + 0.026934636756777763, + 0.017319953069090843, + 0.014846433885395527, + 0.006778566166758537, + 0.00991352740675211, + 0.011385666206479073, + 0.008752236142754555, + 0.006056154612451792, + 0.010490025393664837, + 0.013224207796156406, + 0.011374798603355885, + 0.005780588369816542, + 0.006385389715433121, + 0.009018727578222752, + 0.006116417236626148, + 0.006687157321721315, + 0.012057018466293812, + 0.007794750854372978, + 0.008821815252304077, + 0.0039674355648458, + 0.00841884408146143, + 0.007713711820542812, + 0.005545707885175943, + 0.0035479795187711716, + 0.008931084536015987, + 0.006546166725456715, + 0.005982944741845131, + 0.004127408843487501, + 0.011762773618102074, + 0.012928192503750324, + 0.01130121760070324, + 0.014853045344352722, + 0.016048606485128403, + 0.009801140055060387, + 0.013727427460253239, + 0.010266605764627457, + 0.016289127990603447, + 0.01692168414592743, + 0.011300947517156601, + 0.01253653783351183, + 0.014819149859249592, + 0.016011126339435577, + 0.022227495908737183, + 0.00811366830021143, + 0.011767040938138962, + 0.011601178906857967, + 0.007393372245132923, + 0.00772868562489748, + 0.009105481207370758, + 0.011770131066441536, + 0.011038260534405708, + 0.004446865525096655, + 0.0108269602060318, + 0.02741926535964012, + 0.005023489706218243, + 0.005094544496387243, + 0.007093296851962805, + 0.025948796421289444, + 0.0048688678070902824, + 0.0033216960728168488, + 0.01881108619272709, + 0.02558983862400055, + 0.011268559843301773, + 0.009519562125205994, + 0.009522191248834133, + 0.0367421992123127, + 0.01029633916914463, + 0.016087723895907402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.0019427727675065398 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052475095726549625 + ], + "opposite_logprob_tokens": [ + -5.255247592926025 + ], + "logprob_reward": -0.0052475095726549625, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052475095726549625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20548367500305176, + 0.11122814565896988, + 0.09068601578474045, + 0.09012541174888611, + 0.07319796085357666, + 0.1259586215019226, + 0.09031873941421509, + 0.07352161407470703, + 0.03651636093854904, + 0.10296336561441422 + ], + "credit_map": [ + 0.019265899434685707, + 0.02716672420501709, + 0.015959981828927994, + 0.020883124321699142, + 0.04998162016272545, + 0.029890233650803566, + 0.02738882228732109, + 0.014947277493774891, + 0.012947652488946915, + 0.014428717084228992, + 0.007531262002885342, + 0.012218022719025612, + 0.019039124250411987, + 0.017729533836245537, + 0.016129447147250175, + 0.011204375885426998, + 0.01777183823287487, + 0.01391387078911066, + 0.010040373541414738, + 0.011793314479291439, + 0.007807574700564146, + 0.011831696145236492, + 0.008595620281994343, + 0.008931723423302174, + 0.00942936260253191, + 0.015520303510129452, + 0.008083542808890343, + 0.006713664624840021, + 0.009033586829900742, + 0.02085774391889572, + 0.010177061893045902, + 0.010310150682926178, + 0.011169947683811188, + 0.008980349637567997, + 0.0072945402935147285, + 0.010195121169090271, + 0.009225519374012947, + 0.008256771601736546, + 0.008728002198040485, + 0.009347707964479923, + 0.012706661596894264, + 0.015602052211761475, + 0.013134332373738289, + 0.025819070637226105, + 0.02034853771328926, + 0.01400226540863514, + 0.01418881956487894, + 0.010156871750950813, + 0.010491559281945229, + 0.014497597701847553, + 0.008766290731728077, + 0.007801922038197517, + 0.0118232611566782, + 0.015949785709381104, + 0.014758444391191006, + 0.006229878403246403, + 0.011273970827460289, + 0.011506480164825916, + 0.0073794289492070675, + 0.007894346490502357, + 0.00796009972691536, + 0.011627289466559887, + 0.0098120654001832, + 0.006067927461117506, + 0.005711627192795277, + 0.005696282722055912, + 0.0038661248981952667, + 0.003758990205824375, + 0.004650011193007231, + 0.00582487927749753, + 0.004837571643292904, + 0.002170875668525696, + 0.017107607796788216, + 0.019140098243951797, + 0.007946915924549103, + 0.006908020470291376, + 0.009029204957187176, + 0.02527523785829544, + 0.005917285103350878, + 0.011638998985290527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0052475095726549625 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018173256888985634 + ], + "opposite_logprob_tokens": [ + -4.0181732177734375 + ], + "logprob_reward": -0.018173256888985634, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018173256888985634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15669278800487518, + 0.09486813843250275, + 0.06126519292593002, + 0.06209927797317505, + 0.07967326790094376, + 0.10735903680324554, + 0.08115290105342865, + 0.040920890867710114, + 0.17956919968128204, + 0.13639934360980988 + ], + "credit_map": [ + 0.015031145885586739, + 0.017227545380592346, + 0.012872635386884212, + 0.01176510564982891, + 0.02903287671506405, + 0.019461384043097496, + 0.019327130168676376, + 0.031974948942661285, + 0.011748860590159893, + 0.01345367543399334, + 0.006312791258096695, + 0.009095447137951851, + 0.019811630249023438, + 0.01594545692205429, + 0.012871700339019299, + 0.005628577433526516, + 0.008376742713153362, + 0.007482921238988638, + 0.005994447506964207, + 0.007322205696254969, + 0.009809724055230618, + 0.007479759398847818, + 0.01025290135294199, + 0.004546484909951687, + 0.007370289880782366, + 0.008398576639592648, + 0.005025339778512716, + 0.005640614777803421, + 0.014153405092656612, + 0.006627237889915705, + 0.010934863239526749, + 0.003948946949094534, + 0.00984448567032814, + 0.012150315567851067, + 0.006049314513802528, + 0.008303161710500717, + 0.015725528821349144, + 0.008503809571266174, + 0.013201486319303513, + 0.005895163863897324, + 0.00941616389900446, + 0.012657128274440765, + 0.009992607869207859, + 0.023345308378338814, + 0.016923461109399796, + 0.011831107549369335, + 0.013777334243059158, + 0.009415914304554462, + 0.013275708071887493, + 0.011473085731267929, + 0.007945160381495953, + 0.009186133742332458, + 0.010080902837216854, + 0.010474451817572117, + 0.01344603206962347, + 0.005271414760500193, + 0.0037938759196549654, + 0.004771169740706682, + 0.004687332082539797, + 0.005438508000224829, + 0.006805725395679474, + 0.004792692139744759, + 0.007542470004409552, + 0.003089117119088769, + 0.02078021503984928, + 0.029914524406194687, + 0.01129230484366417, + 0.018305251374840736, + 0.029286900535225868, + 0.025059841573238373, + 0.03438832610845566, + 0.010541819967329502, + 0.021237242966890335, + 0.02636907808482647, + 0.009917140007019043, + 0.010057198815047741, + 0.012870348989963531, + 0.0333358459174633, + 0.008519764989614487, + 0.014092717319726944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.018173256888985634 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024867109023034573 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.0024867109023034573, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024867109023034573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19769607484340668, + 0.09424033015966415, + 0.052195511758327484, + 0.06875451654195786, + 0.04077368602156639, + 0.07449666410684586, + 0.09405362606048584, + 0.0951533243060112, + 0.06127076596021652, + 0.22136548161506653 + ], + "credit_map": [ + 0.020970111712813377, + 0.02237737737596035, + 0.0186868067830801, + 0.01435874868184328, + 0.026709189638495445, + 0.027529630810022354, + 0.02138201706111431, + 0.045682188123464584, + 0.011576212011277676, + 0.015533294528722763, + 0.009123913012444973, + 0.008416635915637016, + 0.014035336673259735, + 0.017460595816373825, + 0.011606923304498196, + 0.006487422157078981, + 0.008528475649654865, + 0.007517850026488304, + 0.005396015010774136, + 0.00572266336530447, + 0.006104040425270796, + 0.0069190505892038345, + 0.006969294976443052, + 0.005038122646510601, + 0.0072113689966499805, + 0.011244653724133968, + 0.006386299151927233, + 0.00595066137611866, + 0.007501725107431412, + 0.016345173120498657, + 0.007555905729532242, + 0.006558725144714117, + 0.006197415757924318, + 0.006964110303670168, + 0.0037856893613934517, + 0.003477885155007243, + 0.0055016763508319855, + 0.0072145406156778336, + 0.005056113936007023, + 0.002576253144070506, + 0.008589720353484154, + 0.009570296853780746, + 0.008575093932449818, + 0.013147213496267796, + 0.011742438189685345, + 0.008409068919718266, + 0.00813818909227848, + 0.006324639078229666, + 0.010105074383318424, + 0.013126485981047153, + 0.010712693445384502, + 0.00992948841303587, + 0.01222629938274622, + 0.01421702653169632, + 0.016869323328137398, + 0.006867235526442528, + 0.016361575573682785, + 0.016064569354057312, + 0.0078109754249453545, + 0.00869553443044424, + 0.00980458129197359, + 0.015026199631392956, + 0.016153449192643166, + 0.00523644033819437, + 0.004903567023575306, + 0.012315567582845688, + 0.0029511626344174147, + 0.0050164079293608665, + 0.008797286078333855, + 0.012098362669348717, + 0.01192998792976141, + 0.003258420852944255, + 0.032638754695653915, + 0.04597447067499161, + 0.013886341825127602, + 0.011921025812625885, + 0.013376072980463505, + 0.06585343927145004, + 0.009837362915277481, + 0.027878008782863617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.0024867109023034573 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, held at eye level, features a plump, cartoonish blue bird, rendered in a classic 2D animation style with crisp lines and vibrant colors. The bird, with its bright sky-blue feathers and soft, creamy-white belly, stands cheerfully in the center of the frame. Its large, round, black eyes sparkle with friendly warmth, and its small, triangular yellow beak opens in a broad, joyful smile. The bird enthusiastically puffs out its chest and spreads its wings to their full extent, not in a gesture of flight, but rather as a grand gesture of welcome, like a person extending their arms in a hug. It stands amidst a lush green meadow dotted with simple white and yellow flowers. The background features a perfectly clear blue sky with a few stylized, fluffy white clouds, all bathed in bright, cheerful sunlight, creating a warm, high-key lighting effect with minimal, soft shadows. The entire scene exudes an atmosphere of pure, innocent joy and open friendliness.", + "seed": 32681, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_1/index_783_seed_32681_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.006438313936814666, + "vlm_logprob": -0.006438313936814666, + "vlm_yes_no_margin": 5.45, + "vlm_token_credit": -0.006438313936814666, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005696847103536129 + ], + "opposite_logprob_tokens": [ + -5.380696773529053 + ], + "logprob_reward": -0.005696847103536129, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.005696847103536129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45209476351737976, + 0.06365575641393661, + 0.023438865318894386, + 0.018025973811745644, + 0.028714507818222046, + 0.012430373579263687, + 0.3189684748649597, + 0.04172500595450401, + 0.01963220350444317, + 0.021314051002264023 + ], + "credit_map": [ + 0.10679493099451065, + 0.10595719516277313, + 0.042638007551431656, + 0.042002707719802856, + 0.04009104147553444, + 0.04233962669968605, + 0.033216167241334915, + 0.039055101573467255, + 0.013887319713830948, + 0.012449567206203938, + 0.007060324773192406, + 0.005018036812543869, + 0.006626012735068798, + 0.005999945104122162, + 0.007257580291479826, + 0.005356974434107542, + 0.0036328451242297888, + 0.003961898852139711, + 0.0025018530432134867, + 0.002706382889300585, + 0.003023627446964383, + 0.0020848566200584173, + 0.0031115449965000153, + 0.0024158575106412172, + 0.0022598186042159796, + 0.0030960862059146166, + 0.002230186015367508, + 0.0016655090730637312, + 0.002229198580607772, + 0.0010763905011117458, + 0.0030443891882896423, + 0.002424395177513361, + 0.0034548291005194187, + 0.0038114150520414114, + 0.004018542356789112, + 0.0026357725728303194, + 0.00423552468419075, + 0.0040045916102826595, + 0.0031627961434423923, + 0.0033910388592630625, + 0.002370980801060796, + 0.003441923763602972, + 0.0008301583002321422, + 0.0010405861539766192, + 0.0013205730356276035, + 0.001436457154341042, + 0.0011330590350553393, + 0.0008566361502744257, + 0.09830909222364426, + 0.1337241530418396, + 0.012014766223728657, + 0.014243087731301785, + 0.02536582760512829, + 0.013951268047094345, + 0.010491552762687206, + 0.010868743993341923, + 0.00519980164244771, + 0.005409785080701113, + 0.004534216597676277, + 0.005359669681638479, + 0.004968305584043264, + 0.006542486604303122, + 0.004781592171639204, + 0.004929151386022568, + 0.002654715906828642, + 0.003146800212562084, + 0.002048807917162776, + 0.002089254790917039, + 0.0025984817184507847, + 0.0019793021492660046, + 0.003144135931506753, + 0.0019707069732248783, + 0.002464247401803732, + 0.0026010845322161913, + 0.0019031917909160256, + 0.002622657921165228, + 0.003350120037794113, + 0.002940343925729394, + 0.0030570998787879944, + 0.0023753067944198847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.005696847103536129 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015160269103944302 + ], + "opposite_logprob_tokens": [ + -6.501515865325928 + ], + "logprob_reward": -0.0015160269103944302, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015160269103944302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26474982500076294, + 0.098281130194664, + 0.056848153471946716, + 0.01680269092321396, + 0.07132985442876816, + 0.025001708418130875, + 0.29378607869148254, + 0.07569074630737305, + 0.045422039926052094, + 0.052087683230638504 + ], + "credit_map": [ + 0.062468189746141434, + 0.06406889855861664, + 0.018693935126066208, + 0.02411065064370632, + 0.03207557648420334, + 0.018099060282111168, + 0.029115749523043633, + 0.016117779538035393, + 0.01778012327849865, + 0.017241720110177994, + 0.012000110000371933, + 0.010188370011746883, + 0.010321914218366146, + 0.009517809376120567, + 0.012288332916796207, + 0.008942759595811367, + 0.006503928452730179, + 0.007802571635693312, + 0.007314142771065235, + 0.006202848628163338, + 0.006964491680264473, + 0.007000449579209089, + 0.009888636879622936, + 0.005171085242182016, + 0.0019372146343812346, + 0.0028237677179276943, + 0.002086547901853919, + 0.0017396643524989486, + 0.0024834834039211273, + 0.0015848904149606824, + 0.002222517505288124, + 0.001924604526720941, + 0.00801848154515028, + 0.009533586911857128, + 0.010442412458360195, + 0.006102814804762602, + 0.00826528575271368, + 0.011732329614460468, + 0.007422447670251131, + 0.009812501259148121, + 0.003348167985677719, + 0.0052206129767000675, + 0.002607380272820592, + 0.00252448208630085, + 0.0032837558537721634, + 0.002625208580866456, + 0.003059853333979845, + 0.0023322459310293198, + 0.09080024808645248, + 0.12289385497570038, + 0.010616609826683998, + 0.013040971010923386, + 0.023098837584257126, + 0.013252819888293743, + 0.00969632063060999, + 0.010386441834270954, + 0.008708569221198559, + 0.00953126884996891, + 0.006858058739453554, + 0.007837614975869656, + 0.011065489612519741, + 0.011418568901717663, + 0.011863077990710735, + 0.008408098481595516, + 0.004671940580010414, + 0.006717613432556391, + 0.0050240266136825085, + 0.006555517204105854, + 0.0061269099824130535, + 0.005702931899577379, + 0.005771585274487734, + 0.004851517267525196, + 0.005621347110718489, + 0.005718647502362728, + 0.004712425637990236, + 0.006454691756516695, + 0.009458690881729126, + 0.005337911657989025, + 0.009821600280702114, + 0.004962368868291378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0015160269103944302 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613016800954938 + ], + "opposite_logprob_tokens": [ + -5.628612995147705 + ], + "logprob_reward": -0.003613016800954938, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613016800954938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3697572350502014, + 0.08926095813512802, + 0.09255515784025192, + 0.04881568253040314, + 0.06433215737342834, + 0.04107516631484032, + 0.09574317187070847, + 0.08542214334011078, + 0.058862172067165375, + 0.05417614430189133 + ], + "credit_map": [ + 0.03710942342877388, + 0.04965692758560181, + 0.04887566342949867, + 0.04536322131752968, + 0.038674548268318176, + 0.05486297607421875, + 0.04069899395108223, + 0.054515447467565536, + 0.012355644255876541, + 0.013370281085371971, + 0.013114968314766884, + 0.009469682350754738, + 0.009107842110097408, + 0.011143849231302738, + 0.010477197356522083, + 0.010221486911177635, + 0.010928815230727196, + 0.01319960504770279, + 0.01284247636795044, + 0.010398052632808685, + 0.011164451017975807, + 0.010506956838071346, + 0.013078943826258183, + 0.010435851290822029, + 0.009926198050379753, + 0.012052363716065884, + 0.004983917810022831, + 0.004400874488055706, + 0.004654971417039633, + 0.004713152069598436, + 0.004660998936742544, + 0.003423206275328994, + 0.008281609043478966, + 0.009498035535216331, + 0.012312673963606358, + 0.006663117092102766, + 0.006305056624114513, + 0.00600880803540349, + 0.005302988924086094, + 0.0099598728120327, + 0.0051216259598731995, + 0.005496581085026264, + 0.0053506651893258095, + 0.005783227272331715, + 0.004792133811861277, + 0.006158808246254921, + 0.004373818635940552, + 0.003998304717242718, + 0.021081188693642616, + 0.027410242706537247, + 0.008356709964573383, + 0.008543269708752632, + 0.0072114123031497, + 0.010817592963576317, + 0.006064844783395529, + 0.006257913541048765, + 0.010628659278154373, + 0.01065637543797493, + 0.011753304861485958, + 0.013315699063241482, + 0.009393180720508099, + 0.007335379254072905, + 0.010127979330718517, + 0.012211569584906101, + 0.007160854525864124, + 0.008358637802302837, + 0.005506969057023525, + 0.00700251804664731, + 0.007146812509745359, + 0.005584617145359516, + 0.009502786211669445, + 0.008598973974585533, + 0.005709089804440737, + 0.006909199059009552, + 0.005192016717046499, + 0.00667575653642416, + 0.009554636664688587, + 0.006322307046502829, + 0.00901844073086977, + 0.004794699139893055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.003613016800954938 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817513071000576 + ], + "opposite_logprob_tokens": [ + -4.01817512512207 + ], + "logprob_reward": -0.01817513071000576, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817513071000576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15953387320041656, + 0.05282789468765259, + 0.1438439041376114, + 0.04241311550140381, + 0.08408581465482712, + 0.04009420424699783, + 0.053812362253665924, + 0.20190726220607758, + 0.15937592089176178, + 0.062105610966682434 + ], + "credit_map": [ + 0.0161605104804039, + 0.023729436099529266, + 0.01549691054970026, + 0.017904549837112427, + 0.021703269332647324, + 0.01997290551662445, + 0.023531554266810417, + 0.021034738048911095, + 0.009538747370243073, + 0.00834580883383751, + 0.004189661704003811, + 0.006511131301522255, + 0.006365150213241577, + 0.003193625481799245, + 0.010415179654955864, + 0.0042685880325734615, + 0.028687093406915665, + 0.033597432076931, + 0.01282044779509306, + 0.014036678709089756, + 0.012323865666985512, + 0.014724713750183582, + 0.0184539332985878, + 0.009199748747050762, + 0.0061295111663639545, + 0.0070383865386247635, + 0.003880762727931142, + 0.007375060115009546, + 0.0050493646413087845, + 0.004649573937058449, + 0.0043376958929002285, + 0.00395275978371501, + 0.013486795127391815, + 0.01440537441521883, + 0.010574529878795147, + 0.008089498616755009, + 0.011902006343007088, + 0.009550399146974087, + 0.009166066534817219, + 0.006911144591867924, + 0.00566104706376791, + 0.005195298232138157, + 0.004875130485743284, + 0.005031673237681389, + 0.0048360261134803295, + 0.005670820362865925, + 0.00467444583773613, + 0.0041497633792459965, + 0.009440145455300808, + 0.012653260491788387, + 0.004933116491883993, + 0.006712725386023521, + 0.005235715769231319, + 0.005110155325382948, + 0.005722876638174057, + 0.004004367161542177, + 0.02848191186785698, + 0.02494279108941555, + 0.016155771911144257, + 0.022794727236032486, + 0.028677554801106453, + 0.02196202054619789, + 0.03886144235730171, + 0.020031047984957695, + 0.01759759709239006, + 0.022097894921898842, + 0.018607191741466522, + 0.019935015588998795, + 0.021472714841365814, + 0.015159664675593376, + 0.023743972182273865, + 0.020761866122484207, + 0.007606209721416235, + 0.0074142455123364925, + 0.006850702688097954, + 0.005974099040031433, + 0.008880776353180408, + 0.007641009986400604, + 0.008541418239474297, + 0.009197152219712734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.01817513071000576 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump, cartoonish blue bird rendered in a classic 2D animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the bird have bright sky-blue feathers, a soft creamy-white belly, large round black eyes, and a small triangular yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bird standing in the center of the frame amidst a lush green meadow with simple white and yellow flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird enthusiastically puff out its chest and spread its wings to their full extent as a gesture of welcome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a perfectly clear blue sky with stylized fluffy white clouds under bright, high-key sunlight with minimal soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031905481591820717 + ], + "opposite_logprob_tokens": [ + -5.753190517425537 + ], + "logprob_reward": -0.0031905481591820717, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031905481591820717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39703381061553955, + 0.11285121738910675, + 0.02617701329290867, + 0.023059187456965446, + 0.036149200052022934, + 0.02832569181919098, + 0.2510438561439514, + 0.05237157270312309, + 0.04212203249335289, + 0.030866408720612526 + ], + "credit_map": [ + 0.11939448863267899, + 0.11691755056381226, + 0.029989194124937057, + 0.027842817828059196, + 0.028996169567108154, + 0.028247613459825516, + 0.025099853053689003, + 0.02054613269865513, + 0.015824444591999054, + 0.017550889402627945, + 0.01874135434627533, + 0.011969737708568573, + 0.01143629290163517, + 0.011811157688498497, + 0.009816008619964123, + 0.015701329335570335, + 0.004664028994739056, + 0.005604931153357029, + 0.0019311553332954645, + 0.0031825269106775522, + 0.0036929305642843246, + 0.0017036994686350226, + 0.0034070617984980345, + 0.0019906789530068636, + 0.002339432016015053, + 0.003387492848560214, + 0.0040743667632341385, + 0.0026259650476276875, + 0.002416824223473668, + 0.002613065531477332, + 0.003222941653802991, + 0.0023790986742824316, + 0.004710377659648657, + 0.00571491289883852, + 0.005071929190307856, + 0.0033562101889401674, + 0.003926475998014212, + 0.005460888613015413, + 0.0032660437282174826, + 0.004642359912395477, + 0.004459583666175604, + 0.005977539345622063, + 0.002743324963375926, + 0.0026896323543041945, + 0.002775941276922822, + 0.004167940933257341, + 0.0024343901313841343, + 0.003077340079471469, + 0.07200594246387482, + 0.0978795662522316, + 0.014067178592085838, + 0.012316212058067322, + 0.01899433135986328, + 0.016174055635929108, + 0.009099229238927364, + 0.010507353581488132, + 0.00545157864689827, + 0.0061721522361040115, + 0.00631349254399538, + 0.007717748638242483, + 0.0046982732601463795, + 0.005512265022844076, + 0.006922205910086632, + 0.009583857841789722, + 0.004941666033118963, + 0.005469787400215864, + 0.005941083654761314, + 0.006557329557836056, + 0.0038721710443496704, + 0.004558023530989885, + 0.004482560325413942, + 0.006299411412328482, + 0.004249574150890112, + 0.0036917864345014095, + 0.0033970396034419537, + 0.0033755633048713207, + 0.003916499204933643, + 0.003994069527834654, + 0.003613203763961792, + 0.0046286736615002155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0031905481591820717 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_11/index_tensor([8691, 8691])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_11/index_tensor([8691, 8691])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1cb6adfe287c87944363dba827f0d32aa0d99542 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_11/index_tensor([8691, 8691])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, close-up tracking shot follows a shaggy white puppy, its pristine white fur speckled with a few fallen leaves. Its obsidian-bright eyes gleam with concentration, its small, wet black nose pressed to the ground, sniffing excitedly through the thick layer of fallen leaves. Soft afternoon sunlight filters through the treetops, casting dappled shadows on the golden, crimson, and brown carpet of fallen leaves, illuminating the fine dust particles hanging in the air. Suddenly, the puppy seems to lock onto its target. It pauses its sniffing and begins digging with its front paws, alternating between rapid and vigorous movements. The soft, dark earth and fallen leaves are flung behind it, revealing its pink paw pads. Its entire body trembles with excitement as it searches intently for the secret buried beneath. The warm tones of the entire scene imbue the scene with a vibrant sense of movement and pure, curious joy.", + "seed": 16688, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_11/index_8691_seed_16688_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08800375149585307, + "vlm_logprob": -0.08800375149585307, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.08800375149585307, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04571957513689995 + ], + "opposite_logprob_tokens": [ + -3.295719623565674 + ], + "logprob_reward": -0.04571957513689995, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04571957513689995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07804573327302933, + 0.04276585206389427, + 0.2240930050611496, + 0.14848048985004425, + 0.1211649551987648, + 0.036882102489471436, + 0.042005643248558044, + 0.18651440739631653, + 0.08422043919563293, + 0.03582741692662239 + ], + "credit_map": [ + 0.004251414444297552, + 0.008169950917363167, + 0.006857797969132662, + 0.015646161511540413, + 0.021027076989412308, + 0.009488875046372414, + 0.006844592746347189, + 0.00575985899195075, + 0.002191834384575486, + 0.004773708060383797, + 0.0034767710603773594, + 0.005225854460150003, + 0.011622881516814232, + 0.004537579137831926, + 0.0061873081140220165, + 0.004749907646328211, + 0.008914368227124214, + 0.06889025121927261, + 0.010672081261873245, + 0.011618809774518013, + 0.018948432058095932, + 0.012742986902594566, + 0.07453793287277222, + 0.017768124118447304, + 0.013165957294404507, + 0.024568427354097366, + 0.01567188836634159, + 0.029598060995340347, + 0.026745758950710297, + 0.011095188558101654, + 0.016294755041599274, + 0.011340446770191193, + 0.011565027758479118, + 0.01915919966995716, + 0.011903834529221058, + 0.01856563240289688, + 0.020966973155736923, + 0.015658998861908913, + 0.014594520442187786, + 0.00875075627118349, + 0.003857225412502885, + 0.005187699105590582, + 0.004151166882365942, + 0.0061442553997039795, + 0.004118690267205238, + 0.005038481671363115, + 0.003802276449277997, + 0.004582303110510111, + 0.005740438587963581, + 0.006412936374545097, + 0.004217049106955528, + 0.004144398495554924, + 0.004300371743738651, + 0.007294448092579842, + 0.005248477216809988, + 0.004647520836442709, + 0.021275579929351807, + 0.022182414308190346, + 0.019711244851350784, + 0.020083313807845116, + 0.023506054654717445, + 0.030992159619927406, + 0.023991728201508522, + 0.024771904572844505, + 0.007623855024576187, + 0.010092051699757576, + 0.005877417977899313, + 0.010939471423625946, + 0.015514449216425419, + 0.008593014441430569, + 0.015457785688340664, + 0.010122392326593399, + 0.004769978113472462, + 0.007079672999680042, + 0.003868646454066038, + 0.007010098081082106, + 0.005478864070028067, + 0.002889618743211031, + 0.0026708871591836214, + 0.0020596468821167946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.04571957513689995 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207282096147537 + ], + "opposite_logprob_tokens": [ + -2.81207275390625 + ], + "logprob_reward": -0.06207282096147537, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207282096147537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15163041651248932, + 0.05458284541964531, + 0.3226218819618225, + 0.12368088960647583, + 0.07345915585756302, + 0.04754127934575081, + 0.05567751079797745, + 0.11173083633184433, + 0.03343396633863449, + 0.025641214102506638 + ], + "credit_map": [ + 0.014782875776290894, + 0.013858001679182053, + 0.02234542742371559, + 0.028612449765205383, + 0.022117622196674347, + 0.018435057252645493, + 0.018208863213658333, + 0.013270122930407524, + 0.002833368256688118, + 0.005817739758640528, + 0.0038758234586566687, + 0.006201770156621933, + 0.014781888574361801, + 0.005558623466640711, + 0.008139495737850666, + 0.007374136708676815, + 0.012986090034246445, + 0.09627816826105118, + 0.013439898379147053, + 0.015802521258592606, + 0.027365075424313545, + 0.019951459020376205, + 0.10850175470113754, + 0.028296897187829018, + 0.008238949812948704, + 0.024937154725193977, + 0.012272077612578869, + 0.02664669044315815, + 0.02144099585711956, + 0.009038854390382767, + 0.011975560337305069, + 0.009130616672337055, + 0.006556231062859297, + 0.01159583032131195, + 0.006948570255190134, + 0.011956917122006416, + 0.010390420444309711, + 0.012082795612514019, + 0.008589624427258968, + 0.005338764749467373, + 0.004624356981366873, + 0.004767101723700762, + 0.005538557656109333, + 0.007508920039981604, + 0.004018570762127638, + 0.007900534197688103, + 0.005864152684807777, + 0.007319084368646145, + 0.004916086792945862, + 0.008497310802340508, + 0.008307534269988537, + 0.00670253811404109, + 0.004988003522157669, + 0.005324199795722961, + 0.009408921003341675, + 0.007532916497439146, + 0.013122932985424995, + 0.012612123973667622, + 0.018267396837472916, + 0.011832475662231445, + 0.009059170261025429, + 0.018905112519860268, + 0.011421285569667816, + 0.016510339453816414, + 0.0032261060550808907, + 0.003269110806286335, + 0.002844291040673852, + 0.00349913677200675, + 0.005740868393331766, + 0.0031520104967057705, + 0.007035552524030209, + 0.0046668886207044125, + 0.005046519450843334, + 0.0045836493372917175, + 0.002428331645205617, + 0.003954136278480291, + 0.004811759572476149, + 0.0015553837874904275, + 0.0016825199127197266, + 0.0015789130702614784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.06207282096147537 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866817593574524 + ], + "opposite_logprob_tokens": [ + -3.048668146133423 + ], + "logprob_reward": -0.04866817593574524, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866817593574524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046247903257608414, + 0.04295758157968521, + 0.23519118130207062, + 0.25042134523391724, + 0.032965246587991714, + 0.1057417094707489, + 0.10197443515062332, + 0.06263778358697891, + 0.09879250824451447, + 0.023070242255926132 + ], + "credit_map": [ + 0.0031923220958560705, + 0.005002036690711975, + 0.0042494856752455235, + 0.008810678496956825, + 0.010861972346901894, + 0.005422359798103571, + 0.004213533829897642, + 0.004495515022426844, + 0.00197577360086143, + 0.005300442688167095, + 0.0024130945093929768, + 0.005734989885240793, + 0.014504383318126202, + 0.004988303408026695, + 0.004250776953995228, + 0.003789816750213504, + 0.008676835335791111, + 0.06994932144880295, + 0.011017652228474617, + 0.010948591865599155, + 0.022277789190411568, + 0.014411955140531063, + 0.07809659093618393, + 0.0198124460875988, + 0.01654590480029583, + 0.052539411932229996, + 0.022598056122660637, + 0.049668338149785995, + 0.05689677223563194, + 0.016551386564970016, + 0.021863695234060287, + 0.01375778391957283, + 0.003726003924384713, + 0.0034294496290385723, + 0.0031452178955078125, + 0.0034435687121003866, + 0.003918818198144436, + 0.008548120968043804, + 0.003802589373663068, + 0.0029514795169234276, + 0.011246428824961185, + 0.013081696815788746, + 0.011429508216679096, + 0.017914438620209694, + 0.009265259839594364, + 0.015584798529744148, + 0.012601494789123535, + 0.014618081972002983, + 0.007183159235864878, + 0.02119106985628605, + 0.014040684327483177, + 0.010329809971153736, + 0.00887064728885889, + 0.009182799607515335, + 0.016331898048520088, + 0.014844377525150776, + 0.007541515864431858, + 0.00659762741997838, + 0.010491381399333477, + 0.00785911176353693, + 0.005301819648593664, + 0.00806458480656147, + 0.006667479407042265, + 0.01011426467448473, + 0.010141951963305473, + 0.01354753877967596, + 0.009211516007781029, + 0.012286669574677944, + 0.014762621372938156, + 0.009457404725253582, + 0.015430849976837635, + 0.013953955844044685, + 0.0032288723159581423, + 0.0034678014926612377, + 0.002627615351229906, + 0.002550124190747738, + 0.003471881616860628, + 0.002879018895328045, + 0.0027872868813574314, + 0.0020576424431055784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.04866817593574524 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021984234917908907 + ], + "opposite_logprob_tokens": [ + -6.127198219299316 + ], + "logprob_reward": -0.0021984234917908907, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021984234917908907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21631336212158203, + 0.083639957010746, + 0.27187076210975647, + 0.062392812222242355, + 0.0614476315677166, + 0.03612030670046806, + 0.11727280169725418, + 0.06600826978683472, + 0.048888783901929855, + 0.03604525327682495 + ], + "credit_map": [ + 0.017338937148451805, + 0.01931655779480934, + 0.02829851023852825, + 0.055164165794849396, + 0.027248341590166092, + 0.02215835079550743, + 0.026345057412981987, + 0.02044345997273922, + 0.0037456650752574205, + 0.010164003819227219, + 0.005202730186283588, + 0.010221822187304497, + 0.026633309200406075, + 0.010123992338776588, + 0.010900348424911499, + 0.006648092530667782, + 0.011493691243231297, + 0.07679589837789536, + 0.01413772627711296, + 0.01540323719382286, + 0.023859402164816856, + 0.01598973199725151, + 0.09009317308664322, + 0.024097934365272522, + 0.004230455029755831, + 0.010650643147528172, + 0.006616092287003994, + 0.006978343240916729, + 0.013290437869727612, + 0.008889418095350266, + 0.006037889514118433, + 0.00569953816011548, + 0.00691103283315897, + 0.008205581456422806, + 0.006178826093673706, + 0.007356614340096712, + 0.008997304365038872, + 0.01165287010371685, + 0.006786690559238195, + 0.005358720198273659, + 0.002345119137316942, + 0.005167212802916765, + 0.005708739627152681, + 0.0045752073638141155, + 0.003875348949804902, + 0.004027318675071001, + 0.005385214928537607, + 0.005036144517362118, + 0.010419906117022038, + 0.02001163922250271, + 0.02004186436533928, + 0.013905921950936317, + 0.01096278801560402, + 0.010050250217318535, + 0.014614037238061428, + 0.017266398295760155, + 0.0064051346853375435, + 0.004990353714674711, + 0.012285571545362473, + 0.008739953860640526, + 0.004978799261152744, + 0.011952881701290607, + 0.0053854151628911495, + 0.011270161718130112, + 0.004640418104827404, + 0.005702836439013481, + 0.00498560257256031, + 0.00496481591835618, + 0.0075021469965577126, + 0.004244921263307333, + 0.010071290656924248, + 0.006776754278689623, + 0.00539802573621273, + 0.005614025983959436, + 0.005312575958669186, + 0.005688793025910854, + 0.0036249065306037664, + 0.004518806003034115, + 0.0027192130219191313, + 0.0031689086463302374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.0021984234917908907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813597619533539 + ], + "opposite_logprob_tokens": [ + -1.4063597917556763 + ], + "logprob_reward": -0.2813597619533539, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813597619533539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05180782079696655, + 0.020044051110744476, + 0.5048661828041077, + 0.11218534409999847, + 0.03009316511452198, + 0.08597142994403839, + 0.046177007257938385, + 0.06595011055469513, + 0.046562984585762024, + 0.03634189814329147 + ], + "credit_map": [ + 0.0047927373088896275, + 0.00483323959633708, + 0.006509691011160612, + 0.006329643540084362, + 0.010599485598504543, + 0.008460105396807194, + 0.005679052788764238, + 0.004603864625096321, + 0.0013919505290687084, + 0.0020328282844275236, + 0.0013905977830290794, + 0.0024454151280224323, + 0.004896173253655434, + 0.0025675168726593256, + 0.002847375115379691, + 0.0024721950758248568, + 0.019698144868016243, + 0.15167631208896637, + 0.020692789927124977, + 0.025067918002605438, + 0.041193895041942596, + 0.03142343461513519, + 0.1720946580171585, + 0.04301906377077103, + 0.008013366721570492, + 0.02290017530322075, + 0.009065673686563969, + 0.018101783469319344, + 0.02911769226193428, + 0.005979040637612343, + 0.012484981678426266, + 0.006522631272673607, + 0.002737873001024127, + 0.004361686762422323, + 0.0033364975824952126, + 0.00462683429941535, + 0.004350753966718912, + 0.0040944539941847324, + 0.003768273163586855, + 0.0028167925775051117, + 0.009876135736703873, + 0.011872118338942528, + 0.009002180770039558, + 0.014965113252401352, + 0.008262679912149906, + 0.011937957257032394, + 0.010624809190630913, + 0.009430432692170143, + 0.004164228215813637, + 0.009284492582082748, + 0.006354709155857563, + 0.0053138816729187965, + 0.003966263961046934, + 0.00462590204551816, + 0.006545240990817547, + 0.005922288633882999, + 0.005932704079896212, + 0.006940785795450211, + 0.009655795991420746, + 0.007026748266071081, + 0.005918401293456554, + 0.013812809251248837, + 0.006394648924469948, + 0.010268214158713818, + 0.004163733217865229, + 0.007159830536693335, + 0.003928929101675749, + 0.006155596114695072, + 0.008653494529426098, + 0.004654603078961372, + 0.005659785121679306, + 0.006187013816088438, + 0.007687568664550781, + 0.00730662839487195, + 0.0030784260015934706, + 0.004779637325555086, + 0.007067497819662094, + 0.002466181293129921, + 0.0021307552233338356, + 0.0018252049339935184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.2813597619533539 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, close-up tracking shot follows a shaggy white puppy, its pristine white fur speckled with a few fallen leaves. Its obsidian-bright eyes gleam with concentration, its small, wet black nose pressed to the ground, sniffing excitedly through the thick layer of fallen leaves. Soft afternoon sunlight filters through the treetops, casting dappled shadows on the golden, crimson, and brown carpet of fallen leaves, illuminating the fine dust particles hanging in the air. Suddenly, the puppy seems to lock onto its target. It pauses its sniffing and begins digging with its front paws, alternating between rapid and vigorous movements. The soft, dark earth and fallen leaves are flung behind it, revealing its pink paw pads. Its entire body trembles with excitement as it searches intently for the secret buried beneath. The warm tones of the entire scene imbue the scene with a vibrant sense of movement and pure, curious joy.", + "seed": 16689, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_11/index_8691_seed_16689_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05742219684179872, + "vlm_logprob": -0.05742219684179872, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.05742219684179872, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03304805979132652 + ], + "opposite_logprob_tokens": [ + -3.783048152923584 + ], + "logprob_reward": -0.03304805979132652, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03304805979132652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11879535764455795, + 0.11168975383043289, + 0.05863652005791664, + 0.33095455169677734, + 0.06267006695270538, + 0.0772794634103775, + 0.037759099155664444, + 0.039370398968458176, + 0.13471272587776184, + 0.02813207544386387 + ], + "credit_map": [ + 0.007894188165664673, + 0.019467443227767944, + 0.008475832641124725, + 0.013288772664964199, + 0.037165116518735886, + 0.01096172071993351, + 0.010202131234109402, + 0.011340147815644741, + 0.008926310576498508, + 0.008563127368688583, + 0.013544687069952488, + 0.00858231820166111, + 0.019073031842708588, + 0.01859177090227604, + 0.017847582697868347, + 0.016560928896069527, + 0.006528465077280998, + 0.00867228489369154, + 0.006203757598996162, + 0.006402082275599241, + 0.007807659916579723, + 0.009949360974133015, + 0.005822212900966406, + 0.007250695489346981, + 0.029886215925216675, + 0.03928016871213913, + 0.024236511439085007, + 0.02415437251329422, + 0.07414375245571136, + 0.058956682682037354, + 0.053523723036050797, + 0.0267731212079525, + 0.00475311791524291, + 0.006430215667933226, + 0.005292760208249092, + 0.004728973377496004, + 0.012835459783673286, + 0.009460825473070145, + 0.00921389739960432, + 0.009954816661775112, + 0.005025175865739584, + 0.008630544878542423, + 0.006064177490770817, + 0.006841043010354042, + 0.008709140121936798, + 0.013301979750394821, + 0.010138384997844696, + 0.018569020554423332, + 0.004029756411910057, + 0.003959826659411192, + 0.00415593758225441, + 0.006682074163109064, + 0.00465638330206275, + 0.004115693271160126, + 0.0060945493169128895, + 0.004064876586198807, + 0.0026220267172902822, + 0.0021841092966496944, + 0.006076275371015072, + 0.011505919508635998, + 0.0023218130227178335, + 0.004041858483105898, + 0.004968681372702122, + 0.005649712402373552, + 0.015016764402389526, + 0.013905253261327744, + 0.01432189904153347, + 0.016227200627326965, + 0.020908648148179054, + 0.022167151793837547, + 0.01668129861354828, + 0.015484514646232128, + 0.0023067023139446974, + 0.0042074560187757015, + 0.0021392875351011753, + 0.005557368043810129, + 0.0033123265020549297, + 0.0037090193945914507, + 0.0038698497228324413, + 0.003030067076906562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.03304805979132652 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308406636118889 + ], + "opposite_logprob_tokens": [ + -3.168084144592285 + ], + "logprob_reward": -0.04308406636118889, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308406636118889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1768149882555008, + 0.1101636067032814, + 0.07131689786911011, + 0.25389474630355835, + 0.0850629061460495, + 0.035591911524534225, + 0.0382394902408123, + 0.039357706904411316, + 0.03410998359322548, + 0.15544778108596802 + ], + "credit_map": [ + 0.014833640307188034, + 0.032493457198143005, + 0.01900637336075306, + 0.010875626467168331, + 0.05394336208701134, + 0.02080414816737175, + 0.00912822037935257, + 0.015730157494544983, + 0.007816797122359276, + 0.010439500212669373, + 0.010943775065243244, + 0.0096741309389472, + 0.022872937843203545, + 0.016982020810246468, + 0.02080596424639225, + 0.010628480464220047, + 0.006203319877386093, + 0.008703842759132385, + 0.008464034646749496, + 0.013217826373875141, + 0.006292859558016062, + 0.014169339090585709, + 0.005207885522395372, + 0.009057790040969849, + 0.022384023293852806, + 0.027926642447710037, + 0.01907511241734028, + 0.018774541094899178, + 0.05724024027585983, + 0.04381899535655975, + 0.044219065457582474, + 0.02045612409710884, + 0.006810882594436407, + 0.011888825334608555, + 0.00906766764819622, + 0.009156639687716961, + 0.017656145617365837, + 0.010031231679022312, + 0.010994341224431992, + 0.009457173757255077, + 0.0028567470144480467, + 0.005772140342742205, + 0.004350731614977121, + 0.003469771007075906, + 0.003502280917018652, + 0.005721392575651407, + 0.0034551932476460934, + 0.006463654339313507, + 0.0038739473093301058, + 0.0041661872528493404, + 0.0038622268475592136, + 0.004623003304004669, + 0.005706201307475567, + 0.003799833357334137, + 0.007724094670265913, + 0.004483997356146574, + 0.0035412232391536236, + 0.00330529036000371, + 0.004829570651054382, + 0.005049640312790871, + 0.0046074772253632545, + 0.005982737522572279, + 0.006007603369653225, + 0.006034160032868385, + 0.003124834271147847, + 0.006421198602765799, + 0.0019654452335089445, + 0.008677154779434204, + 0.0036764293909072876, + 0.005194841884076595, + 0.0023955265060067177, + 0.0026545492000877857, + 0.009702715091407299, + 0.027010230347514153, + 0.009785705246031284, + 0.03420327231287956, + 0.019596170634031296, + 0.017156308516860008, + 0.02129032462835312, + 0.016703058034181595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.04308406636118889 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866284132003784 + ], + "opposite_logprob_tokens": [ + -3.0486629009246826 + ], + "logprob_reward": -0.04866284132003784, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866284132003784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28260722756385803, + 0.11896267533302307, + 0.041981857270002365, + 0.1926460713148117, + 0.05383434519171715, + 0.06026819348335266, + 0.09863686561584473, + 0.04567275568842888, + 0.058588139712810516, + 0.0468018613755703 + ], + "credit_map": [ + 0.02139621041715145, + 0.0474960096180439, + 0.024029625579714775, + 0.020803220570087433, + 0.08849088102579117, + 0.02876628376543522, + 0.01940433867275715, + 0.03222067281603813, + 0.008886107243597507, + 0.010162051767110825, + 0.011769051663577557, + 0.00956167746335268, + 0.024876760318875313, + 0.01640574261546135, + 0.024256886914372444, + 0.01304439827799797, + 0.004421197343617678, + 0.006112800911068916, + 0.004143136087805033, + 0.0035739631857722998, + 0.005803142208606005, + 0.007884472608566284, + 0.004160228651016951, + 0.005882915575057268, + 0.015128077007830143, + 0.02222701534628868, + 0.012017210945487022, + 0.014101375825703144, + 0.04082580283284187, + 0.03802572935819626, + 0.035447217524051666, + 0.01487363688647747, + 0.004906921647489071, + 0.005444591864943504, + 0.007145275827497244, + 0.004345097579061985, + 0.009196871891617775, + 0.008671487681567669, + 0.007473476696759462, + 0.006650618743151426, + 0.004743096884340048, + 0.0075432658195495605, + 0.005226160399615765, + 0.005340422969311476, + 0.005996679421514273, + 0.014055401086807251, + 0.006528748665004969, + 0.010834415443241596, + 0.008264610543847084, + 0.00913096871227026, + 0.010294576175510883, + 0.0141608202829957, + 0.010457491502165794, + 0.009675133042037487, + 0.023706892505288124, + 0.01294636819511652, + 0.00422892440110445, + 0.004026360809803009, + 0.0065274545922875404, + 0.00747319869697094, + 0.004760188516229391, + 0.0051386915147304535, + 0.007802347186952829, + 0.0057155899703502655, + 0.005450312979519367, + 0.008938483893871307, + 0.005070769228041172, + 0.008161692880094051, + 0.007889365777373314, + 0.010856925509870052, + 0.006079196929931641, + 0.006141393445432186, + 0.003787981579080224, + 0.007888348773121834, + 0.0031751322094351053, + 0.009488549083471298, + 0.005909768398851156, + 0.003809016663581133, + 0.007147707510739565, + 0.0055953580886125565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.04866284132003784 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00194110709708184 + ], + "opposite_logprob_tokens": [ + -6.251941204071045 + ], + "logprob_reward": -0.00194110709708184, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00194110709708184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13368311524391174, + 0.3348275125026703, + 0.02816702052950859, + 0.04090464860200882, + 0.04740682244300842, + 0.12514536082744598, + 0.03144083917140961, + 0.10275033116340637, + 0.04492153599858284, + 0.11075281351804733 + ], + "credit_map": [ + 0.009666726924479008, + 0.022918710485100746, + 0.009058346040546894, + 0.012934370897710323, + 0.04483989253640175, + 0.013976636342704296, + 0.009327598847448826, + 0.010960816405713558, + 0.030450880527496338, + 0.027070997282862663, + 0.054011791944503784, + 0.03325774148106575, + 0.03576326742768288, + 0.055188458412885666, + 0.034504640847444534, + 0.06457973271608353, + 0.0032449711579829454, + 0.0032844252418726683, + 0.0030349688604474068, + 0.004426402971148491, + 0.0027691018767654896, + 0.0037141311913728714, + 0.003835883690044284, + 0.003857133211567998, + 0.003248092019930482, + 0.0029981411062180996, + 0.003509089583531022, + 0.003388080047443509, + 0.006071990821510553, + 0.01192785706371069, + 0.005501833278685808, + 0.004259560257196426, + 0.0046221669763326645, + 0.005463994108140469, + 0.006222127005457878, + 0.004738560412079096, + 0.00959829706698656, + 0.006443359889090061, + 0.0052751884795725346, + 0.00504312664270401, + 0.01012424472719431, + 0.027149543166160583, + 0.01503042597323656, + 0.013800648041069508, + 0.010655914433300495, + 0.023317044600844383, + 0.011066646315157413, + 0.014000890776515007, + 0.002392400288954377, + 0.0025875966530293226, + 0.004316793289035559, + 0.005505889654159546, + 0.003188349073752761, + 0.0038803848437964916, + 0.005123501643538475, + 0.004445922095328569, + 0.005841610487550497, + 0.004529057536274195, + 0.01550503820180893, + 0.03026207536458969, + 0.006137605756521225, + 0.009769832715392113, + 0.01253065187484026, + 0.01817444898188114, + 0.005390384700149298, + 0.003801995422691107, + 0.005353673826903105, + 0.006368497386574745, + 0.006110871210694313, + 0.007042738609015942, + 0.005507252179086208, + 0.005346120800822973, + 0.006764581426978111, + 0.01821974664926529, + 0.006843128707259893, + 0.024572599679231644, + 0.013238413259387016, + 0.014675457961857319, + 0.0157198254019022, + 0.010719054378569126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.00194110709708184 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037490963935852 + ], + "opposite_logprob_tokens": [ + -1.9103748798370361 + ], + "logprob_reward": -0.16037490963935852, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037490963935852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09827303886413574, + 0.06019505485892296, + 0.05895543470978737, + 0.11448805034160614, + 0.08000896126031876, + 0.3012729585170746, + 0.033650241792201996, + 0.10456875711679459, + 0.06603068113327026, + 0.08255679905414581 + ], + "credit_map": [ + 0.011226657778024673, + 0.01047400664538145, + 0.013814859092235565, + 0.007990951649844646, + 0.019821682944893837, + 0.015501725487411022, + 0.008188235573470592, + 0.01125493273139, + 0.005989251658320427, + 0.005199410021305084, + 0.009957224130630493, + 0.00843263603746891, + 0.008388319052755833, + 0.007749610114842653, + 0.00641274731606245, + 0.008065861649811268, + 0.006356490310281515, + 0.010384457185864449, + 0.0074619874358177185, + 0.005626847967505455, + 0.009550021030008793, + 0.008371056988835335, + 0.004624098539352417, + 0.006580480840057135, + 0.010537339374423027, + 0.0098387086763978, + 0.011344285681843758, + 0.00915120355784893, + 0.023757440969347954, + 0.016943350434303284, + 0.02250616066157818, + 0.010409567505121231, + 0.005431093275547028, + 0.009586600586771965, + 0.009469357319176197, + 0.006292339880019426, + 0.014016468077898026, + 0.010257630608975887, + 0.013957290910184383, + 0.010998189449310303, + 0.022457541897892952, + 0.06936665624380112, + 0.028634754940867424, + 0.025196095928549767, + 0.026592455804347992, + 0.05938146635890007, + 0.033843424171209335, + 0.03580058366060257, + 0.003082173876464367, + 0.005355191417038441, + 0.003501419210806489, + 0.003975607920438051, + 0.004016866907477379, + 0.003870655782520771, + 0.006531007122248411, + 0.0033173197880387306, + 0.007859798148274422, + 0.0054635293781757355, + 0.015866778790950775, + 0.022242553532123566, + 0.009101279079914093, + 0.011631629429757595, + 0.01669614017009735, + 0.015707047656178474, + 0.007118404842913151, + 0.007067031692713499, + 0.006802873685956001, + 0.009824811480939388, + 0.008681810460984707, + 0.012424628250300884, + 0.007174673955887556, + 0.006936453282833099, + 0.0049660298973321915, + 0.012946913950145245, + 0.005766202230006456, + 0.01887955702841282, + 0.00906201172620058, + 0.009785140864551067, + 0.012362435460090637, + 0.008788514882326126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 7 + ], + "selected_reward": -0.16037490963935852 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, close-up tracking shot follows a shaggy white puppy, its pristine white fur speckled with a few fallen leaves. Its obsidian-bright eyes gleam with concentration, its small, wet black nose pressed to the ground, sniffing excitedly through the thick layer of fallen leaves. Soft afternoon sunlight filters through the treetops, casting dappled shadows on the golden, crimson, and brown carpet of fallen leaves, illuminating the fine dust particles hanging in the air. Suddenly, the puppy seems to lock onto its target. It pauses its sniffing and begins digging with its front paws, alternating between rapid and vigorous movements. The soft, dark earth and fallen leaves are flung behind it, revealing its pink paw pads. Its entire body trembles with excitement as it searches intently for the secret buried beneath. The warm tones of the entire scene imbue the scene with a vibrant sense of movement and pure, curious joy.", + "seed": 16690, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_11/index_8691_seed_16690_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.059506577299907805, + "vlm_logprob": -0.059506577299907805, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.059506577299907805, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04133337736129761 + ], + "opposite_logprob_tokens": [ + -3.4163334369659424 + ], + "logprob_reward": -0.04133337736129761, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04133337736129761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1223689466714859, + 0.113848477602005, + 0.069202721118927, + 0.05444340035319328, + 0.06757648289203644, + 0.0870458111166954, + 0.1625848263502121, + 0.05578004941344261, + 0.017605673521757126, + 0.24954357743263245 + ], + "credit_map": [ + 0.009662123396992683, + 0.017658600583672523, + 0.012738888151943684, + 0.01156279630959034, + 0.023848168551921844, + 0.010108748450875282, + 0.019902145490050316, + 0.01688748225569725, + 0.011512796394526958, + 0.018385233357548714, + 0.008815046399831772, + 0.00973969791084528, + 0.024762770161032677, + 0.010510222055017948, + 0.013898261822760105, + 0.016224456951022148, + 0.006704166065901518, + 0.015741202980279922, + 0.007502930238842964, + 0.009321731515228748, + 0.00924539566040039, + 0.009898762218654156, + 0.006264549680054188, + 0.004523990675806999, + 0.006085654255002737, + 0.011567353270947933, + 0.006651030387729406, + 0.0058924490585923195, + 0.007495718076825142, + 0.004549314733594656, + 0.006261026952415705, + 0.005940855946391821, + 0.007012983784079552, + 0.015525216236710548, + 0.007307533174753189, + 0.006739902775734663, + 0.008501245640218258, + 0.00768715376034379, + 0.009048675186932087, + 0.005753776058554649, + 0.008337724953889847, + 0.01560315303504467, + 0.006393918301910162, + 0.008911723271012306, + 0.011654356494545937, + 0.014634454622864723, + 0.01107232179492712, + 0.010438170284032822, + 0.017606597393751144, + 0.02709258906543255, + 0.01759563572704792, + 0.020972074940800667, + 0.027884773910045624, + 0.014342674985527992, + 0.01825154945254326, + 0.018838943913578987, + 0.0056516551412642, + 0.00588020496070385, + 0.004014214500784874, + 0.0042532384395599365, + 0.009437968023121357, + 0.013135649263858795, + 0.006553218234330416, + 0.006853904575109482, + 0.0026209265924990177, + 0.002524434356018901, + 0.0017938623204827309, + 0.0008780427160672843, + 0.0022151449229568243, + 0.0029663913883268833, + 0.0017512686317786574, + 0.002855603815987706, + 0.03131541237235069, + 0.021187884733080864, + 0.009836677461862564, + 0.008130103349685669, + 0.0715353712439537, + 0.08839301764965057, + 0.007145291659981012, + 0.011999818496406078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.04133337736129761 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029823314398527145 + ], + "opposite_logprob_tokens": [ + -3.5298233032226562 + ], + "logprob_reward": -0.029823314398527145, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029823314398527145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18380650877952576, + 0.06285230070352554, + 0.05166637524962425, + 0.036110904067754745, + 0.09934744983911514, + 0.05002903938293457, + 0.03034980595111847, + 0.08413396030664444, + 0.016691483557224274, + 0.385012149810791 + ], + "credit_map": [ + 0.016831928864121437, + 0.023059960454702377, + 0.03399095684289932, + 0.021376537159085274, + 0.020617516711354256, + 0.015725599601864815, + 0.024312371388077736, + 0.02789163403213024, + 0.007380862254649401, + 0.006694483570754528, + 0.011303845793008804, + 0.0052855960093438625, + 0.007462226320058107, + 0.007268535438925028, + 0.007521986961364746, + 0.00993476714938879, + 0.004583395551890135, + 0.011825485154986382, + 0.005705184768885374, + 0.005981190130114555, + 0.0060031539760529995, + 0.0059679835103452206, + 0.005482507403939962, + 0.0061174719594419, + 0.003915456123650074, + 0.004176623187959194, + 0.0039046602323651314, + 0.004549834877252579, + 0.00797137152403593, + 0.0028448638040572405, + 0.004357985220849514, + 0.004390109796077013, + 0.008130963891744614, + 0.015747856348752975, + 0.010283919051289558, + 0.012563822790980339, + 0.010654198005795479, + 0.010187293402850628, + 0.01858053170144558, + 0.013198871165513992, + 0.005183021537959576, + 0.006888012867420912, + 0.004489494021981955, + 0.005037193186581135, + 0.006227409932762384, + 0.009397958405315876, + 0.00588664086535573, + 0.006919306702911854, + 0.0025519465561956167, + 0.0060022673569619656, + 0.002541967201977968, + 0.0036964172031730413, + 0.00431834626942873, + 0.003211171831935644, + 0.004441757686436176, + 0.0035859302151948214, + 0.0059878891333937645, + 0.0071133701130747795, + 0.005203778389841318, + 0.005177867133170366, + 0.02735258638858795, + 0.018830344080924988, + 0.007547769229859114, + 0.006920353509485722, + 0.002537699416279793, + 0.0014202217571437359, + 0.002353327814489603, + 0.0016515611205250025, + 0.0014691123506054282, + 0.002001208020374179, + 0.003452796721830964, + 0.0018055563559755683, + 0.06719409674406052, + 0.029247604310512543, + 0.013351758010685444, + 0.010883420705795288, + 0.09158533066511154, + 0.1445944607257843, + 0.009682091884315014, + 0.01847340539097786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.029823314398527145 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329319901764393 + ], + "opposite_logprob_tokens": [ + -3.7732932567596436 + ], + "logprob_reward": -0.02329319901764393, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329319901764393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06836885213851929, + 0.025834612548351288, + 0.04377548024058342, + 0.10537900030612946, + 0.44535601139068604, + 0.060431599617004395, + 0.03982265666127205, + 0.06908661127090454, + 0.016836468130350113, + 0.12510864436626434 + ], + "credit_map": [ + 0.00430678715929389, + 0.011113782413303852, + 0.00774040911346674, + 0.005860249977558851, + 0.012865743599832058, + 0.0060507068410515785, + 0.013374117203056812, + 0.007057058624923229, + 0.0030788355506956577, + 0.004613899625837803, + 0.002870537806302309, + 0.002070228336378932, + 0.0030038682743906975, + 0.003751559415832162, + 0.004146597348153591, + 0.0022990864235907793, + 0.0038815452717244625, + 0.00951392948627472, + 0.004910389892756939, + 0.007242907304316759, + 0.00548730231821537, + 0.005182378925383091, + 0.004876045044511557, + 0.002680980833247304, + 0.011649235151708126, + 0.01307628769427538, + 0.011231943964958191, + 0.014329444617033005, + 0.021165136247873306, + 0.007782230619341135, + 0.013088621199131012, + 0.013056107796728611, + 0.03862130269408226, + 0.06862032413482666, + 0.05078881233930588, + 0.05059681832790375, + 0.051073070615530014, + 0.049661412835121155, + 0.0715406984090805, + 0.06445358693599701, + 0.004143241327255964, + 0.015020507387816906, + 0.004182877019047737, + 0.006001779343932867, + 0.015743060037493706, + 0.007084706798195839, + 0.003707681782543659, + 0.004547744523733854, + 0.0036580825690180063, + 0.0068512726575136185, + 0.0029179104603827, + 0.0034942138008773327, + 0.007140272296965122, + 0.005195756908506155, + 0.005185555201023817, + 0.005379595328122377, + 0.005015267990529537, + 0.012010663747787476, + 0.004810684360563755, + 0.006282035261392593, + 0.017281901091337204, + 0.01314057782292366, + 0.005908088758587837, + 0.004637392237782478, + 0.0017950074980035424, + 0.002246352145448327, + 0.0019624135456979275, + 0.0014348531840369105, + 0.0020596690010279417, + 0.001903730328194797, + 0.0029293496627360582, + 0.0025050928816199303, + 0.011534428223967552, + 0.012021603062748909, + 0.004519531968981028, + 0.004228534176945686, + 0.04259034991264343, + 0.040334198623895645, + 0.004191287327557802, + 0.005688713863492012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.02329319901764393 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00151305110193789 + ], + "opposite_logprob_tokens": [ + -6.5015130043029785 + ], + "logprob_reward": -0.00151305110193789, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.00151305110193789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1742856353521347, + 0.053301852196455, + 0.033647216856479645, + 0.04112847521901131, + 0.11332396417856216, + 0.11778262257575989, + 0.04832587391138077, + 0.1050361767411232, + 0.029287906363606453, + 0.2838802635669708 + ], + "credit_map": [ + 0.01767018996179104, + 0.024147801101207733, + 0.028087083250284195, + 0.021490471437573433, + 0.020967038348317146, + 0.013465005904436111, + 0.025049014016985893, + 0.02340902015566826, + 0.006087245419621468, + 0.008235235698521137, + 0.005721377208828926, + 0.003690935205668211, + 0.0076466756872832775, + 0.005440112669020891, + 0.007795070298016071, + 0.008685200475156307, + 0.003398215863853693, + 0.00575366010889411, + 0.004805992357432842, + 0.004933081567287445, + 0.0043660905212163925, + 0.0032180382404476404, + 0.0038595802616328, + 0.0033125581685453653, + 0.004774746019393206, + 0.005403985269367695, + 0.0037452825345098972, + 0.0076922886073589325, + 0.007656857371330261, + 0.0031848999205976725, + 0.004012265708297491, + 0.004658149089664221, + 0.010156911797821522, + 0.01997668296098709, + 0.014578861184418201, + 0.013572955504059792, + 0.012071840465068817, + 0.012227020226418972, + 0.015122116543352604, + 0.015617575496435165, + 0.008733026683330536, + 0.02413528598845005, + 0.0052130986005067825, + 0.007582557387650013, + 0.034595951437950134, + 0.016712820157408714, + 0.009609215892851353, + 0.011200659908354282, + 0.004506021272391081, + 0.007506490685045719, + 0.004363828804343939, + 0.004786629229784012, + 0.007686626631766558, + 0.007577295880764723, + 0.005373827647417784, + 0.006525150965899229, + 0.007832121104001999, + 0.007876872085034847, + 0.0066194380633533, + 0.007447727490216494, + 0.029577406123280525, + 0.023221418261528015, + 0.009608885273337364, + 0.01285230927169323, + 0.004421139135956764, + 0.003310101106762886, + 0.003975792322307825, + 0.0031113186851143837, + 0.002787807723507285, + 0.003912476357072592, + 0.004384088795632124, + 0.003385182935744524, + 0.027274593710899353, + 0.028712265193462372, + 0.009894742630422115, + 0.009672390297055244, + 0.09825127571821213, + 0.08812231570482254, + 0.009360743686556816, + 0.012591954320669174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.00151305110193789 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156994462013245 + ], + "opposite_logprob_tokens": [ + -1.70156991481781 + ], + "logprob_reward": -0.20156994462013245, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156994462013245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11030130833387375, + 0.04175999388098717, + 0.03555968031287193, + 0.09285428375005722, + 0.34266242384910583, + 0.07082603126764297, + 0.09529118239879608, + 0.061640676110982895, + 0.032922688871622086, + 0.11618178337812424 + ], + "credit_map": [ + 0.009985529817640781, + 0.0150798000395298, + 0.013480366207659245, + 0.012546953745186329, + 0.015602890402078629, + 0.008417543023824692, + 0.02145029418170452, + 0.013737933710217476, + 0.0048517873510718346, + 0.005008648615330458, + 0.006101659499108791, + 0.003325526136904955, + 0.00549926096573472, + 0.005495547316968441, + 0.0043356288224458694, + 0.007141934707760811, + 0.004010610282421112, + 0.006806326564401388, + 0.0042747072875499725, + 0.004196720663458109, + 0.0044968924485147, + 0.004122702404856682, + 0.004275971092283726, + 0.0033757505007088184, + 0.009529225528240204, + 0.014865180477499962, + 0.0071132187731564045, + 0.010453247465193272, + 0.022041887044906616, + 0.006290895398706198, + 0.011896785348653793, + 0.010663839988410473, + 0.027029026299715042, + 0.07439325004816055, + 0.033025916665792465, + 0.04152630269527435, + 0.03513285145163536, + 0.03483615070581436, + 0.060477789491415024, + 0.03624112159013748, + 0.007634032052010298, + 0.010538092814385891, + 0.007536097429692745, + 0.0057931868359446526, + 0.013715597800910473, + 0.011401784606277943, + 0.006983939558267593, + 0.0072232987731695175, + 0.006302203517407179, + 0.015364096499979496, + 0.007955046370625496, + 0.00889118667691946, + 0.01840992085635662, + 0.015408804640173912, + 0.01225708331912756, + 0.010702845640480518, + 0.004808098077774048, + 0.008079422637820244, + 0.0046555399894714355, + 0.009501430206000805, + 0.010478608310222626, + 0.009178470820188522, + 0.008028627373278141, + 0.006910480093210936, + 0.005136585794389248, + 0.004847314674407244, + 0.004116317257285118, + 0.00299282930791378, + 0.0039022231940180063, + 0.004233756568282843, + 0.004485429730266333, + 0.0032082325778901577, + 0.017857249826192856, + 0.011508877389132977, + 0.0041566286236047745, + 0.004395364783704281, + 0.043173421174287796, + 0.02591470070183277, + 0.004614582750946283, + 0.00456096138805151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.20156994462013245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, close-up tracking shot follows a shaggy white puppy, its pristine white fur speckled with a few fallen leaves. Its obsidian-bright eyes gleam with concentration, its small, wet black nose pressed to the ground, sniffing excitedly through the thick layer of fallen leaves. Soft afternoon sunlight filters through the treetops, casting dappled shadows on the golden, crimson, and brown carpet of fallen leaves, illuminating the fine dust particles hanging in the air. Suddenly, the puppy seems to lock onto its target. It pauses its sniffing and begins digging with its front paws, alternating between rapid and vigorous movements. The soft, dark earth and fallen leaves are flung behind it, revealing its pink paw pads. Its entire body trembles with excitement as it searches intently for the secret buried beneath. The warm tones of the entire scene imbue the scene with a vibrant sense of movement and pure, curious joy.", + "seed": 16691, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_11/index_8691_seed_16691_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.053023916226811706, + "vlm_logprob": -0.053023916226811706, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.053023916226811706, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021848460659384727 + ], + "opposite_logprob_tokens": [ + -4.271848678588867 + ], + "logprob_reward": -0.021848460659384727, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.021848460659384727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13224968314170837, + 0.0462728887796402, + 0.05282352492213249, + 0.1037830039858818, + 0.2852274179458618, + 0.06727953255176544, + 0.1195206493139267, + 0.1052628681063652, + 0.04971841722726822, + 0.03786195442080498 + ], + "credit_map": [ + 0.010646460577845573, + 0.02245614491403103, + 0.01143172662705183, + 0.009981227107346058, + 0.028461989015340805, + 0.01726800948381424, + 0.02155252732336521, + 0.0104516027495265, + 0.003489058930426836, + 0.008798853494226933, + 0.002849582117050886, + 0.003202539635822177, + 0.010800055228173733, + 0.007523738779127598, + 0.004946921020746231, + 0.004662145394831896, + 0.0076925805769860744, + 0.006067920941859484, + 0.007809806615114212, + 0.007061006035655737, + 0.007443554233759642, + 0.006274646148085594, + 0.005325698759406805, + 0.005148311145603657, + 0.01858679950237274, + 0.014042865484952927, + 0.014631160534918308, + 0.008623387664556503, + 0.013453454710543156, + 0.01558571495115757, + 0.010833227075636387, + 0.00802640151232481, + 0.01410833839327097, + 0.08874266594648361, + 0.014452910050749779, + 0.012398136779665947, + 0.03066430054605007, + 0.09916365146636963, + 0.014003980904817581, + 0.011693454347550869, + 0.0059064943343400955, + 0.005735021084547043, + 0.010134696029126644, + 0.006503558252006769, + 0.01478827465325594, + 0.007877741940319538, + 0.006274404935538769, + 0.010059349238872528, + 0.014374151825904846, + 0.011965322308242321, + 0.009406903758645058, + 0.036683447659015656, + 0.006436599884182215, + 0.008133705705404282, + 0.012193793430924416, + 0.02032674103975296, + 0.005428334232419729, + 0.03341779112815857, + 0.00719257490709424, + 0.004529785830527544, + 0.01420949213206768, + 0.008843474090099335, + 0.023635094985365868, + 0.00800632406026125, + 0.006616631057113409, + 0.011807303875684738, + 0.007846830412745476, + 0.0061390888877213, + 0.004780557472258806, + 0.004281757865101099, + 0.005155419930815697, + 0.0030908319167792797, + 0.003534032264724374, + 0.0041723353788256645, + 0.005004514940083027, + 0.005356485024094582, + 0.005473901983350515, + 0.004619227722287178, + 0.005967249628156424, + 0.0037342123687267303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.021848460659384727 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812013566493988 + ], + "opposite_logprob_tokens": [ + -3.2881200313568115 + ], + "logprob_reward": -0.03812013566493988, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812013566493988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3951414227485657, + 0.11488404124975204, + 0.042946215718984604, + 0.03949408233165741, + 0.10658121854066849, + 0.0395718514919281, + 0.09816600382328033, + 0.09505172073841095, + 0.04050276800990105, + 0.02766072377562523 + ], + "credit_map": [ + 0.017444439232349396, + 0.054606202989816666, + 0.014027404598891735, + 0.014837516471743584, + 0.1383131742477417, + 0.11717308312654495, + 0.025660187005996704, + 0.013079402036964893, + 0.004925186280161142, + 0.006154310889542103, + 0.00654687499627471, + 0.003389725461602211, + 0.03864451125264168, + 0.04201972112059593, + 0.005095082335174084, + 0.008108628913760185, + 0.004202153068035841, + 0.004767191130667925, + 0.0036801465321332216, + 0.006356914062052965, + 0.00572254927828908, + 0.0043273563496768475, + 0.006909145507961512, + 0.006980759557336569, + 0.007095266133546829, + 0.004664340056478977, + 0.005748377647250891, + 0.0037718056701123714, + 0.005013344343751669, + 0.0050325216725468636, + 0.004486478865146637, + 0.0036819481756538153, + 0.005477974656969309, + 0.03059310093522072, + 0.007598766591399908, + 0.005694659426808357, + 0.012586652301251888, + 0.03266162797808647, + 0.006572296842932701, + 0.005396141204982996, + 0.002872087061405182, + 0.006134620867669582, + 0.0056821261532604694, + 0.002794150495901704, + 0.006392577197402716, + 0.008027501404285431, + 0.00278939469717443, + 0.004879394546151161, + 0.006794904358685017, + 0.00630429619923234, + 0.007420073729008436, + 0.02435186132788658, + 0.0040699150413274765, + 0.006830927450209856, + 0.014829817228019238, + 0.027564212679862976, + 0.007360788993537426, + 0.024872230365872383, + 0.008116881363093853, + 0.00610353983938694, + 0.012115499004721642, + 0.007409157697111368, + 0.01867869310081005, + 0.010394931770861149, + 0.0045109400525689125, + 0.006838473491370678, + 0.0066996486857533455, + 0.005184933543205261, + 0.003343701595440507, + 0.005650498438626528, + 0.004576642531901598, + 0.0036979294382035732, + 0.002720578107982874, + 0.003325433237478137, + 0.003931493032723665, + 0.002922014333307743, + 0.0038254219107329845, + 0.0032513942569494247, + 0.004082430619746447, + 0.00360195804387331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.03812013566493988 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306933656334877 + ], + "opposite_logprob_tokens": [ + -3.168069362640381 + ], + "logprob_reward": -0.04306933656334877, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306933656334877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20972317457199097, + 0.14806631207466125, + 0.0581219382584095, + 0.09089082479476929, + 0.024208327755331993, + 0.014293421059846878, + 0.18460486829280853, + 0.20212894678115845, + 0.024807719513773918, + 0.04315447807312012 + ], + "credit_map": [ + 0.007427912671118975, + 0.01567809469997883, + 0.009949839673936367, + 0.009167933836579323, + 0.07245872914791107, + 0.07301389425992966, + 0.014476743526756763, + 0.007550023030489683, + 0.004022797103971243, + 0.01760462112724781, + 0.004790056962519884, + 0.004454384557902813, + 0.04236384108662605, + 0.0603252574801445, + 0.005563165061175823, + 0.008942194283008575, + 0.005450642667710781, + 0.008783334866166115, + 0.005249901674687862, + 0.006367824971675873, + 0.010053527541458607, + 0.008245495148003101, + 0.007899717427790165, + 0.006071493495255709, + 0.01826588250696659, + 0.010753374546766281, + 0.014420287683606148, + 0.006915170233696699, + 0.011424885131418705, + 0.012506282888352871, + 0.010058005340397358, + 0.006546926684677601, + 0.0020770919509232044, + 0.002632957184687257, + 0.001981808803975582, + 0.0038289690855890512, + 0.004999010358005762, + 0.003882157849147916, + 0.0027169871609658003, + 0.0020893465261906385, + 0.001443705870769918, + 0.0014048421289771795, + 0.0020257162395864725, + 0.0017692323308438063, + 0.0026067420840263367, + 0.001733560930006206, + 0.001491702743805945, + 0.001817918848246336, + 0.021581564098596573, + 0.014923730865120888, + 0.015422212891280651, + 0.05587105453014374, + 0.012847223319113255, + 0.01637989841401577, + 0.01838040165603161, + 0.02919878251850605, + 0.013536255806684494, + 0.058260947465896606, + 0.014556393027305603, + 0.011287102475762367, + 0.025206878781318665, + 0.018206899985671043, + 0.042406339198350906, + 0.01866813376545906, + 0.0036417730152606964, + 0.005353573244065046, + 0.003819198813289404, + 0.002912577474489808, + 0.0025680926628410816, + 0.0022955676540732384, + 0.0024001433048397303, + 0.0018167926464229822, + 0.003348852274939418, + 0.005593376699835062, + 0.0039669848047196865, + 0.006351425778120756, + 0.007049960549920797, + 0.006575206760317087, + 0.006005001254379749, + 0.004263666924089193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.04306933656334877 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017146660247817636 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017146660247817636, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017146660247817636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27068009972572327, + 0.06981540471315384, + 0.05924620479345322, + 0.033872127532958984, + 0.22735027968883514, + 0.060632653534412384, + 0.11586742103099823, + 0.10052543878555298, + 0.02198086306452751, + 0.04002953693270683 + ], + "credit_map": [ + 0.028145121410489082, + 0.07944326102733612, + 0.016407717019319534, + 0.013611826114356518, + 0.03377074375748634, + 0.06289365142583847, + 0.020111095160245895, + 0.016296686604619026, + 0.004403666127473116, + 0.017774619162082672, + 0.0032485683914273977, + 0.0031739859841763973, + 0.010807446204125881, + 0.023578541353344917, + 0.0036476491950452328, + 0.0031809282954782248, + 0.006706609856337309, + 0.00793042778968811, + 0.007042431738227606, + 0.008453219197690487, + 0.010043811984360218, + 0.0067258491180837154, + 0.007091491483151913, + 0.005252365954220295, + 0.003464222652837634, + 0.005348348990082741, + 0.0034030971582978964, + 0.0033100920263677835, + 0.004906065762042999, + 0.0063769216649234295, + 0.0041907294653356075, + 0.002872651442885399, + 0.009324364364147186, + 0.07581135630607605, + 0.012245659716427326, + 0.009493441320955753, + 0.023463400080800056, + 0.07937055826187134, + 0.008564099669456482, + 0.009077402763068676, + 0.005451650358736515, + 0.006799055729061365, + 0.011628790758550167, + 0.004365425556898117, + 0.008366860449314117, + 0.0067030880600214005, + 0.004755516070872545, + 0.012562265619635582, + 0.012566604651510715, + 0.008052388206124306, + 0.01143806241452694, + 0.02790737897157669, + 0.006793967913836241, + 0.012285674922168255, + 0.013085813261568546, + 0.023737531155347824, + 0.008539029397070408, + 0.024523429572582245, + 0.007504694163799286, + 0.01038917526602745, + 0.011284705251455307, + 0.009467357769608498, + 0.01642303727567196, + 0.012394008226692677, + 0.00403574900701642, + 0.0030164311174303293, + 0.0031927975360304117, + 0.003106475807726383, + 0.0018432991346344352, + 0.0027004575822502375, + 0.002168121514841914, + 0.0019175339257344604, + 0.0049068438820540905, + 0.004978447686880827, + 0.0053001148626208305, + 0.004996600095182657, + 0.005587645806372166, + 0.0037157374899834394, + 0.007166511379182339, + 0.0033776387572288513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0017146660247817636 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle close-up tracking shot following a shaggy white puppy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the puppy's fur white with a few fallen leaves speckled on it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the puppy pause its sniffing to dig with its front paws, revealing pink paw pads?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a ground covered in a thick layer of golden, crimson, and brown fallen leaves with dappled sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the puppy's body tremble with excitement while it searches intently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603669822216034 + ], + "opposite_logprob_tokens": [ + -1.9103670120239258 + ], + "logprob_reward": -0.1603669822216034, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603669822216034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13963130116462708, + 0.025759976357221603, + 0.10589874535799026, + 0.03400749713182449, + 0.2089555561542511, + 0.03039564937353134, + 0.27822253108024597, + 0.0740928053855896, + 0.0454210489988327, + 0.05761497840285301 + ], + "credit_map": [ + 0.011724932119250298, + 0.022123631089925766, + 0.011702561751008034, + 0.012333887629210949, + 0.02074667438864708, + 0.032250914722681046, + 0.01966232992708683, + 0.009086348116397858, + 0.00214915769174695, + 0.003389237215742469, + 0.0023358138278126717, + 0.0019882211927324533, + 0.00355381122790277, + 0.006447272840887308, + 0.0027703847736120224, + 0.003126075491309166, + 0.013624822720885277, + 0.010229733772575855, + 0.014253637753427029, + 0.018177520483732224, + 0.015456822700798512, + 0.015273003838956356, + 0.009833096526563168, + 0.0090500982478261, + 0.004614615812897682, + 0.004256709944456816, + 0.005442514084279537, + 0.002595009282231331, + 0.005199916660785675, + 0.005573354195803404, + 0.004073841962963343, + 0.0022515340242534876, + 0.008064603433012962, + 0.06881862878799438, + 0.009811406023800373, + 0.00999364908784628, + 0.020754879340529442, + 0.07478508353233337, + 0.00831686146557331, + 0.008410421200096607, + 0.003103828988969326, + 0.004099425859749317, + 0.0038613122887909412, + 0.0020114353392273188, + 0.00480526452884078, + 0.0062346383929252625, + 0.002671258058398962, + 0.00360848312266171, + 0.03863061964511871, + 0.02521049603819847, + 0.023763902485370636, + 0.07864660024642944, + 0.021050874143838882, + 0.02344835363328457, + 0.03068198636174202, + 0.0367896668612957, + 0.007901364006102085, + 0.01632651500403881, + 0.005523144267499447, + 0.005651378538459539, + 0.008358730003237724, + 0.008869408629834652, + 0.012193292379379272, + 0.009268970228731632, + 0.00445396825671196, + 0.005548824090510607, + 0.005965431220829487, + 0.0060868458822369576, + 0.004695196636021137, + 0.008860516361892223, + 0.005232366733253002, + 0.004577895626425743, + 0.007017933297902346, + 0.006367532070726156, + 0.0057906839065253735, + 0.0080214599147439, + 0.008396457880735397, + 0.007874415256083012, + 0.009508815594017506, + 0.004637675825506449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.1603669822216034 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_12/index_tensor([8050, 8050])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_12/index_tensor([8050, 8050])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dd8f6be8a1116489da30c315fffb4b7e6d9a4f02 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_12/index_tensor([8050, 8050])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot with the precision of ultra-high-speed photography, this extreme close-up utilizes a static camera to depict the climactic moment of impact. A projectile, fired from an ornate M16 rifle, its gleaming, intricately engraved gold receiver partially visible at the edge of the frame, strikes a dark, cold-rolled steel surface. The camera employs a razor-sharp depth of field, maintaining impeccable, sharp focus solely on the bullet. As it makes contact, the entire event unfolds in breathtaking slow motion; the copper tip curls instantly, violently impacting the solid steel. This catastrophic collision transforms kinetic energy into a brilliant, explosive shower of incandescent sparks, blasting outward in a fiery starburst pattern, their white-hot trails starkly illuminating the scene against a softly blurred, dark, and indistinct background. The now-damaged projectile, spinning to a halt and tumbling away from its point of impact, its twisted metal surface captures dramatic key lighting as it gracefully ricochets through the air, creating a mesmerizing and powerful atmosphere of controlled destruction.", + "seed": 1419, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_12/index_8050_seed_1419_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.22990563102066516, + "vlm_logprob": -0.22990563102066516, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.22990563102066516, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.905570387840271 + ], + "opposite_logprob_tokens": [ + -0.530570387840271 + ], + "logprob_reward": -0.905570387840271, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.905570387840271, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09013096988201141, + 0.042870718985795975, + 0.026303553953766823, + 0.0950855016708374, + 0.04225415363907814, + 0.07858327031135559, + 0.16634373366832733, + 0.2831154763698578, + 0.1514468789100647, + 0.023865772411227226 + ], + "credit_map": [ + 0.008245803415775299, + 0.01720641739666462, + 0.009297607466578484, + 0.011094697751104832, + 0.011628782376646996, + 0.007443294394761324, + 0.017632432281970978, + 0.007581941317766905, + 0.006868608295917511, + 0.006061017978936434, + 0.006013805046677589, + 0.0044464911334216595, + 0.0022577736526727676, + 0.004398411139845848, + 0.0055745914578437805, + 0.007250023540109396, + 0.003218278754502535, + 0.003470039926469326, + 0.003429925302043557, + 0.0035539944656193256, + 0.0020776079036295414, + 0.00433120783418417, + 0.0028119429480284452, + 0.0034105563536286354, + 0.01176125556230545, + 0.014206347987055779, + 0.013492200523614883, + 0.00577034754678607, + 0.007591362576931715, + 0.01727585680782795, + 0.013252688571810722, + 0.011735438369214535, + 0.0046351514756679535, + 0.005032602231949568, + 0.005146600306034088, + 0.006014764774590731, + 0.005113659892231226, + 0.004500215873122215, + 0.0066835335455834866, + 0.005127623677253723, + 0.008255052380263805, + 0.014604452066123486, + 0.010871515609323978, + 0.0075944592244923115, + 0.00759120425209403, + 0.008939226157963276, + 0.0096110999584198, + 0.01111625600606203, + 0.017305508255958557, + 0.03092813678085804, + 0.023036330938339233, + 0.015493839979171753, + 0.017562493681907654, + 0.013050653971731663, + 0.027721650898456573, + 0.021245108917355537, + 0.027825329452753067, + 0.05156886950135231, + 0.03780932351946831, + 0.023653771728277206, + 0.028803622350096703, + 0.029160453006625175, + 0.045738354325294495, + 0.03855574503540993, + 0.014654417522251606, + 0.03237977996468544, + 0.01644621416926384, + 0.015628060325980186, + 0.020532570779323578, + 0.015943603590130806, + 0.01618158631026745, + 0.01968064159154892, + 0.0025286045856773853, + 0.0029306511860340834, + 0.0028439112938940525, + 0.002837494481354952, + 0.0033390072640031576, + 0.0031617959029972553, + 0.0032971252221614122, + 0.00292718387208879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.905570387840271 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16046994924545288 + ], + "opposite_logprob_tokens": [ + -1.9104700088500977 + ], + "logprob_reward": -0.16046994924545288, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16046994924545288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11027564853429794, + 0.06533852964639664, + 0.14352789521217346, + 0.08436568081378937, + 0.0879627913236618, + 0.05307634547352791, + 0.16709479689598083, + 0.14353099465370178, + 0.1234426349401474, + 0.021384725347161293 + ], + "credit_map": [ + 0.010123254731297493, + 0.019740941002964973, + 0.014060276560485363, + 0.01721550151705742, + 0.012040076777338982, + 0.01026761531829834, + 0.016506506130099297, + 0.010321474634110928, + 0.009320213459432125, + 0.012591676786541939, + 0.007871951907873154, + 0.006009374745190144, + 0.00489038834348321, + 0.005502328742295504, + 0.008508146740496159, + 0.010644444264471531, + 0.03129541873931885, + 0.03655313327908516, + 0.010068868286907673, + 0.011463341303169727, + 0.012463530525565147, + 0.01639215089380741, + 0.01067750807851553, + 0.014613949693739414, + 0.006185681559145451, + 0.012821953743696213, + 0.02290632762014866, + 0.00488855317234993, + 0.00439047534018755, + 0.009664240293204784, + 0.00748025719076395, + 0.01602819375693798, + 0.006905337329953909, + 0.012323996052145958, + 0.01349017582833767, + 0.016060642898082733, + 0.007329768501222134, + 0.011075835675001144, + 0.012424904853105545, + 0.008352131582796574, + 0.0055839139968156815, + 0.009476851671934128, + 0.008326773531734943, + 0.005260745529085398, + 0.005293551832437515, + 0.005539807491004467, + 0.005224043969064951, + 0.00837065652012825, + 0.01744251884520054, + 0.03406272083520889, + 0.020839225500822067, + 0.013888495974242687, + 0.019833194091916084, + 0.012420883402228355, + 0.026789290830492973, + 0.02181847020983696, + 0.014233396388590336, + 0.02461388148367405, + 0.02006949670612812, + 0.012061933055520058, + 0.013023077510297298, + 0.015151722356677055, + 0.02222393825650215, + 0.022153548896312714, + 0.012252725660800934, + 0.02622753381729126, + 0.01099047064781189, + 0.013026419095695019, + 0.01742643676698208, + 0.012888829223811626, + 0.01436273567378521, + 0.01626748777925968, + 0.002459472045302391, + 0.002766976598650217, + 0.0033708626870065928, + 0.0022664943244308233, + 0.00219749822281301, + 0.0023955546785146, + 0.0026286772917956114, + 0.0032991881016641855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.16046994924545288 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867817088961601 + ], + "opposite_logprob_tokens": [ + -3.048678159713745 + ], + "logprob_reward": -0.04867817088961601, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867817088961601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07841438055038452, + 0.07364676147699356, + 0.13973651826381683, + 0.11559418588876724, + 0.09579043835401535, + 0.10845871269702911, + 0.20966017246246338, + 0.08620594441890717, + 0.0288547333329916, + 0.06363816559314728 + ], + "credit_map": [ + 0.009553197771310806, + 0.011200389824807644, + 0.008734559640288353, + 0.010068518109619617, + 0.011016001924872398, + 0.009158474393188953, + 0.0091931764036417, + 0.009490063413977623, + 0.010981854051351547, + 0.01452562678605318, + 0.011069459840655327, + 0.009182030335068703, + 0.0040267049334943295, + 0.004527210257947445, + 0.008191925473511219, + 0.011141953989863396, + 0.02704218402504921, + 0.03889851272106171, + 0.014964698813855648, + 0.008843531832098961, + 0.01037999615073204, + 0.013282021507620811, + 0.010000117123126984, + 0.016325445845723152, + 0.011413966305553913, + 0.017419369891285896, + 0.01292995922267437, + 0.00991620123386383, + 0.010260294191539288, + 0.018812105059623718, + 0.012900575995445251, + 0.021941717714071274, + 0.008763568475842476, + 0.012416644021868706, + 0.013764097355306149, + 0.012187808752059937, + 0.009353010915219784, + 0.012260361574590206, + 0.011661786586046219, + 0.015383166261017323, + 0.012086804956197739, + 0.019808586686849594, + 0.013295941986143589, + 0.009891952387988567, + 0.01154907513409853, + 0.010955522768199444, + 0.011066262610256672, + 0.019804565235972404, + 0.027428802102804184, + 0.03719586506485939, + 0.034220609813928604, + 0.023261364549398422, + 0.015160406939685345, + 0.01781565509736538, + 0.027339404448866844, + 0.027238059788942337, + 0.00799765344709158, + 0.01195396576076746, + 0.016042334958910942, + 0.008886192925274372, + 0.00821495521813631, + 0.007894707843661308, + 0.012128710746765137, + 0.013087422586977482, + 0.002875053323805332, + 0.004676575306802988, + 0.0032069196458905935, + 0.004057940561324358, + 0.0032548331655561924, + 0.0028303752187639475, + 0.003212581854313612, + 0.004740454256534576, + 0.0076543246395885944, + 0.00726824626326561, + 0.011058307252824306, + 0.005621017888188362, + 0.005447418428957462, + 0.006340860389173031, + 0.007357263937592506, + 0.012890725396573544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.04867817088961601 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014220215380191803 + ], + "opposite_logprob_tokens": [ + -4.264220237731934 + ], + "logprob_reward": -0.014220215380191803, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014220215380191803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06868557631969452, + 0.07457840442657471, + 0.05308457463979721, + 0.1286391019821167, + 0.05781625956296921, + 0.05650734901428223, + 0.11484576761722565, + 0.2506585717201233, + 0.14401976764202118, + 0.051164597272872925 + ], + "credit_map": [ + 0.008440191857516766, + 0.011785902082920074, + 0.00865088403224945, + 0.009756078943610191, + 0.008102516643702984, + 0.005782278720289469, + 0.0096741309389472, + 0.006493587512522936, + 0.010228886269032955, + 0.01644803024828434, + 0.010616922751069069, + 0.007341863587498665, + 0.004634181037545204, + 0.006166856270283461, + 0.008899620734155178, + 0.010242043994367123, + 0.010777749121189117, + 0.013889999128878117, + 0.004648260772228241, + 0.004638077225536108, + 0.003886695485562086, + 0.005555347073823214, + 0.003722111228853464, + 0.005966332741081715, + 0.01848077028989792, + 0.020075945183634758, + 0.014270393177866936, + 0.009971581399440765, + 0.010169295594096184, + 0.024336060509085655, + 0.01715010218322277, + 0.014184955507516861, + 0.00472898269072175, + 0.006824096664786339, + 0.008147128857672215, + 0.009345641359686852, + 0.006055261939764023, + 0.005695523228496313, + 0.008026274852454662, + 0.008993344381451607, + 0.005497636739164591, + 0.009576213546097279, + 0.008188383653759956, + 0.005398728884756565, + 0.0051397900097072124, + 0.005509106907993555, + 0.008081031031906605, + 0.009116457775235176, + 0.011880864389240742, + 0.018581151962280273, + 0.015467584133148193, + 0.013512353412806988, + 0.013389254920184612, + 0.011897824704647064, + 0.016052505001425743, + 0.014064228162169456, + 0.023590918630361557, + 0.04037661850452423, + 0.038683969527482986, + 0.023345546796917915, + 0.02248726785182953, + 0.02616642788052559, + 0.04255637899041176, + 0.03345146030187607, + 0.014194373041391373, + 0.03141431882977486, + 0.016507146880030632, + 0.014586398378014565, + 0.01996535435318947, + 0.01506409328430891, + 0.014466407708823681, + 0.017821678891777992, + 0.0055918884463608265, + 0.006203614640980959, + 0.009188667871057987, + 0.0059667411260306835, + 0.006281084381043911, + 0.005864499602466822, + 0.005081175826489925, + 0.00698692724108696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.014220215380191803 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058943174779415 + ], + "opposite_logprob_tokens": [ + -3.8955893516540527 + ], + "logprob_reward": -0.02058943174779415, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058943174779415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08917655795812607, + 0.07006508857011795, + 0.25309842824935913, + 0.040981002151966095, + 0.08169832080602646, + 0.034574251621961594, + 0.06915082782506943, + 0.2195948213338852, + 0.10746601223945618, + 0.03419466316699982 + ], + "credit_map": [ + 0.009223682805895805, + 0.018090426921844482, + 0.01024576649069786, + 0.011196247301995754, + 0.009450241923332214, + 0.00859690923243761, + 0.012849720194935799, + 0.009523574262857437, + 0.010194934904575348, + 0.015248916111886501, + 0.010049225762486458, + 0.006363701540976763, + 0.0038295607082545757, + 0.004541235044598579, + 0.009726449847221375, + 0.010111070238053799, + 0.04316137358546257, + 0.05589321628212929, + 0.027038613334298134, + 0.024014679715037346, + 0.018015675246715546, + 0.030911484733223915, + 0.023145627230405807, + 0.030917758122086525, + 0.004013551399111748, + 0.005152184050530195, + 0.006222234573215246, + 0.0047818757593631744, + 0.0035881439689546824, + 0.006036136765033007, + 0.006400180049240589, + 0.004786697681993246, + 0.00881155114620924, + 0.011664078570902348, + 0.01000817958265543, + 0.010555535554885864, + 0.007569663226604462, + 0.011469528079032898, + 0.011571631766855717, + 0.0100481566041708, + 0.0038892487064003944, + 0.005409713368862867, + 0.005037419032305479, + 0.0034714119974523783, + 0.005367905832827091, + 0.0029684165492653847, + 0.004410749301314354, + 0.004019390791654587, + 0.00715322932228446, + 0.011789081618189812, + 0.009363055229187012, + 0.005663159769028425, + 0.006212357897311449, + 0.005025908350944519, + 0.011403980664908886, + 0.012540057301521301, + 0.021200601011514664, + 0.039990417659282684, + 0.025557054206728935, + 0.017229730263352394, + 0.022066546604037285, + 0.02399851754307747, + 0.03826040402054787, + 0.03129156678915024, + 0.012151230126619339, + 0.02367212437093258, + 0.008506839163601398, + 0.014961380511522293, + 0.013279332779347897, + 0.011235933750867844, + 0.010810577310621738, + 0.01284859050065279, + 0.0034771354403346777, + 0.003784024389460683, + 0.004651936702430248, + 0.0038853001315146685, + 0.004314206074923277, + 0.005123374983668327, + 0.0038428304251283407, + 0.005115855019539595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.02058943174779415 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with the precision of ultra-high-speed photography, this extreme close-up utilizes a static camera to depict the climactic moment of impact. A projectile, fired from an ornate M16 rifle, its gleaming, intricately engraved gold receiver partially visible at the edge of the frame, strikes a dark, cold-rolled steel surface. The camera employs a razor-sharp depth of field, maintaining impeccable, sharp focus solely on the bullet. As it makes contact, the entire event unfolds in breathtaking slow motion; the copper tip curls instantly, violently impacting the solid steel. This catastrophic collision transforms kinetic energy into a brilliant, explosive shower of incandescent sparks, blasting outward in a fiery starburst pattern, their white-hot trails starkly illuminating the scene against a softly blurred, dark, and indistinct background. The now-damaged projectile, spinning to a halt and tumbling away from its point of impact, its twisted metal surface captures dramatic key lighting as it gracefully ricochets through the air, creating a mesmerizing and powerful atmosphere of controlled destruction.", + "seed": 1420, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_12/index_8050_seed_1420_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.082299380376935, + "vlm_logprob": -0.082299380376935, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.082299380376935, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24182289838790894 + ], + "opposite_logprob_tokens": [ + -1.6168229579925537 + ], + "logprob_reward": -0.24182289838790894, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24182289838790894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03462597727775574, + 0.06110133230686188, + 0.09301543235778809, + 0.532537043094635, + 0.04490141198039055, + 0.0879136323928833, + 0.06337279826402664, + 0.04496338218450546, + 0.023698659613728523, + 0.013870327733457088 + ], + "credit_map": [ + 0.0038126001600176096, + 0.006548348348587751, + 0.004218065645545721, + 0.005807168316096067, + 0.004023291170597076, + 0.0027409973554313183, + 0.00451307650655508, + 0.002962426282465458, + 0.004652156960219145, + 0.010284828022122383, + 0.0050774868577718735, + 0.0052581727504730225, + 0.01811150833964348, + 0.0032340416219085455, + 0.009513789787888527, + 0.00496934587135911, + 0.008260635659098625, + 0.013364213518798351, + 0.013028590939939022, + 0.011566722765564919, + 0.012987430207431316, + 0.011647041887044907, + 0.012294361367821693, + 0.00986642949283123, + 0.06110704690217972, + 0.17062152922153473, + 0.14280179142951965, + 0.055229753255844116, + 0.04403255879878998, + 0.015638206154108047, + 0.017920590937137604, + 0.025185564532876015, + 0.004969585686922073, + 0.007457354571670294, + 0.0056755635887384415, + 0.004821066744625568, + 0.005487046204507351, + 0.006533435545861721, + 0.005207084119319916, + 0.004750275518745184, + 0.006047806236892939, + 0.010490872897207737, + 0.014428123831748962, + 0.011096800677478313, + 0.010857298970222473, + 0.011033639311790466, + 0.013486068695783615, + 0.010473025031387806, + 0.005741628352552652, + 0.008544509299099445, + 0.013194230385124683, + 0.010143222287297249, + 0.005509797018021345, + 0.006961507257074118, + 0.006750498432666063, + 0.006527408957481384, + 0.005129852797836065, + 0.0065190535970032215, + 0.0052634733729064465, + 0.009974107146263123, + 0.00391386216506362, + 0.004699986893683672, + 0.0038999561220407486, + 0.005563090089708567, + 0.0028792615048587322, + 0.0025457392912358046, + 0.004522606264799833, + 0.0026965243741869926, + 0.0021404651924967766, + 0.003133510472252965, + 0.0028914345894008875, + 0.002889118855819106, + 0.001335677457973361, + 0.0024151005782186985, + 0.0019886083900928497, + 0.0026578924153000116, + 0.0010338471038267016, + 0.0013793990947306156, + 0.0016299973940476775, + 0.0014298052992671728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.24182289838790894 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008294761180878 + ], + "opposite_logprob_tokens": [ + -2.695082902908325 + ], + "logprob_reward": -0.07008294761180878, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008294761180878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.296598345041275, + 0.11258527636528015, + 0.01493404433131218, + 0.03058876097202301, + 0.022645195946097374, + 0.3089231252670288, + 0.06334058195352554, + 0.08906160295009613, + 0.041788194328546524, + 0.019534900784492493 + ], + "credit_map": [ + 0.02990330010652542, + 0.04697969928383827, + 0.030841993167996407, + 0.037862908095121384, + 0.03880210965871811, + 0.02740737982094288, + 0.03760404884815216, + 0.04719691723585129, + 0.010155641473829746, + 0.013137626461684704, + 0.021704941987991333, + 0.014486751519143581, + 0.015737898647785187, + 0.009901552461087704, + 0.010104916989803314, + 0.017355943098664284, + 0.0013545932015404105, + 0.00172680604737252, + 0.0019242503913119435, + 0.0024686185643076897, + 0.002776727546006441, + 0.0014383606612682343, + 0.0015987642109394073, + 0.0016459237085655332, + 0.0026079623494297266, + 0.0032915431074798107, + 0.004629918374121189, + 0.006027305033057928, + 0.003971496596932411, + 0.002324205357581377, + 0.0031640271190553904, + 0.004572304431349039, + 0.0019810074009001255, + 0.003319758689031005, + 0.0032889575231820345, + 0.002789733000099659, + 0.0024248301051557064, + 0.0024189127143472433, + 0.0022275198716670275, + 0.004194476641714573, + 0.021859219297766685, + 0.05074085295200348, + 0.05651240050792694, + 0.04389353469014168, + 0.0235025305300951, + 0.04892941564321518, + 0.022591527551412582, + 0.040893636643886566, + 0.004727482330054045, + 0.008562954142689705, + 0.011185117065906525, + 0.00884510762989521, + 0.006220473442226648, + 0.005758434534072876, + 0.010158094577491283, + 0.007882918231189251, + 0.005948799662292004, + 0.015076594427227974, + 0.011445378884673119, + 0.03243986889719963, + 0.002235103864222765, + 0.0028667065780609846, + 0.003174161072820425, + 0.015874994918704033, + 0.00498977629467845, + 0.005561159458011389, + 0.00570153770968318, + 0.0047505744732916355, + 0.005044503137469292, + 0.005503153894096613, + 0.00522986426949501, + 0.005007624160498381, + 0.0019999691285192966, + 0.0023665742482990026, + 0.004303859546780586, + 0.0023073868360370398, + 0.0022057790774852037, + 0.002015045378357172, + 0.0023370005656033754, + 0.001999284839257598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.07008294761180878 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207573413848877 + ], + "opposite_logprob_tokens": [ + -2.812075614929199 + ], + "logprob_reward": -0.06207573413848877, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207573413848877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14558252692222595, + 0.056984059512615204, + 0.0529300794005394, + 0.14425712823867798, + 0.07091024518013, + 0.302150696516037, + 0.05976438522338867, + 0.036417849361896515, + 0.10610923171043396, + 0.024893781170248985 + ], + "credit_map": [ + 0.014919286593794823, + 0.022130336612462997, + 0.013451461680233479, + 0.016794705763459206, + 0.0190386101603508, + 0.014759664423763752, + 0.01754123531281948, + 0.026947224512696266, + 0.007826907560229301, + 0.007595454808324575, + 0.008008790202438831, + 0.007880378514528275, + 0.008422549813985825, + 0.0054956465028226376, + 0.00650267256423831, + 0.005251662340015173, + 0.006038534455001354, + 0.007371313404291868, + 0.008945474401116371, + 0.0073492033407092094, + 0.008527944795787334, + 0.0028104388620704412, + 0.004673589952290058, + 0.0072135827504098415, + 0.014258312061429024, + 0.036515265703201294, + 0.03725679591298103, + 0.02283484674990177, + 0.012238210067152977, + 0.005952286999672651, + 0.006276397965848446, + 0.008925016969442368, + 0.006129071116447449, + 0.012281821109354496, + 0.013127562589943409, + 0.00757091399282217, + 0.006611023098230362, + 0.009872937574982643, + 0.008044101297855377, + 0.007272814866155386, + 0.020446384325623512, + 0.04929463192820549, + 0.059194643050432205, + 0.040839746594429016, + 0.0219756867736578, + 0.050760261714458466, + 0.020325997844338417, + 0.03931334614753723, + 0.00441393069922924, + 0.008629536256194115, + 0.008354047313332558, + 0.00768940057605505, + 0.006018093321472406, + 0.006217881105840206, + 0.011204726994037628, + 0.007236772682517767, + 0.002648415043950081, + 0.003080048132687807, + 0.0064147054217755795, + 0.011812286451458931, + 0.0017619645223021507, + 0.0022609292063862085, + 0.0026680310256779194, + 0.005771468859165907, + 0.011874357238411903, + 0.013765290379524231, + 0.013691725209355354, + 0.012896613217890263, + 0.010084542445838451, + 0.014989210292696953, + 0.013906974345445633, + 0.014900517649948597, + 0.0026653644163161516, + 0.003593438072130084, + 0.003938548732548952, + 0.0036962516605854034, + 0.0021444717422127724, + 0.002755141817033291, + 0.0029915575869381428, + 0.0031090076081454754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.06207573413848877 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014225034043192863 + ], + "opposite_logprob_tokens": [ + -4.264225006103516 + ], + "logprob_reward": -0.014225034043192863, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014225034043192863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05776131898164749, + 0.060389433056116104, + 0.038585927337408066, + 0.2458958923816681, + 0.041785627603530884, + 0.0977032333612442, + 0.1062631830573082, + 0.2792711555957794, + 0.01810610480606556, + 0.05423809587955475 + ], + "credit_map": [ + 0.006478911265730858, + 0.011888213455677032, + 0.007174496538937092, + 0.008125984109938145, + 0.008458119817078114, + 0.004272768273949623, + 0.005893629975616932, + 0.005469196941703558, + 0.004895180929452181, + 0.009715099819004536, + 0.004431745503097773, + 0.0055128890089690685, + 0.018508508801460266, + 0.003991677891463041, + 0.009656538255512714, + 0.0036777902860194445, + 0.003456100355833769, + 0.005676488392055035, + 0.0061633228324353695, + 0.004550067707896233, + 0.00493493489921093, + 0.003518525743857026, + 0.005782473366707563, + 0.004504012875258923, + 0.02801685594022274, + 0.07756663858890533, + 0.06409226357936859, + 0.02304941974580288, + 0.021303987130522728, + 0.00987655483186245, + 0.01008656620979309, + 0.011903603561222553, + 0.004565445240586996, + 0.007258521392941475, + 0.007111445069313049, + 0.0050827013328671455, + 0.0033646177034825087, + 0.004578583408147097, + 0.0033674975857138634, + 0.006456816103309393, + 0.006001313682645559, + 0.012964744120836258, + 0.014212515205144882, + 0.013029503636062145, + 0.012841041199862957, + 0.009223893284797668, + 0.016338828951120377, + 0.013091396540403366, + 0.0075236172415316105, + 0.015363628044724464, + 0.016464315354824066, + 0.013804293237626553, + 0.011161673814058304, + 0.00903340894728899, + 0.021753158420324326, + 0.01115909218788147, + 0.014700307510793209, + 0.03743482381105423, + 0.0437421016395092, + 0.1092786118388176, + 0.006124240346252918, + 0.00716964527964592, + 0.008725056424736977, + 0.05209637060761452, + 0.002091027097776532, + 0.0017151052597910166, + 0.0029911522287875414, + 0.002228884492069483, + 0.0014082861598581076, + 0.002785974880680442, + 0.002686151070520282, + 0.0021995222195982933, + 0.005533414892852306, + 0.008488338440656662, + 0.010485070757567883, + 0.010480569675564766, + 0.003350917249917984, + 0.004730260465294123, + 0.0073808543384075165, + 0.003788665635511279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.014225034043192863 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329028770327568 + ], + "opposite_logprob_tokens": [ + -3.7732903957366943 + ], + "logprob_reward": -0.02329028770327568, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329028770327568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22580371797084808, + 0.030130213126540184, + 0.020021310076117516, + 0.07637286931276321, + 0.025115003809332848, + 0.39848271012306213, + 0.036187347024679184, + 0.05737277865409851, + 0.10761032998561859, + 0.02290370501577854 + ], + "credit_map": [ + 0.02289576269686222, + 0.03440709412097931, + 0.021487664431333542, + 0.025626415386795998, + 0.028699427843093872, + 0.025791136547923088, + 0.02889954298734665, + 0.03799668326973915, + 0.002638353267684579, + 0.004321322776377201, + 0.0037753938231617212, + 0.003776748199015856, + 0.0071622165851294994, + 0.0019279175903648138, + 0.003944368567317724, + 0.0025838916189968586, + 0.0017554330406710505, + 0.002644268097355962, + 0.0027127740904688835, + 0.002595145022496581, + 0.004076712764799595, + 0.0011206887429580092, + 0.001832496258430183, + 0.003283791011199355, + 0.008580130524933338, + 0.022633343935012817, + 0.01796790212392807, + 0.006155084352940321, + 0.007829053327441216, + 0.004858100321143866, + 0.003488748101517558, + 0.004860507324337959, + 0.0018127226503565907, + 0.003321721451357007, + 0.00287718721665442, + 0.002807185286656022, + 0.003319110721349716, + 0.0024379915557801723, + 0.004697269294410944, + 0.0038418143521994352, + 0.025480888783931732, + 0.0669403001666069, + 0.07217618077993393, + 0.052920278161764145, + 0.031543318182229996, + 0.07017935812473297, + 0.03139199689030647, + 0.047850366681814194, + 0.0025690265465527773, + 0.004318860825151205, + 0.0048028347082436085, + 0.005161639302968979, + 0.003821408376097679, + 0.003578717587515712, + 0.0068930271081626415, + 0.005041833966970444, + 0.004044658504426479, + 0.01189314853399992, + 0.013464635238051414, + 0.010107566602528095, + 0.0033311708830296993, + 0.004238910507410765, + 0.004701029974967241, + 0.005591662600636482, + 0.010708460584282875, + 0.014566784724593163, + 0.018017103895545006, + 0.017385907471179962, + 0.013299118727445602, + 0.012370387092232704, + 0.00981124583631754, + 0.011451326310634613, + 0.002167742233723402, + 0.0034827927593141794, + 0.0029732196126133204, + 0.004614198114722967, + 0.002068765228614211, + 0.0022701190318912268, + 0.003223537700250745, + 0.002103329636156559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.02329028770327568 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with the precision of ultra-high-speed photography, this extreme close-up utilizes a static camera to depict the climactic moment of impact. A projectile, fired from an ornate M16 rifle, its gleaming, intricately engraved gold receiver partially visible at the edge of the frame, strikes a dark, cold-rolled steel surface. The camera employs a razor-sharp depth of field, maintaining impeccable, sharp focus solely on the bullet. As it makes contact, the entire event unfolds in breathtaking slow motion; the copper tip curls instantly, violently impacting the solid steel. This catastrophic collision transforms kinetic energy into a brilliant, explosive shower of incandescent sparks, blasting outward in a fiery starburst pattern, their white-hot trails starkly illuminating the scene against a softly blurred, dark, and indistinct background. The now-damaged projectile, spinning to a halt and tumbling away from its point of impact, its twisted metal surface captures dramatic key lighting as it gracefully ricochets through the air, creating a mesmerizing and powerful atmosphere of controlled destruction.", + "seed": 1421, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_12/index_8050_seed_1421_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.15880286768078805, + "vlm_logprob": -0.15880286768078805, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.15880286768078805, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.647777259349823 + ], + "opposite_logprob_tokens": [ + -0.772777259349823 + ], + "logprob_reward": -0.647777259349823, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.647777259349823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031252842396497726, + 0.024875206872820854, + 0.042759303003549576, + 0.1143278032541275, + 0.13045400381088257, + 0.08582749217748642, + 0.1448964625597, + 0.19353553652763367, + 0.17177337408065796, + 0.06029798462986946 + ], + "credit_map": [ + 0.0035843166988343, + 0.0033987730275839567, + 0.004851638339459896, + 0.005089998710900545, + 0.003337952308356762, + 0.003852701047435403, + 0.003225921653211117, + 0.003911541309207678, + 0.0022503992076963186, + 0.0031707186717540026, + 0.002955677220597863, + 0.005987732205539942, + 0.0021850692573934793, + 0.002974253613501787, + 0.002418163698166609, + 0.002933192066848278, + 0.005512418691068888, + 0.004354423843324184, + 0.00422766525298357, + 0.0064792572520673275, + 0.003818353172391653, + 0.006496971473097801, + 0.0062405760399997234, + 0.0056296344846487045, + 0.009213536977767944, + 0.027214886620640755, + 0.04273397848010063, + 0.010749719105660915, + 0.007259938400238752, + 0.004894865211099386, + 0.00738505320623517, + 0.004875831305980682, + 0.016056085005402565, + 0.023265084251761436, + 0.01855645328760147, + 0.01664484106004238, + 0.01676912233233452, + 0.015011045150458813, + 0.012821020558476448, + 0.011330352164804935, + 0.010724062100052834, + 0.014120692387223244, + 0.012360596098005772, + 0.011116379871964455, + 0.009280786849558353, + 0.009781774133443832, + 0.009125344455242157, + 0.009317856281995773, + 0.016731008887290955, + 0.020174559205770493, + 0.02324874885380268, + 0.02756507322192192, + 0.01080851536244154, + 0.010870987549424171, + 0.019007733091711998, + 0.016489842906594276, + 0.02446071058511734, + 0.0382230319082737, + 0.028151651844382286, + 0.02545565366744995, + 0.01982698030769825, + 0.01486780121922493, + 0.017397664487361908, + 0.025152048096060753, + 0.024559667333960533, + 0.03090743161737919, + 0.02579652890563011, + 0.019108721986413002, + 0.018648020923137665, + 0.016265535727143288, + 0.018065674230456352, + 0.018421804532408714, + 0.006281673442572355, + 0.008277369663119316, + 0.008028995245695114, + 0.007409811485558748, + 0.009701956063508987, + 0.006977911572903395, + 0.006264818366616964, + 0.007355447858572006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.647777259349823 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07904396951198578 + ], + "opposite_logprob_tokens": [ + -2.5790438652038574 + ], + "logprob_reward": -0.07904396951198578, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07904396951198578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05149633437395096, + 0.028578998520970345, + 0.12381263822317123, + 0.4504338502883911, + 0.055755868554115295, + 0.053184863179922104, + 0.14298032224178314, + 0.05571334436535835, + 0.013751793652772903, + 0.024292010813951492 + ], + "credit_map": [ + 0.005046510137617588, + 0.0069405995309352875, + 0.00860814843326807, + 0.010921192355453968, + 0.0039653778076171875, + 0.004825976677238941, + 0.006012717727571726, + 0.005175815895199776, + 0.0020505308639258146, + 0.004190467298030853, + 0.003324799006804824, + 0.004002626985311508, + 0.00380828813649714, + 0.0028800014406442642, + 0.004649255890399218, + 0.003673027502372861, + 0.016774412244558334, + 0.012691093608736992, + 0.01157595869153738, + 0.017707036808133125, + 0.013231790624558926, + 0.020559459924697876, + 0.017759006470441818, + 0.013513876125216484, + 0.028487998992204666, + 0.12929797172546387, + 0.18612311780452728, + 0.032281484454870224, + 0.019864803180098534, + 0.012910207733511925, + 0.02696576714515686, + 0.014502494595944881, + 0.005178562365472317, + 0.008751139044761658, + 0.007940396666526794, + 0.006853782571852207, + 0.009911486878991127, + 0.005819506477564573, + 0.0051909214816987514, + 0.006110072135925293, + 0.004445652477443218, + 0.009294592775404453, + 0.006920783314853907, + 0.008188405074179173, + 0.005578755401074886, + 0.005062291398644447, + 0.007261831779032946, + 0.00643254816532135, + 0.014822529628872871, + 0.017358576878905296, + 0.021802276372909546, + 0.029679207131266594, + 0.011710547842085361, + 0.013331365771591663, + 0.01698058284819126, + 0.01729523204267025, + 0.008400466293096542, + 0.012454576790332794, + 0.008238495327532291, + 0.0058299177326262, + 0.005021116230636835, + 0.004003819078207016, + 0.005422496236860752, + 0.006342456676065922, + 0.001104741240851581, + 0.0015490135410800576, + 0.0016598901711404324, + 0.0017729238606989384, + 0.002480026101693511, + 0.001633276930078864, + 0.0018536824500188231, + 0.0016982393572106957, + 0.0024257234763354063, + 0.002747846767306328, + 0.003140875371173024, + 0.004625421948730946, + 0.003705873154103756, + 0.0025711648631840944, + 0.0021715916227549314, + 0.0029035136103630066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.07904396951198578 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023299720138311386 + ], + "opposite_logprob_tokens": [ + -3.7732996940612793 + ], + "logprob_reward": -0.023299720138311386, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023299720138311386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061256375163793564, + 0.031055571511387825, + 0.15735206007957458, + 0.07682309299707413, + 0.05884939804673195, + 0.039085984230041504, + 0.04156070575118065, + 0.41923651099205017, + 0.0470765084028244, + 0.06770384311676025 + ], + "credit_map": [ + 0.005633211694657803, + 0.0061206878162920475, + 0.00891081802546978, + 0.011190878227353096, + 0.004824298899620771, + 0.009518238715827465, + 0.0070199607871472836, + 0.008038273081183434, + 0.0025741667486727238, + 0.007077282294631004, + 0.0033608772791922092, + 0.004713291302323341, + 0.003617499954998493, + 0.0031989170238375664, + 0.0037913403939455748, + 0.0027221955824643373, + 0.018566947430372238, + 0.014272915199398994, + 0.017176080495119095, + 0.037550315260887146, + 0.012093443423509598, + 0.016142435371875763, + 0.016737351194024086, + 0.02481256052851677, + 0.00819843728095293, + 0.01884598657488823, + 0.023379826918244362, + 0.00910838134586811, + 0.0037501719780266285, + 0.0036384076811373234, + 0.006113532930612564, + 0.0037883399054408073, + 0.007095394656062126, + 0.008654308505356312, + 0.00895019806921482, + 0.009401061572134495, + 0.00896481517702341, + 0.0071517243050038815, + 0.004092726856470108, + 0.004539167508482933, + 0.00553878303617239, + 0.004680914804339409, + 0.005573510657995939, + 0.004036200698465109, + 0.004926147870719433, + 0.005399814806878567, + 0.005314106121659279, + 0.0036165034398436546, + 0.004012336488813162, + 0.0059089441783726215, + 0.005176130216568708, + 0.005992820020765066, + 0.004167948849499226, + 0.005596478935331106, + 0.0060190861113369465, + 0.0046869614161551, + 0.05081569775938988, + 0.07050487399101257, + 0.058537986129522324, + 0.05386317893862724, + 0.04044414311647415, + 0.03439581021666527, + 0.03920254483819008, + 0.07147224247455597, + 0.006725985556840897, + 0.008395458571612835, + 0.006301706191152334, + 0.0050130002200603485, + 0.006163217127323151, + 0.003195241093635559, + 0.005777146201580763, + 0.0055047497153282166, + 0.007412772625684738, + 0.007642838638275862, + 0.009245388209819794, + 0.01085286121815443, + 0.01056712307035923, + 0.006983959581702948, + 0.0069922273978590965, + 0.008006667718291283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.023299720138311386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020606713369488716 + ], + "opposite_logprob_tokens": [ + -3.895606756210327 + ], + "logprob_reward": -0.020606713369488716, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020606713369488716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03790883347392082, + 0.07005107402801514, + 0.04294528067111969, + 0.09102804213762283, + 0.14960935711860657, + 0.0400470569729805, + 0.08488107472658157, + 0.283616304397583, + 0.05320121720433235, + 0.1467117816209793 + ], + "credit_map": [ + 0.003634461434558034, + 0.0044032372534275055, + 0.0042776125483214855, + 0.007131871767342091, + 0.0025571887381374836, + 0.005602100398391485, + 0.004472764208912849, + 0.0058295936323702335, + 0.006790938321501017, + 0.015535683371126652, + 0.009107737801969051, + 0.016224011778831482, + 0.0049926843494176865, + 0.00450426759198308, + 0.007812126073986292, + 0.005083617754280567, + 0.004564026370644569, + 0.0042677889578044415, + 0.004266107454895973, + 0.007607440929859877, + 0.0039987582713365555, + 0.006166920531541109, + 0.007015211973339319, + 0.005059019662439823, + 0.011883318424224854, + 0.015323418192565441, + 0.022120457142591476, + 0.013127033598721027, + 0.007222200743854046, + 0.008339683525264263, + 0.007897146046161652, + 0.005114774685353041, + 0.01449553482234478, + 0.022415725514292717, + 0.02092941291630268, + 0.02111002430319786, + 0.029837269335985184, + 0.019393524155020714, + 0.012609089724719524, + 0.008818773552775383, + 0.003239393699914217, + 0.005057628732174635, + 0.005478446371853352, + 0.007367212325334549, + 0.004341773223131895, + 0.004815822467207909, + 0.0050862510688602924, + 0.0046605258248746395, + 0.007903038524091244, + 0.009818855673074722, + 0.010977326892316341, + 0.01521970797330141, + 0.011613539420068264, + 0.006400017067790031, + 0.014474482275545597, + 0.008474104106426239, + 0.03051356039941311, + 0.04714614897966385, + 0.0375785268843174, + 0.03975449129939079, + 0.03336944803595543, + 0.02774570696055889, + 0.031068768352270126, + 0.03643964231014252, + 0.007090938743203878, + 0.009679782204329967, + 0.007855718955397606, + 0.005799198057502508, + 0.006211425177752972, + 0.0059155928902328014, + 0.005354426335543394, + 0.0052941287867724895, + 0.015944037586450577, + 0.017622310668230057, + 0.02074626088142395, + 0.026681190356612206, + 0.021207042038440704, + 0.015873286873102188, + 0.0127261346206069, + 0.015911510214209557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.020606713369488716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328667603433132 + ], + "opposite_logprob_tokens": [ + -3.7732865810394287 + ], + "logprob_reward": -0.02328667603433132, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328667603433132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05233257636427879, + 0.02669638954102993, + 0.040710628032684326, + 0.25100940465927124, + 0.036203645169734955, + 0.08783455193042755, + 0.05278278887271881, + 0.06661960482597351, + 0.03496590629220009, + 0.35084450244903564 + ], + "credit_map": [ + 0.004664387088268995, + 0.0056723421439528465, + 0.006639195140451193, + 0.009482275694608688, + 0.004618596751242876, + 0.005933509673923254, + 0.007590379565954208, + 0.007731890305876732, + 0.0031100893393158913, + 0.0059237186796963215, + 0.0035889374557882547, + 0.00579665694385767, + 0.002028117189183831, + 0.002065735636278987, + 0.0023895022459328175, + 0.0017936363583430648, + 0.005170947872102261, + 0.004165403079241514, + 0.004887797404080629, + 0.0071839396841824055, + 0.0035172232892364264, + 0.005727722775191069, + 0.004547587130218744, + 0.005510011687874794, + 0.01788397505879402, + 0.06920921057462692, + 0.09871792793273926, + 0.020418576896190643, + 0.011264016851782799, + 0.007902317680418491, + 0.01646142080426216, + 0.009151985868811607, + 0.005723562091588974, + 0.004727648571133614, + 0.004081937484443188, + 0.004447817802429199, + 0.0041304477490484715, + 0.004552432801574469, + 0.002777669345960021, + 0.005762134678661823, + 0.009140459820628166, + 0.02127002365887165, + 0.012136765755712986, + 0.00983859971165657, + 0.007086328696459532, + 0.008038492873311043, + 0.011042593978345394, + 0.009281291626393795, + 0.004823107738047838, + 0.006698913406580687, + 0.006805790588259697, + 0.00904475525021553, + 0.008563187904655933, + 0.0037389665376394987, + 0.00801442377269268, + 0.00509364390745759, + 0.006837914232164621, + 0.009260027669370174, + 0.00872561614960432, + 0.010170995257794857, + 0.007070409599691629, + 0.007494915276765823, + 0.007584245875477791, + 0.009475493803620338, + 0.004157147835940123, + 0.005262666381895542, + 0.004533377476036549, + 0.00456216000020504, + 0.004695267416536808, + 0.0036691436544060707, + 0.00475115142762661, + 0.0033349955920130014, + 0.038331203162670135, + 0.04018642380833626, + 0.04849936068058014, + 0.060758672654628754, + 0.05467604845762253, + 0.03770597651600838, + 0.032253388315439224, + 0.03843345865607262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.02328667603433132 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with the precision of ultra-high-speed photography, this extreme close-up utilizes a static camera to depict the climactic moment of impact. A projectile, fired from an ornate M16 rifle, its gleaming, intricately engraved gold receiver partially visible at the edge of the frame, strikes a dark, cold-rolled steel surface. The camera employs a razor-sharp depth of field, maintaining impeccable, sharp focus solely on the bullet. As it makes contact, the entire event unfolds in breathtaking slow motion; the copper tip curls instantly, violently impacting the solid steel. This catastrophic collision transforms kinetic energy into a brilliant, explosive shower of incandescent sparks, blasting outward in a fiery starburst pattern, their white-hot trails starkly illuminating the scene against a softly blurred, dark, and indistinct background. The now-damaged projectile, spinning to a halt and tumbling away from its point of impact, its twisted metal surface captures dramatic key lighting as it gracefully ricochets through the air, creating a mesmerizing and powerful atmosphere of controlled destruction.", + "seed": 1422, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_12/index_8050_seed_1422_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09858967512845992, + "vlm_logprob": -0.09858967512845992, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.09858967512845992, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32722941040992737 + ], + "opposite_logprob_tokens": [ + -1.327229380607605 + ], + "logprob_reward": -0.32722941040992737, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32722941040992737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1487276256084442, + 0.03698694333434105, + 0.0836746022105217, + 0.15496864914894104, + 0.1316317617893219, + 0.09465554356575012, + 0.17053422331809998, + 0.020303189754486084, + 0.07293953746557236, + 0.08557792007923126 + ], + "credit_map": [ + 0.011807464063167572, + 0.010567273944616318, + 0.024519165977835655, + 0.025321735069155693, + 0.015268146060407162, + 0.015136513859033585, + 0.025951586663722992, + 0.020155737176537514, + 0.0039012280758470297, + 0.004401421640068293, + 0.004656486678868532, + 0.005191833712160587, + 0.0055684116668999195, + 0.004900022409856319, + 0.004574712365865707, + 0.00379282608628273, + 0.005924338940531015, + 0.011170989833772182, + 0.012498520314693451, + 0.011497734114527702, + 0.010355209931731224, + 0.011731884442269802, + 0.010872085578739643, + 0.00962384045124054, + 0.01819056086242199, + 0.01889597438275814, + 0.03313202038407326, + 0.019551469013094902, + 0.010660477913916111, + 0.023330437019467354, + 0.014883594587445259, + 0.0163241159170866, + 0.010100167244672775, + 0.014542609453201294, + 0.031678568571805954, + 0.010995002463459969, + 0.008431805297732353, + 0.037814028561115265, + 0.008789814077317715, + 0.009279774501919746, + 0.007054593879729509, + 0.023345327004790306, + 0.020633069798350334, + 0.00884237140417099, + 0.008810577914118767, + 0.011427109129726887, + 0.007886107079684734, + 0.0066563826985657215, + 0.014668839983642101, + 0.06251537799835205, + 0.024413110688328743, + 0.009086696431040764, + 0.010370025411248207, + 0.034932732582092285, + 0.0071211038157343864, + 0.007426325231790543, + 0.0015247468836605549, + 0.004429664928466082, + 0.0023139789700508118, + 0.0023359148763120174, + 0.0025101378560066223, + 0.0035676315892487764, + 0.002053255680948496, + 0.0015678597846999764, + 0.00545399310067296, + 0.011158895678818226, + 0.01388342771679163, + 0.007300204131752253, + 0.006161182653158903, + 0.017659639939665794, + 0.005478894338011742, + 0.0058432999067008495, + 0.006803224328905344, + 0.01454436406493187, + 0.01830035075545311, + 0.008049767464399338, + 0.009537782520055771, + 0.01339025516062975, + 0.008589379489421844, + 0.006362795364111662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.32722941040992737 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008461654186249 + ], + "opposite_logprob_tokens": [ + -2.695084571838379 + ], + "logprob_reward": -0.07008461654186249, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008461654186249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17855040729045868, + 0.0563693605363369, + 0.05433952808380127, + 0.051922935992479324, + 0.187251478433609, + 0.048269037157297134, + 0.1064455658197403, + 0.1483016312122345, + 0.043985843658447266, + 0.12456422299146652 + ], + "credit_map": [ + 0.011934015899896622, + 0.01964731141924858, + 0.030284922569990158, + 0.030227795243263245, + 0.02329085022211075, + 0.022303996607661247, + 0.02332536317408085, + 0.017536159604787827, + 0.0057842060923576355, + 0.00626433314755559, + 0.0066421134397387505, + 0.009066893719136715, + 0.00401431554928422, + 0.0047773877158761024, + 0.01167740486562252, + 0.008142704144120216, + 0.003668047720566392, + 0.006582292728126049, + 0.008230898529291153, + 0.006554036866873503, + 0.005383257754147053, + 0.008112126030027866, + 0.006518194917589426, + 0.009290676563978195, + 0.004004945047199726, + 0.009069676510989666, + 0.006377462297677994, + 0.0072600278072059155, + 0.003730224911123514, + 0.007965766824781895, + 0.006317113526165485, + 0.007197719532996416, + 0.011034161783754826, + 0.02934710681438446, + 0.04326770082116127, + 0.011746739968657494, + 0.010463418439030647, + 0.06767011433839798, + 0.006077049300074577, + 0.007645187433809042, + 0.004138650372624397, + 0.006514190696179867, + 0.00884551927447319, + 0.005454117897897959, + 0.004997662268579006, + 0.008834750391542912, + 0.005815054755657911, + 0.0036690954584628344, + 0.009196399711072445, + 0.03891510143876076, + 0.01445127371698618, + 0.004873679019510746, + 0.0054556443355977535, + 0.022346053272485733, + 0.006141953635960817, + 0.0050654630176723, + 0.009379604831337929, + 0.04626546800136566, + 0.009875582531094551, + 0.011034940369427204, + 0.014944161288440228, + 0.03930489346385002, + 0.009249108843505383, + 0.0082478616386652, + 0.0036253381986171007, + 0.008755951188504696, + 0.0059082224033772945, + 0.0055984011851251125, + 0.004424469079822302, + 0.007106017787009478, + 0.004846467636525631, + 0.0037209750153124332, + 0.009584472514688969, + 0.010721824131906033, + 0.03602687641978264, + 0.010328677482903004, + 0.01205342635512352, + 0.03267736732959747, + 0.005487028509378433, + 0.007684546988457441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.07008461654186249 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867374151945114 + ], + "opposite_logprob_tokens": [ + -3.048673629760742 + ], + "logprob_reward": -0.04867374151945114, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867374151945114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05795057862997055, + 0.02818594127893448, + 0.10901452600955963, + 0.1263597458600998, + 0.4222944378852844, + 0.029532551765441895, + 0.08907492458820343, + 0.04754015803337097, + 0.022771349176764488, + 0.0672757476568222 + ], + "credit_map": [ + 0.004464101046323776, + 0.0065459515899419785, + 0.00928542297333479, + 0.011583325453102589, + 0.005444129928946495, + 0.006871349178254604, + 0.008296405896544456, + 0.00545989815145731, + 0.0031511434353888035, + 0.0030311807058751583, + 0.0031859860755503178, + 0.004877003375440836, + 0.0020104944705963135, + 0.0035301251336932182, + 0.004872661549597979, + 0.003527347929775715, + 0.008432033471763134, + 0.01416217815130949, + 0.020185282453894615, + 0.01474014949053526, + 0.01001038122922182, + 0.016927795484662056, + 0.009301464073359966, + 0.01525525376200676, + 0.00978807732462883, + 0.014650462195277214, + 0.01924316957592964, + 0.02110888436436653, + 0.008719278499484062, + 0.017541859298944473, + 0.013399946503341198, + 0.02190808393061161, + 0.02537047490477562, + 0.06724971532821655, + 0.09961424767971039, + 0.024427538737654686, + 0.022292764857411385, + 0.15442828834056854, + 0.01112839113920927, + 0.017783047631382942, + 0.002304814290255308, + 0.0050409287214279175, + 0.004563977941870689, + 0.003474737983196974, + 0.002277562627568841, + 0.005600041709840298, + 0.0032815304584801197, + 0.0029889605939388275, + 0.00752016669139266, + 0.02390260249376297, + 0.012204441241919994, + 0.011153488419950008, + 0.005652768071740866, + 0.0149161322042346, + 0.005291896406561136, + 0.008433436043560505, + 0.0027736597694456577, + 0.013180244714021683, + 0.0038580915424972773, + 0.0037588514387607574, + 0.005257903598248959, + 0.013404474593698978, + 0.0031322294380515814, + 0.0021747087594121695, + 0.001689703669399023, + 0.004881022963672876, + 0.003056278685107827, + 0.0024959163274616003, + 0.001719065592624247, + 0.005180317908525467, + 0.001970177749171853, + 0.0017788674449548125, + 0.004815756808966398, + 0.007471807301044464, + 0.019578641280531883, + 0.006126240827143192, + 0.007022905629128218, + 0.014807892963290215, + 0.003806740278378129, + 0.003645773744210601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.04867374151945114 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020612670108675957 + ], + "opposite_logprob_tokens": [ + -3.8956127166748047 + ], + "logprob_reward": -0.020612670108675957, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020612670108675957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039065830409526825, + 0.039928656071424484, + 0.09694737941026688, + 0.1821630448102951, + 0.1443958580493927, + 0.04640212282538414, + 0.1354088932275772, + 0.06620725989341736, + 0.12167664617300034, + 0.12780427932739258 + ], + "credit_map": [ + 0.0024226736277341843, + 0.0037616128101944923, + 0.005313112400472164, + 0.006180241238325834, + 0.004613909404724836, + 0.0039772517047822475, + 0.006562567315995693, + 0.006234461907297373, + 0.004801981616765261, + 0.003997378516942263, + 0.005005415063351393, + 0.006415440235286951, + 0.002968747401610017, + 0.004547556396573782, + 0.008302970789372921, + 0.0038891658186912537, + 0.007916918024420738, + 0.013075125403702259, + 0.01986902952194214, + 0.012280108407139778, + 0.008865071460604668, + 0.015513604506850243, + 0.007702693808823824, + 0.011724828742444515, + 0.020715195685625076, + 0.026753218844532967, + 0.03758429363369942, + 0.01884828694164753, + 0.013115418143570423, + 0.03619326651096344, + 0.016421224921941757, + 0.012532142922282219, + 0.008669635280966759, + 0.020884819328784943, + 0.033850084990262985, + 0.009494239464402199, + 0.008133267983794212, + 0.049506355077028275, + 0.006386147812008858, + 0.007471292279660702, + 0.0032929459121078253, + 0.010968875139951706, + 0.005877700634300709, + 0.004687352105975151, + 0.003929982427507639, + 0.010128257796168327, + 0.004374907352030277, + 0.0031420974992215633, + 0.011419084854424, + 0.04666973277926445, + 0.025088580325245857, + 0.007107838056981564, + 0.009732100181281567, + 0.022116832435131073, + 0.007252048701047897, + 0.006022681016474962, + 0.004319971427321434, + 0.01713312231004238, + 0.006477608811110258, + 0.004692752845585346, + 0.007612888235598803, + 0.018123053014278412, + 0.004271566867828369, + 0.0035762919578701258, + 0.006928897462785244, + 0.022144222632050514, + 0.023734740912914276, + 0.006087584421038628, + 0.008481498807668686, + 0.046647198498249054, + 0.0031398709397763014, + 0.004512637387961149, + 0.009161513298749924, + 0.009717817418277264, + 0.0383453443646431, + 0.012452461756765842, + 0.011450008489191532, + 0.03352148085832596, + 0.0054200137965381145, + 0.007735644467175007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.020612670108675957 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme close-up with a static position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a projectile fired from an ornate M16 rifle with a gleaming, intricately engraved gold receiver visible at the edge of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the projectile strike a dark, cold-rolled steel surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the focus razor-sharp solely on the bullet while the background is softly blurred and indistinct?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the copper tip to curl instantly, creating an explosive shower of incandescent sparks in a fiery starburst pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347937062382698 + ], + "opposite_logprob_tokens": [ + -3.6513478755950928 + ], + "logprob_reward": -0.026347937062382698, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347937062382698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09342536330223083, + 0.03876180201768875, + 0.15298080444335938, + 0.1109539344906807, + 0.09119682759046555, + 0.04385270178318024, + 0.11909430474042892, + 0.11044126749038696, + 0.12038595974445343, + 0.11890700459480286 + ], + "credit_map": [ + 0.007359382696449757, + 0.011053279973566532, + 0.016183821484446526, + 0.018898485228419304, + 0.008514928631484509, + 0.009464651346206665, + 0.0105681037530303, + 0.011382711119949818, + 0.003942364361137152, + 0.0040448629297316074, + 0.0044771479442715645, + 0.006932365708053112, + 0.004045550245791674, + 0.003699662396684289, + 0.006262530107051134, + 0.005357317626476288, + 0.010630524717271328, + 0.020253226161003113, + 0.024415597319602966, + 0.021918848156929016, + 0.021484483033418655, + 0.0168685894459486, + 0.018141696229577065, + 0.019267840310931206, + 0.012002477422356606, + 0.017177298665046692, + 0.017242485657334328, + 0.015415194444358349, + 0.007409222424030304, + 0.017465131357312202, + 0.009828298352658749, + 0.014413819648325443, + 0.005941323935985565, + 0.015086283907294273, + 0.019901970401406288, + 0.005699465982615948, + 0.005516829900443554, + 0.029891274869441986, + 0.004475449211895466, + 0.004684224724769592, + 0.0036205255892127752, + 0.006466887425631285, + 0.006466211285442114, + 0.006007907912135124, + 0.004732139408588409, + 0.008333002217113972, + 0.004754976369440556, + 0.0034710527397692204, + 0.00906720757484436, + 0.038010358810424805, + 0.026797974482178688, + 0.006903647445142269, + 0.009323003701865673, + 0.015560025349259377, + 0.0077942791394889355, + 0.005637817550450563, + 0.006875528022646904, + 0.03656300902366638, + 0.017491092905402184, + 0.007528586313128471, + 0.01029941439628601, + 0.021383680403232574, + 0.0057458593510091305, + 0.0045540956780314445, + 0.007433322258293629, + 0.020503854379057884, + 0.02329539693892002, + 0.006545877084136009, + 0.00865019392222166, + 0.044275276362895966, + 0.004123001825064421, + 0.005559038836508989, + 0.009557212702929974, + 0.01346332486718893, + 0.03150876984000206, + 0.010976900346577168, + 0.011522016488015652, + 0.023624293506145477, + 0.009675241075456142, + 0.008579247631132603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.026347937062382698 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_14/index_tensor([4209, 4209])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_14/index_tensor([4209, 4209])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6ff5f04cdba5b4d624fb10881cf8e075b3fe467c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_14/index_tensor([4209, 4209])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In realistic documentary style, the camera rotates slowly and deliberately from a low angle, maintaining tight focus on the side-by-side cockpit of a Kamov Ka-52 attack helicopter. The powerful aircraft flies sideways, strafing with incredible stability just meters above the arid, sun-scorched ground. Its military camouflage, a muted green and tan pattern on its metal fuselage, is coated in a thin layer of dust in the harsh midday sun. The distinctive coaxial propellers, a blur of counter-rotating motion, violently churn the air, kicking up a vast, turbulent cloud of brown dust and small debris that billows out and engulfs the ground below. As the helicopter glides sideways, a brief, bright burst of fire from its side-mounted cannons erupts, the impact kicking up a secondary dust plume in the distance. The camera's slow, arcing movement keeps the cockpit perfectly framed by the chaotic swirl of dust, capturing the raw mechanical force of the maneuver with the tension and visceral immediacy of a frontline shot.", + "seed": 9632, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_14/index_4209_seed_9632_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07820497816428543, + "vlm_logprob": -0.07820497816428543, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.07820497816428543, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1539744734764099 + ], + "opposite_logprob_tokens": [ + -2.2789745330810547 + ], + "logprob_reward": -0.1539744734764099, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1539744734764099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12844988703727722, + 0.09432418644428253, + 0.07731001824140549, + 0.04442129284143448, + 0.06056492403149605, + 0.09356774389743805, + 0.22229698300361633, + 0.1893264502286911, + 0.06512004882097244, + 0.024618517607450485 + ], + "credit_map": [ + 0.01139359176158905, + 0.013655556365847588, + 0.016502488404512405, + 0.01696990430355072, + 0.02374107763171196, + 0.02049228549003601, + 0.011182775720953941, + 0.014512205496430397, + 0.007054303307086229, + 0.007609323598444462, + 0.017906004562973976, + 0.013093938119709492, + 0.010824919678270817, + 0.011090830899775028, + 0.012290031649172306, + 0.014454835094511509, + 0.007707298267632723, + 0.009732477366924286, + 0.009161091409623623, + 0.012846156023442745, + 0.008956276811659336, + 0.011601119302213192, + 0.010243413038551807, + 0.007062182761728764, + 0.0044824229553341866, + 0.006732562091201544, + 0.006156054325401783, + 0.007547540590167046, + 0.00514857517555356, + 0.005894818343222141, + 0.004426753614097834, + 0.004032565746456385, + 0.00574768241494894, + 0.007668815553188324, + 0.009069343097507954, + 0.010819641873240471, + 0.006311750505119562, + 0.004073684103786945, + 0.008621735498309135, + 0.00825226865708828, + 0.01117755938321352, + 0.010485502891242504, + 0.012496779672801495, + 0.012060652486979961, + 0.012140551581978798, + 0.007773484569042921, + 0.01957397721707821, + 0.007859227247536182, + 0.020930735394358635, + 0.018087025731801987, + 0.03573014959692955, + 0.03131162375211716, + 0.022178007289767265, + 0.01457233726978302, + 0.027776015922427177, + 0.051711101084947586, + 0.02771056815981865, + 0.04817303642630577, + 0.02488458715379238, + 0.022339999675750732, + 0.020431555807590485, + 0.01433718204498291, + 0.016166886314749718, + 0.015282627195119858, + 0.007066756021231413, + 0.005898956675082445, + 0.015618840232491493, + 0.009332550689578056, + 0.005541960243135691, + 0.004976550582796335, + 0.005304927006363869, + 0.011379508301615715, + 0.003256362397223711, + 0.002773683052510023, + 0.002904321765527129, + 0.0031865923665463924, + 0.003869175212457776, + 0.0027830738108605146, + 0.002858526771888137, + 0.0029867826960980892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.1539744734764099 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018276149407029152 + ], + "opposite_logprob_tokens": [ + -4.018276214599609 + ], + "logprob_reward": -0.018276149407029152, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018276149407029152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15696118772029877, + 0.10564932227134705, + 0.08168074488639832, + 0.056341785937547684, + 0.13953138887882233, + 0.11001512408256531, + 0.12617380917072296, + 0.1007235124707222, + 0.08897706121206284, + 0.03394605964422226 + ], + "credit_map": [ + 0.01181715726852417, + 0.016670458018779755, + 0.028399834409356117, + 0.010062345303595066, + 0.037444062530994415, + 0.021885139867663383, + 0.010958455502986908, + 0.019723715260624886, + 0.012240954674780369, + 0.013656945899128914, + 0.02007727138698101, + 0.01219528540968895, + 0.010042861104011536, + 0.013996534049510956, + 0.008730112574994564, + 0.014709356240928173, + 0.008950412273406982, + 0.006799344439059496, + 0.016825059428811073, + 0.011778258718550205, + 0.006382830440998077, + 0.009266825392842293, + 0.010204214602708817, + 0.011473801918327808, + 0.00518003199249506, + 0.008837353438138962, + 0.008268778212368488, + 0.010542504489421844, + 0.006295835133641958, + 0.005964313633739948, + 0.004945214372128248, + 0.006307756528258324, + 0.015239240601658821, + 0.024718981236219406, + 0.02282821014523506, + 0.013623859733343124, + 0.022595064714550972, + 0.010029916651546955, + 0.013644329272210598, + 0.016851792111992836, + 0.013505811803042889, + 0.010593957267701626, + 0.01970766857266426, + 0.015808284282684326, + 0.010686160996556282, + 0.007052750792354345, + 0.02072881907224655, + 0.01193167082965374, + 0.012380504049360752, + 0.01124507375061512, + 0.02277626097202301, + 0.01667827181518078, + 0.010451988317072392, + 0.008244147524237633, + 0.015093067660927773, + 0.02930450066924095, + 0.009913327172398567, + 0.009123671799898148, + 0.01616356335580349, + 0.013421358540654182, + 0.011320220306515694, + 0.00669466145336628, + 0.01364185567945242, + 0.020444849506020546, + 0.009999694302678108, + 0.00800422765314579, + 0.0124174440279603, + 0.0089200958609581, + 0.007283455692231655, + 0.006132157053798437, + 0.009745288640260696, + 0.026474695652723312, + 0.004085255321115255, + 0.004617873579263687, + 0.003830225206911564, + 0.0037635699845850468, + 0.005037057679146528, + 0.004808428231626749, + 0.0033260593190789223, + 0.004477592650800943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.018276149407029152 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012569139711558819 + ], + "opposite_logprob_tokens": [ + -4.387568950653076 + ], + "logprob_reward": -0.012569139711558819, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012569139711558819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1645788550376892, + 0.07254772633314133, + 0.06389215588569641, + 0.033987872302532196, + 0.05719631165266037, + 0.14186373353004456, + 0.12182515859603882, + 0.21180710196495056, + 0.0858607068657875, + 0.046440355479717255 + ], + "credit_map": [ + 0.015324225649237633, + 0.015733959153294563, + 0.017751766368746758, + 0.030810456722974777, + 0.021575791761279106, + 0.021069439128041267, + 0.019067293033003807, + 0.02324592135846615, + 0.00589319784194231, + 0.008486534468829632, + 0.015130712650716305, + 0.008604716509580612, + 0.00917497742921114, + 0.008192675188183784, + 0.007914443500339985, + 0.009150469675660133, + 0.0054527572356164455, + 0.005776746664196253, + 0.00858014915138483, + 0.016798976808786392, + 0.006791149266064167, + 0.00608520582318306, + 0.007419556844979525, + 0.0069876196794211864, + 0.0024738835636526346, + 0.003498170990496874, + 0.004552701022475958, + 0.0078171007335186, + 0.0033360382076352835, + 0.004875383339822292, + 0.003741328837350011, + 0.0036932637449353933, + 0.005883998703211546, + 0.008575577288866043, + 0.007516192737966776, + 0.01091744750738144, + 0.008652717806398869, + 0.004588519223034382, + 0.0060600535944104195, + 0.0050018043257296085, + 0.018005073070526123, + 0.01882905699312687, + 0.01921345852315426, + 0.01570487953722477, + 0.017117280513048172, + 0.009164422750473022, + 0.030828358605504036, + 0.01300120260566473, + 0.01171719841659069, + 0.012287726625800133, + 0.020403746515512466, + 0.022523950785398483, + 0.013459774665534496, + 0.007105537224560976, + 0.01605844497680664, + 0.018268780782818794, + 0.029224416241049767, + 0.04966389760375023, + 0.031181039288640022, + 0.02372216247022152, + 0.01958564855158329, + 0.014990485273301601, + 0.021842289716005325, + 0.02159714885056019, + 0.009987897239625454, + 0.00633171945810318, + 0.013902101665735245, + 0.009914147667586803, + 0.005894254893064499, + 0.005385041236877441, + 0.007641750853508711, + 0.026803800836205482, + 0.0058915335685014725, + 0.006467011757194996, + 0.0047845738008618355, + 0.006733851972967386, + 0.008236415684223175, + 0.006617430131882429, + 0.0037919487804174423, + 0.0039175874553620815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.012569139711558819 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645033273845911 + ], + "opposite_logprob_tokens": [ + -5.379644870758057 + ], + "logprob_reward": -0.004645033273845911, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645033273845911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28774377703666687, + 0.15115807950496674, + 0.0885959193110466, + 0.01926085166633129, + 0.05944129452109337, + 0.0850292295217514, + 0.10194143652915955, + 0.12303204834461212, + 0.05376827344298363, + 0.030029121786355972 + ], + "credit_map": [ + 0.02265632152557373, + 0.021271908655762672, + 0.05921639874577522, + 0.026839477941393852, + 0.06249217316508293, + 0.04206002876162529, + 0.023495646193623543, + 0.029711833223700523, + 0.011598890647292137, + 0.014906528405845165, + 0.03773099556565285, + 0.019305339083075523, + 0.014921581372618675, + 0.018457947298884392, + 0.01589956320822239, + 0.018337225541472435, + 0.010061601176857948, + 0.007341747637838125, + 0.013918239623308182, + 0.012088156305253506, + 0.010316314175724983, + 0.010286248289048672, + 0.011466395109891891, + 0.013117223046720028, + 0.0017503970302641392, + 0.0022748266346752644, + 0.002420647069811821, + 0.0023494239430874586, + 0.0024389999452978373, + 0.00414906395599246, + 0.002063170075416565, + 0.0018143221968784928, + 0.005487421527504921, + 0.006799422204494476, + 0.010174163617193699, + 0.012298991903662682, + 0.006341831758618355, + 0.00487693352624774, + 0.006606917828321457, + 0.006855614949017763, + 0.011806030757725239, + 0.011596770957112312, + 0.015645762905478477, + 0.006655196659266949, + 0.00900485459715128, + 0.008719533681869507, + 0.010589908808469772, + 0.011011171154677868, + 0.010469943284988403, + 0.009636875241994858, + 0.016110219061374664, + 0.01572607085108757, + 0.010608421638607979, + 0.00767688499763608, + 0.011813738383352757, + 0.019899288192391396, + 0.013312573544681072, + 0.014494274742901325, + 0.01986188068985939, + 0.014122994616627693, + 0.013931925408542156, + 0.009057645685970783, + 0.017575092613697052, + 0.020675666630268097, + 0.005679969675838947, + 0.005656916182488203, + 0.009591367095708847, + 0.008458191528916359, + 0.005375453270971775, + 0.003124834969639778, + 0.004920667968690395, + 0.010960872285068035, + 0.003501803847029805, + 0.0033491384238004684, + 0.0031717983074486256, + 0.006317406427115202, + 0.004921846091747284, + 0.003435146063566208, + 0.0032984197605401278, + 0.0020335621666163206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.004645033273845911 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015600949525833 + ], + "opposite_logprob_tokens": [ + -1.701560139656067 + ], + "logprob_reward": -0.2015600949525833, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015600949525833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12516234815120697, + 0.08319716900587082, + 0.09892647713422775, + 0.06287656724452972, + 0.06749765574932098, + 0.2079291194677353, + 0.09938281029462814, + 0.1410910189151764, + 0.060408636927604675, + 0.053528204560279846 + ], + "credit_map": [ + 0.010263008065521717, + 0.010732090100646019, + 0.022674359381198883, + 0.011741921305656433, + 0.028359999880194664, + 0.0183230172842741, + 0.010687315836548805, + 0.012380641885101795, + 0.00923141185194254, + 0.00909640546888113, + 0.014118813909590244, + 0.008157594129443169, + 0.009332808665931225, + 0.013892506249248981, + 0.0067223189398646355, + 0.012645313516259193, + 0.010668331757187843, + 0.010433758608996868, + 0.016816960647702217, + 0.012151088565587997, + 0.01074871327728033, + 0.009129319339990616, + 0.01148010790348053, + 0.017498206347227097, + 0.006345009431242943, + 0.009415012784302235, + 0.00944805983453989, + 0.008847125805914402, + 0.006022998597472906, + 0.010554524138569832, + 0.0049707056023180485, + 0.007273131515830755, + 0.0069166189059615135, + 0.009070876985788345, + 0.011170739307999611, + 0.013616068288683891, + 0.0068876068107783794, + 0.005043064244091511, + 0.008119173347949982, + 0.006673507858067751, + 0.02649660035967827, + 0.025856779888272285, + 0.03317289799451828, + 0.026364196091890335, + 0.021987400949001312, + 0.014287763275206089, + 0.03886191546916962, + 0.020901577547192574, + 0.01060541346669197, + 0.014357906766235828, + 0.01261376403272152, + 0.015015642158687115, + 0.010143011808395386, + 0.009278005920350552, + 0.012173107825219631, + 0.015195963904261589, + 0.016085606068372726, + 0.02335597760975361, + 0.017802758142352104, + 0.016911450773477554, + 0.01966296322643757, + 0.011101003736257553, + 0.01873140223324299, + 0.017439866438508034, + 0.007765678223222494, + 0.00852852314710617, + 0.012696354649960995, + 0.008186079561710358, + 0.005205947440117598, + 0.0047148470766842365, + 0.005712124519050121, + 0.007599083241075277, + 0.005772629752755165, + 0.005527832079678774, + 0.004629828035831451, + 0.011515543796122074, + 0.009061469696462154, + 0.007182335946708918, + 0.005340955685824156, + 0.00449761189520359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.2015600949525833 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In realistic documentary style, the camera rotates slowly and deliberately from a low angle, maintaining tight focus on the side-by-side cockpit of a Kamov Ka-52 attack helicopter. The powerful aircraft flies sideways, strafing with incredible stability just meters above the arid, sun-scorched ground. Its military camouflage, a muted green and tan pattern on its metal fuselage, is coated in a thin layer of dust in the harsh midday sun. The distinctive coaxial propellers, a blur of counter-rotating motion, violently churn the air, kicking up a vast, turbulent cloud of brown dust and small debris that billows out and engulfs the ground below. As the helicopter glides sideways, a brief, bright burst of fire from its side-mounted cannons erupts, the impact kicking up a secondary dust plume in the distance. The camera's slow, arcing movement keeps the cockpit perfectly framed by the chaotic swirl of dust, capturing the raw mechanical force of the maneuver with the tension and visceral immediacy of a frontline shot.", + "seed": 9633, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_14/index_4209_seed_9633_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08024462331086397, + "vlm_logprob": -0.08024462331086397, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.08024462331086397, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08225329965353012 + ], + "opposite_logprob_tokens": [ + -2.7072532176971436 + ], + "logprob_reward": -0.08225329965353012, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08225329965353012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1938454806804657, + 0.07900875061750412, + 0.11175096035003662, + 0.12193378806114197, + 0.07624632865190506, + 0.08515201508998871, + 0.11890535056591034, + 0.07985452562570572, + 0.04266154021024704, + 0.09064126014709473 + ], + "credit_map": [ + 0.02140062302350998, + 0.02929607778787613, + 0.035046037286520004, + 0.024431677535176277, + 0.034390516579151154, + 0.013837624341249466, + 0.02005098946392536, + 0.0153919393196702, + 0.01225220039486885, + 0.01162999402731657, + 0.015233651734888554, + 0.006676345132291317, + 0.00866810791194439, + 0.009898095391690731, + 0.008511225692927837, + 0.006139128934592009, + 0.017073610797524452, + 0.017223678529262543, + 0.010596624575555325, + 0.013914521783590317, + 0.016148466616868973, + 0.015792613849043846, + 0.010656372644007206, + 0.01034506969153881, + 0.014151724986732006, + 0.01031251810491085, + 0.009804803878068924, + 0.0118104862049222, + 0.02460881881415844, + 0.01759483851492405, + 0.02080170251429081, + 0.012848899699747562, + 0.013200435787439346, + 0.006544655188918114, + 0.009023847058415413, + 0.007381564471870661, + 0.00823111180216074, + 0.011527539230883121, + 0.006798987742513418, + 0.013538191094994545, + 0.01733376458287239, + 0.00722374627366662, + 0.012080655433237553, + 0.0054332236759364605, + 0.008818347007036209, + 0.017137181013822556, + 0.007151743862777948, + 0.009973357431590557, + 0.015738777816295624, + 0.021906835958361626, + 0.01209887582808733, + 0.015293275937438011, + 0.019031424075365067, + 0.014759416691958904, + 0.009583852253854275, + 0.010492888279259205, + 0.013025745749473572, + 0.008465304970741272, + 0.007021327503025532, + 0.012885529547929764, + 0.007744967471808195, + 0.01070050522685051, + 0.012891273014247417, + 0.007119874469935894, + 0.0070366328582167625, + 0.003913164138793945, + 0.0038682003505527973, + 0.0033749397844076157, + 0.007249135989695787, + 0.006617154460400343, + 0.004114813636988401, + 0.006487501785159111, + 0.01113967876881361, + 0.012808498926460743, + 0.010542885400354862, + 0.00918283686041832, + 0.011718039400875568, + 0.012710027396678925, + 0.01048673503100872, + 0.012052560225129128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.08225329965353012 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020628785714507103 + ], + "opposite_logprob_tokens": [ + -3.8956286907196045 + ], + "logprob_reward": -0.020628785714507103, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020628785714507103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07673373073339462, + 0.0718877837061882, + 0.06966155767440796, + 0.07558265328407288, + 0.14387834072113037, + 0.19696193933486938, + 0.13077720999717712, + 0.13049374520778656, + 0.05469459295272827, + 0.049328431487083435 + ], + "credit_map": [ + 0.008375330828130245, + 0.01074623316526413, + 0.01022880245000124, + 0.009795968420803547, + 0.014229055494070053, + 0.008033396676182747, + 0.008009836077690125, + 0.0073151132091879845, + 0.010932224802672863, + 0.010496011935174465, + 0.009566607885062695, + 0.0075680771842598915, + 0.006660148035734892, + 0.008362183347344398, + 0.009437989443540573, + 0.00886454340070486, + 0.00820315070450306, + 0.010011741891503334, + 0.006704418919980526, + 0.009087246842682362, + 0.009028066881000996, + 0.008535494096577168, + 0.010643121786415577, + 0.007448316086083651, + 0.011915144510567188, + 0.007873594760894775, + 0.0061072250828146935, + 0.005148037802428007, + 0.013608665205538273, + 0.012216325849294662, + 0.009042207151651382, + 0.009671458043158054, + 0.029987722635269165, + 0.011371786706149578, + 0.01615593396127224, + 0.011303669773042202, + 0.016082029789686203, + 0.02417144551873207, + 0.011642259545624256, + 0.02316349744796753, + 0.04873928055167198, + 0.016616173088550568, + 0.017338352277874947, + 0.014005553908646107, + 0.020410623401403427, + 0.048977091908454895, + 0.013618013821542263, + 0.0172568466514349, + 0.018299313262104988, + 0.018420273438096046, + 0.013612690381705761, + 0.021026935428380966, + 0.01701480709016323, + 0.016989050433039665, + 0.013616164214909077, + 0.011797974817454815, + 0.017218146473169327, + 0.009481224231421947, + 0.01325041800737381, + 0.008265319280326366, + 0.010431686416268349, + 0.019761623814702034, + 0.020392099395394325, + 0.031693220138549805, + 0.005603819154202938, + 0.004731677006930113, + 0.007035448681563139, + 0.006425326224416494, + 0.007416461128741503, + 0.007108449470251799, + 0.008811259642243385, + 0.0075621516443789005, + 0.006027047522366047, + 0.006907441187649965, + 0.006898127030581236, + 0.004085356369614601, + 0.005962412804365158, + 0.006271573714911938, + 0.008028400130569935, + 0.005148070864379406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.020628785714507103 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086884886026382 + ], + "opposite_logprob_tokens": [ + -4.511086940765381 + ], + "logprob_reward": -0.011086884886026382, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086884886026382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10279574245214462, + 0.11723342537879944, + 0.0905710756778717, + 0.13431087136268616, + 0.04483386129140854, + 0.15948016941547394, + 0.08256680518388748, + 0.16614526510238647, + 0.03061584383249283, + 0.0714469626545906 + ], + "credit_map": [ + 0.01547038834542036, + 0.01575174182653427, + 0.014425846748054028, + 0.012186994776129723, + 0.016770794987678528, + 0.009705386124551296, + 0.009268749505281448, + 0.009215843863785267, + 0.015699615702033043, + 0.01575644686818123, + 0.013629162684082985, + 0.010113447904586792, + 0.014973567798733711, + 0.012005672790110111, + 0.013901659287512302, + 0.021153846755623817, + 0.009095856919884682, + 0.010557596571743488, + 0.008741388097405434, + 0.014151456765830517, + 0.015345540829002857, + 0.011773376725614071, + 0.01287993136793375, + 0.00802592746913433, + 0.013467107899487019, + 0.00875126477330923, + 0.013406360521912575, + 0.018290838226675987, + 0.022422239184379578, + 0.016979221254587173, + 0.025971457362174988, + 0.015022373758256435, + 0.0059400093741714954, + 0.005005303304642439, + 0.004555388353765011, + 0.004661187529563904, + 0.008834346197545528, + 0.006629576440900564, + 0.0036233675200492144, + 0.0055846842005848885, + 0.03844184800982475, + 0.016292709857225418, + 0.016106978058815002, + 0.01163558755069971, + 0.017319662496447563, + 0.036275580525398254, + 0.0097025316208601, + 0.013705267570912838, + 0.01151997409760952, + 0.011369100771844387, + 0.007312309462577105, + 0.01006900705397129, + 0.016583655029535294, + 0.00910442229360342, + 0.0066591608338057995, + 0.00994917657226324, + 0.01865316741168499, + 0.013375414535403252, + 0.014084254391491413, + 0.03322724625468254, + 0.01572476141154766, + 0.022111253812909126, + 0.025921329855918884, + 0.023047838360071182, + 0.003344594966620207, + 0.0031344874296337366, + 0.002779204398393631, + 0.002858224557712674, + 0.004646894987672567, + 0.003968878649175167, + 0.005356074310839176, + 0.004527487326413393, + 0.008755072951316833, + 0.01005096361041069, + 0.005536937154829502, + 0.005333038978278637, + 0.010280496440827847, + 0.019938690587878227, + 0.005827220156788826, + 0.005724536720663309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.011086884886026382 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950710736215115 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950710736215115, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950710736215115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10745912045240402, + 0.05714451149106026, + 0.052448514848947525, + 0.08723288774490356, + 0.07420209050178528, + 0.36544206738471985, + 0.09197920560836792, + 0.043966274708509445, + 0.04740925878286362, + 0.07271604239940643 + ], + "credit_map": [ + 0.014654803089797497, + 0.013964383862912655, + 0.0188017375767231, + 0.012202898971736431, + 0.016031885519623756, + 0.011659213341772556, + 0.010453355498611927, + 0.009690852835774422, + 0.008967208676040173, + 0.010940286330878735, + 0.008473561145365238, + 0.006064120680093765, + 0.005297101568430662, + 0.006465813145041466, + 0.005587512161582708, + 0.005348916165530682, + 0.0067709656432271, + 0.0068391235545277596, + 0.004864975810050964, + 0.006977804470807314, + 0.006549632176756859, + 0.006353428121656179, + 0.007827400229871273, + 0.006265189964324236, + 0.011625285260379314, + 0.00927656702697277, + 0.007656577974557877, + 0.007115596905350685, + 0.018182123079895973, + 0.01375612709671259, + 0.009829835966229439, + 0.009790783748030663, + 0.009217537939548492, + 0.010362339206039906, + 0.00555666908621788, + 0.011309986934065819, + 0.014997493475675583, + 0.01064907293766737, + 0.006802928633987904, + 0.005306072998791933, + 0.08726880699396133, + 0.040700048208236694, + 0.035729050636291504, + 0.02829260565340519, + 0.04164993762969971, + 0.08014894276857376, + 0.02159012481570244, + 0.03006259724497795, + 0.012996231205761433, + 0.00897467415779829, + 0.008445528335869312, + 0.008449086919426918, + 0.016034290194511414, + 0.014400235377252102, + 0.009575020521879196, + 0.013104152865707874, + 0.008582462556660175, + 0.004443291109055281, + 0.005271862726658583, + 0.004524936433881521, + 0.005795782431960106, + 0.0065214913338422775, + 0.005064438562840223, + 0.0037620177026838064, + 0.005745942238718271, + 0.005101913586258888, + 0.005501303821802139, + 0.0044662863947451115, + 0.005706479772925377, + 0.0068389349617064, + 0.006515581626445055, + 0.007532817777246237, + 0.007961330004036427, + 0.008366629481315613, + 0.008913133293390274, + 0.006674912292510271, + 0.007511231116950512, + 0.01811639778316021, + 0.01001669280230999, + 0.005155719816684723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.005950710736215115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28130343556404114 + ], + "opposite_logprob_tokens": [ + -1.4063034057617188 + ], + "logprob_reward": -0.28130343556404114, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28130343556404114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11386509239673615, + 0.0481652095913887, + 0.052426621317863464, + 0.045381706207990646, + 0.05513247847557068, + 0.23579171299934387, + 0.16918565332889557, + 0.08526194095611572, + 0.0384010411798954, + 0.1563885360956192 + ], + "credit_map": [ + 0.01652398705482483, + 0.014152239076793194, + 0.01655503921210766, + 0.016894889995455742, + 0.015491562895476818, + 0.009582191705703735, + 0.01365899108350277, + 0.011006178334355354, + 0.0064399088732898235, + 0.007372275926172733, + 0.009398628026247025, + 0.005245971027761698, + 0.004432846792042255, + 0.005260573234409094, + 0.0053660073317587376, + 0.004648997914046049, + 0.007080290000885725, + 0.006095374468713999, + 0.005781563930213451, + 0.008974388241767883, + 0.006179147399961948, + 0.0070449113845825195, + 0.004939856007695198, + 0.0063310847617685795, + 0.007046759128570557, + 0.004825789947062731, + 0.004064061678946018, + 0.006505976431071758, + 0.006407301407307386, + 0.005877393763512373, + 0.005501083098351955, + 0.005153336096554995, + 0.011414505541324615, + 0.005630597472190857, + 0.0054131909273564816, + 0.007157075684517622, + 0.00720610935240984, + 0.008556957356631756, + 0.003411887213587761, + 0.0063421521335840225, + 0.05228254199028015, + 0.024505378678441048, + 0.025142842903733253, + 0.01728578843176365, + 0.028384611010551453, + 0.04932938143610954, + 0.015302619896829128, + 0.023558542132377625, + 0.02553826943039894, + 0.022516775876283646, + 0.01717124879360199, + 0.020070495083928108, + 0.030803894624114037, + 0.022498495876789093, + 0.015263106673955917, + 0.015323368832468987, + 0.01095986645668745, + 0.010366170667111874, + 0.01012040488421917, + 0.011604064144194126, + 0.009414409287273884, + 0.012175413779914379, + 0.012225058861076832, + 0.008396554738283157, + 0.005505767650902271, + 0.003486445639282465, + 0.005342422053217888, + 0.0029903349932283163, + 0.004714381881058216, + 0.005550570785999298, + 0.005686716642230749, + 0.005124399438500404, + 0.02021891623735428, + 0.010246671736240387, + 0.012191368266940117, + 0.011378311552107334, + 0.010706642642617226, + 0.05711502581834793, + 0.018587389960885048, + 0.01594420149922371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.28130343556404114 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In realistic documentary style, the camera rotates slowly and deliberately from a low angle, maintaining tight focus on the side-by-side cockpit of a Kamov Ka-52 attack helicopter. The powerful aircraft flies sideways, strafing with incredible stability just meters above the arid, sun-scorched ground. Its military camouflage, a muted green and tan pattern on its metal fuselage, is coated in a thin layer of dust in the harsh midday sun. The distinctive coaxial propellers, a blur of counter-rotating motion, violently churn the air, kicking up a vast, turbulent cloud of brown dust and small debris that billows out and engulfs the ground below. As the helicopter glides sideways, a brief, bright burst of fire from its side-mounted cannons erupts, the impact kicking up a secondary dust plume in the distance. The camera's slow, arcing movement keeps the cockpit perfectly framed by the chaotic swirl of dust, capturing the raw mechanical force of the maneuver with the tension and visceral immediacy of a frontline shot.", + "seed": 9634, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_14/index_4209_seed_9634_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.047325033508241174, + "vlm_logprob": -0.047325033508241174, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.047325033508241174, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08091980963945389 + ], + "opposite_logprob_tokens": [ + -2.7059197425842285 + ], + "logprob_reward": -0.08091980963945389, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08091980963945389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19641514122486115, + 0.20660027861595154, + 0.09864963591098785, + 0.03537774458527565, + 0.0842687115073204, + 0.0864541307091713, + 0.08528139442205429, + 0.06085512414574623, + 0.07648396492004395, + 0.06961385160684586 + ], + "credit_map": [ + 0.02593238279223442, + 0.031611911952495575, + 0.01684647984802723, + 0.020104438066482544, + 0.026492420583963394, + 0.030925599858164787, + 0.019879039376974106, + 0.02462286688387394, + 0.026297030970454216, + 0.03781124949455261, + 0.016660690307617188, + 0.019384225830435753, + 0.045402225106954575, + 0.030438683927059174, + 0.017020396888256073, + 0.013585777953267097, + 0.013962841592729092, + 0.01161983609199524, + 0.012229620479047298, + 0.00741971842944622, + 0.020638400688767433, + 0.016088446602225304, + 0.00656398618593812, + 0.010126793757081032, + 0.00405400525778532, + 0.004040549509227276, + 0.004311629105359316, + 0.0034426546189934015, + 0.005612518172711134, + 0.005440967157483101, + 0.0035579935647547245, + 0.004917429760098457, + 0.007390272803604603, + 0.013196012936532497, + 0.00975747499614954, + 0.00946817360818386, + 0.01377657800912857, + 0.012826019898056984, + 0.009711641818284988, + 0.00814253930002451, + 0.012576314620673656, + 0.009438208304345608, + 0.019721610471606255, + 0.009962048381567001, + 0.007425917778164148, + 0.009852620773017406, + 0.008584548719227314, + 0.008892865851521492, + 0.019760612398386, + 0.012939270585775375, + 0.006998521275818348, + 0.00796441175043583, + 0.00854815449565649, + 0.0126051539555192, + 0.008715488947927952, + 0.00774978194385767, + 0.010808216407895088, + 0.009250830858945847, + 0.009049874730408192, + 0.006580092012882233, + 0.006495407782495022, + 0.007098858710378408, + 0.006194645073264837, + 0.005377194844186306, + 0.013686994090676308, + 0.009926828555762768, + 0.009005296044051647, + 0.010974960401654243, + 0.0043602194637060165, + 0.008632020093500614, + 0.010028754360973835, + 0.00986888725310564, + 0.011652817018330097, + 0.009594055823981762, + 0.009206417016685009, + 0.006998531520366669, + 0.006046191323548555, + 0.008984212763607502, + 0.008686387911438942, + 0.008445237763226032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.08091980963945389 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014239019714295864 + ], + "opposite_logprob_tokens": [ + -4.2642388343811035 + ], + "logprob_reward": -0.014239019714295864, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014239019714295864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08226486295461655, + 0.07060270011425018, + 0.0409059077501297, + 0.05693071335554123, + 0.03586249053478241, + 0.11455417424440384, + 0.23696190118789673, + 0.25831857323646545, + 0.04671940952539444, + 0.056879304349422455 + ], + "credit_map": [ + 0.012720576487481594, + 0.014990967698395252, + 0.010302798822522163, + 0.009556343778967857, + 0.008342479355633259, + 0.011194723658263683, + 0.0076708062551915646, + 0.007486171089112759, + 0.00886722281575203, + 0.010472584515810013, + 0.010041012428700924, + 0.0058800335973501205, + 0.011954599991440773, + 0.010677355341613293, + 0.005000232253223658, + 0.007709651254117489, + 0.005164561327546835, + 0.004379266873002052, + 0.005694716237485409, + 0.0035820230841636658, + 0.009280570782721043, + 0.005807782988995314, + 0.0040160431526601315, + 0.002980937948450446, + 0.005340352188795805, + 0.005020176991820335, + 0.005489531438797712, + 0.00378809729591012, + 0.01090174075216055, + 0.011631694622337818, + 0.005672670900821686, + 0.009086450561881065, + 0.0030369774904102087, + 0.00547064607962966, + 0.0034345327876508236, + 0.0047661722637712955, + 0.00713591231033206, + 0.005820747464895248, + 0.0024172638077288866, + 0.003780236467719078, + 0.012718910351395607, + 0.009907132014632225, + 0.036038562655448914, + 0.013541404157876968, + 0.008442439138889313, + 0.01372245792299509, + 0.012277347035706043, + 0.007905920036137104, + 0.06498018652200699, + 0.056642305105924606, + 0.019138002768158913, + 0.013793342746794224, + 0.014097136445343494, + 0.043669067323207855, + 0.012053243815898895, + 0.01258862018585205, + 0.08557618409395218, + 0.040425579994916916, + 0.04377313703298569, + 0.016691414639353752, + 0.015571405179798603, + 0.021515732631087303, + 0.01705273427069187, + 0.017712391912937164, + 0.007971937768161297, + 0.005354918539524078, + 0.005546238273382187, + 0.005647525656968355, + 0.003540949895977974, + 0.007613813038915396, + 0.0042844051495194435, + 0.00675961934030056, + 0.013427511788904667, + 0.008477795869112015, + 0.007945562712848186, + 0.005033691879361868, + 0.004073305055499077, + 0.0050095245242118835, + 0.004350998438894749, + 0.008560914546251297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.014239019714295864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794510900974274 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.009794510900974274, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794510900974274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22239059209823608, + 0.1878880262374878, + 0.156652569770813, + 0.03422348201274872, + 0.05949940159916878, + 0.09459348767995834, + 0.07417052239179611, + 0.10483824461698532, + 0.024261726066470146, + 0.04148189350962639 + ], + "credit_map": [ + 0.031982000917196274, + 0.03010447509586811, + 0.026125315576791763, + 0.023338299244642258, + 0.025953801348805428, + 0.040113646537065506, + 0.019528130069375038, + 0.02524491772055626, + 0.02172684110701084, + 0.028545869514346123, + 0.029738930985331535, + 0.021313253790140152, + 0.03308850899338722, + 0.02110014110803604, + 0.015282483771443367, + 0.017092006281018257, + 0.02480250783264637, + 0.015800053253769875, + 0.02270554006099701, + 0.012073798105120659, + 0.025462601333856583, + 0.023875029757618904, + 0.017027458176016808, + 0.014905599877238274, + 0.003169519826769829, + 0.004457531962543726, + 0.0038782604970037937, + 0.0027985521592199802, + 0.004657173529267311, + 0.005087953060865402, + 0.004161391407251358, + 0.00601310096681118, + 0.005024048034101725, + 0.010951702482998371, + 0.005715448409318924, + 0.007675961125642061, + 0.008054501377046108, + 0.007132577244192362, + 0.006343111861497164, + 0.008602052927017212, + 0.013305418193340302, + 0.008898554369807243, + 0.02472795359790325, + 0.009257519617676735, + 0.009126109071075916, + 0.013013310730457306, + 0.0070748478174209595, + 0.0091897863894701, + 0.016512690111994743, + 0.010919246822595596, + 0.007369052618741989, + 0.005530588328838348, + 0.008635657839477062, + 0.011702435091137886, + 0.0054101599380373955, + 0.008090694434940815, + 0.02923573926091194, + 0.014402143657207489, + 0.016965096816420555, + 0.008993140421807766, + 0.007986859418451786, + 0.00828465260565281, + 0.00965296570211649, + 0.009317656978964806, + 0.0042534093372523785, + 0.003105610143393278, + 0.0027778856456279755, + 0.0030248246621340513, + 0.002253754995763302, + 0.0034364666789770126, + 0.0030879401601850986, + 0.002321834908798337, + 0.006197065114974976, + 0.004619525279849768, + 0.007316139992326498, + 0.005024287849664688, + 0.0038764432538300753, + 0.003950871527194977, + 0.004841187037527561, + 0.005656380672007799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.009794510900974274 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639100283384323 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639100283384323, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639100283384323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17562204599380493, + 0.1250573843717575, + 0.18060240149497986, + 0.03212957829236984, + 0.025595664978027344, + 0.06699652224779129, + 0.15067361295223236, + 0.14181232452392578, + 0.04757716879248619, + 0.05393333360552788 + ], + "credit_map": [ + 0.024699512869119644, + 0.03825850784778595, + 0.023793647065758705, + 0.02174646221101284, + 0.01685263030230999, + 0.021647261455655098, + 0.015332809649407864, + 0.0132911941036582, + 0.015339993871748447, + 0.021719565615057945, + 0.01101533230394125, + 0.00885308813303709, + 0.032245710492134094, + 0.01857590302824974, + 0.008642911911010742, + 0.008664862252771854, + 0.022388147190213203, + 0.020662862807512283, + 0.0129283107817173, + 0.015188785269856453, + 0.051246367394924164, + 0.032360728830099106, + 0.011308958753943443, + 0.01451822742819786, + 0.0032774326391518116, + 0.003587678773328662, + 0.003563998267054558, + 0.002542511560022831, + 0.004782528150826693, + 0.004977705888450146, + 0.004190365318208933, + 0.005207354202866554, + 0.002275473903864622, + 0.003983496688306332, + 0.00251819659024477, + 0.0036240722984075546, + 0.004551996476948261, + 0.003409573808312416, + 0.0021221747156232595, + 0.0031106777023524046, + 0.008979206904768944, + 0.007394595071673393, + 0.009350747801363468, + 0.006971322000026703, + 0.007960226386785507, + 0.007769717834889889, + 0.007484355475753546, + 0.011086340062320232, + 0.03332060948014259, + 0.02545883134007454, + 0.013168057426810265, + 0.012947051785886288, + 0.019001847133040428, + 0.020896214991807938, + 0.010736553929746151, + 0.015144454315304756, + 0.0466553159058094, + 0.022433757781982422, + 0.024700716137886047, + 0.008963149972259998, + 0.00909499917179346, + 0.012083575129508972, + 0.009257427416741848, + 0.008623368106782436, + 0.007902010343968868, + 0.00564195029437542, + 0.005516884382814169, + 0.006806273013353348, + 0.003449296345934272, + 0.006943072658032179, + 0.005592033267021179, + 0.0057256449945271015, + 0.008952049538493156, + 0.006893574260175228, + 0.006111628841608763, + 0.007424707058817148, + 0.004331185948103666, + 0.005111293867230415, + 0.005774135701358318, + 0.009334756061434746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.004639100283384323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703272700309753 + ], + "opposite_logprob_tokens": [ + -2.12703275680542 + ], + "logprob_reward": -0.12703272700309753, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703272700309753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15686939656734467, + 0.11291541159152985, + 0.07279307395219803, + 0.05061712488532066, + 0.052132219076156616, + 0.12241306900978088, + 0.12554974853992462, + 0.24673081934452057, + 0.018483154475688934, + 0.04149596020579338 + ], + "credit_map": [ + 0.02271840162575245, + 0.02536517195403576, + 0.01860048435628414, + 0.021873589605093002, + 0.018550269305706024, + 0.02105332724750042, + 0.01438236515969038, + 0.01432577520608902, + 0.014786997810006142, + 0.018281547352671623, + 0.020801816135644913, + 0.014243281446397305, + 0.014151151292026043, + 0.011075416579842567, + 0.009326823055744171, + 0.010248382575809956, + 0.013023420237004757, + 0.006666807923465967, + 0.008495268411934376, + 0.004441641736775637, + 0.016034157946705818, + 0.012329748831689358, + 0.005660282913595438, + 0.006141753401607275, + 0.0046193720772862434, + 0.004550958517938852, + 0.0038044406101107597, + 0.0033572574611753225, + 0.011739042587578297, + 0.010432328097522259, + 0.005133875645697117, + 0.006979856174439192, + 0.005051897838711739, + 0.010594655759632587, + 0.005248752888292074, + 0.0063146501779556274, + 0.007841444574296474, + 0.006291292142122984, + 0.00433890288695693, + 0.006450625602155924, + 0.0194470826536417, + 0.020033907145261765, + 0.00848215539008379, + 0.011662169359624386, + 0.015041538514196873, + 0.015743184834718704, + 0.01336432620882988, + 0.018638718873262405, + 0.026970069855451584, + 0.023536602035164833, + 0.014319055713713169, + 0.011466423980891705, + 0.010861215181648731, + 0.017258182168006897, + 0.012844879180192947, + 0.008293322287499905, + 0.08059059083461761, + 0.03729172423481941, + 0.041208695620298386, + 0.017478851601481438, + 0.015794793143868446, + 0.019564950838685036, + 0.018225615844130516, + 0.016575602814555168, + 0.0037076377775520086, + 0.002457793103531003, + 0.002604132052510977, + 0.0021054940298199654, + 0.000906819652300328, + 0.0019893578719347715, + 0.002186169847846031, + 0.002525751944631338, + 0.005461443215608597, + 0.004156873095780611, + 0.006052838172763586, + 0.005879770033061504, + 0.003962562885135412, + 0.004906801972538233, + 0.005724871531128883, + 0.005350799765437841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.12703272700309753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In realistic documentary style, the camera rotates slowly and deliberately from a low angle, maintaining tight focus on the side-by-side cockpit of a Kamov Ka-52 attack helicopter. The powerful aircraft flies sideways, strafing with incredible stability just meters above the arid, sun-scorched ground. Its military camouflage, a muted green and tan pattern on its metal fuselage, is coated in a thin layer of dust in the harsh midday sun. The distinctive coaxial propellers, a blur of counter-rotating motion, violently churn the air, kicking up a vast, turbulent cloud of brown dust and small debris that billows out and engulfs the ground below. As the helicopter glides sideways, a brief, bright burst of fire from its side-mounted cannons erupts, the impact kicking up a secondary dust plume in the distance. The camera's slow, arcing movement keeps the cockpit perfectly framed by the chaotic swirl of dust, capturing the raw mechanical force of the maneuver with the tension and visceral immediacy of a frontline shot.", + "seed": 9635, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_14/index_4209_seed_9635_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0569466033950448, + "vlm_logprob": -0.0569466033950448, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.0569466033950448, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07432340830564499 + ], + "opposite_logprob_tokens": [ + -2.8243234157562256 + ], + "logprob_reward": -0.07432340830564499, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07432340830564499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22262150049209595, + 0.17295093834400177, + 0.0902320072054863, + 0.15610527992248535, + 0.07697898149490356, + 0.052708279341459274, + 0.09819451719522476, + 0.04520201310515404, + 0.05023226886987686, + 0.034774232655763626 + ], + "credit_map": [ + 0.02759127877652645, + 0.028354624286293983, + 0.03727438673377037, + 0.02249838225543499, + 0.019813813269138336, + 0.03069707378745079, + 0.025178823620080948, + 0.03121310845017433, + 0.025461671873927116, + 0.021409817039966583, + 0.04471421241760254, + 0.01253802701830864, + 0.012193424627184868, + 0.025844493880867958, + 0.010309035889804363, + 0.020480258390307426, + 0.009933874011039734, + 0.011657556518912315, + 0.01106718834489584, + 0.010213036090135574, + 0.009483253583312035, + 0.010744131170213223, + 0.010875873267650604, + 0.016257086768746376, + 0.02115422487258911, + 0.024544280022382736, + 0.031600870192050934, + 0.009989932179450989, + 0.021634094417095184, + 0.023133590817451477, + 0.009894859977066517, + 0.014153440482914448, + 0.013302839361131191, + 0.01101688388735056, + 0.009299228899180889, + 0.0077322316356003284, + 0.007660979405045509, + 0.010919110849499702, + 0.008227837271988392, + 0.008819867856800556, + 0.006315332371741533, + 0.005840736906975508, + 0.0051972889341413975, + 0.005186953581869602, + 0.006757709663361311, + 0.007061521522700787, + 0.010949579998850822, + 0.005399154499173164, + 0.01704319380223751, + 0.01346813328564167, + 0.013825837522745132, + 0.01284080557525158, + 0.007668096572160721, + 0.01816956140100956, + 0.007289455272257328, + 0.007889428175985813, + 0.006646144203841686, + 0.008232458494603634, + 0.006211128085851669, + 0.007284503895789385, + 0.004852840211242437, + 0.0037096550222486258, + 0.004899680148810148, + 0.0033656007144600153, + 0.008560740388929844, + 0.005073943175375462, + 0.009949387051165104, + 0.0056860195472836494, + 0.006849369965493679, + 0.004434253089129925, + 0.0053817154839634895, + 0.004296839237213135, + 0.0036578993313014507, + 0.00430641183629632, + 0.005378542002290487, + 0.004505978897213936, + 0.003953211475163698, + 0.004724267404526472, + 0.0049291555769741535, + 0.00331876496784389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.07432340830564499 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016109894961118698 + ], + "opposite_logprob_tokens": [ + -4.141109943389893 + ], + "logprob_reward": -0.016109894961118698, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016109894961118698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24669091403484344, + 0.15253837406635284, + 0.11514108628034592, + 0.047782253473997116, + 0.051917072385549545, + 0.05240962654352188, + 0.16497108340263367, + 0.03067793883383274, + 0.09675118327140808, + 0.04112052172422409 + ], + "credit_map": [ + 0.030475685372948647, + 0.034041665494441986, + 0.04319636523723602, + 0.02157742902636528, + 0.018402790650725365, + 0.0465412512421608, + 0.02462630718946457, + 0.027829408645629883, + 0.01672673225402832, + 0.019152101129293442, + 0.021805549040436745, + 0.012907830998301506, + 0.01673778146505356, + 0.0374365895986557, + 0.013752670958638191, + 0.014019109308719635, + 0.016856104135513306, + 0.01856808550655842, + 0.024534769356250763, + 0.009096063673496246, + 0.011868099682033062, + 0.015586509369313717, + 0.007913714274764061, + 0.010717734694480896, + 0.007852005772292614, + 0.008097481913864613, + 0.012342964299023151, + 0.004051533527672291, + 0.0028800766449421644, + 0.004772715270519257, + 0.00355894910171628, + 0.004226521123200655, + 0.009595231153070927, + 0.006348070222884417, + 0.005365594755858183, + 0.006036668084561825, + 0.00729857524856925, + 0.005802033469080925, + 0.006220509298145771, + 0.005250385031104088, + 0.006951836869120598, + 0.00512914452701807, + 0.005373668856918812, + 0.006825026590377092, + 0.007512149401009083, + 0.0069906823337078094, + 0.00678484421223402, + 0.0068422709591686726, + 0.029354829341173172, + 0.019214866682887077, + 0.021543649956583977, + 0.018199466168880463, + 0.012101768516004086, + 0.01510107796639204, + 0.014837722294032574, + 0.034617695957422256, + 0.0043622273951768875, + 0.006198245100677013, + 0.004042746499180794, + 0.004784044809639454, + 0.002889069262892008, + 0.0029548427555710077, + 0.003559727920219302, + 0.0018870317144319415, + 0.018936442211270332, + 0.011537204496562481, + 0.0183999165892601, + 0.015270944684743881, + 0.008668833412230015, + 0.007812810130417347, + 0.009002665057778358, + 0.007122368551790714, + 0.005042139906436205, + 0.006365114822983742, + 0.005631520412862301, + 0.0036874914076179266, + 0.005779807921499014, + 0.005722332280129194, + 0.004640989471226931, + 0.004251124337315559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.016109894961118698 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791914373636246 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.009791914373636246, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791914373636246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29647448658943176, + 0.08467178046703339, + 0.10338200628757477, + 0.03896050900220871, + 0.03942129760980606, + 0.05020212009549141, + 0.28293558955192566, + 0.019571810960769653, + 0.06992466747760773, + 0.014455718919634819 + ], + "credit_map": [ + 0.0338265635073185, + 0.040794361382722855, + 0.052906591445207596, + 0.025691289454698563, + 0.021951524540781975, + 0.05739273875951767, + 0.030512792989611626, + 0.033398617058992386, + 0.010233626700937748, + 0.011610401794314384, + 0.01888025552034378, + 0.007586388383060694, + 0.0072076814249157906, + 0.015703601762652397, + 0.006287329364567995, + 0.007162500638514757, + 0.01313190907239914, + 0.01772170700132847, + 0.010351217351853848, + 0.010143125429749489, + 0.014194981195032597, + 0.011709177866578102, + 0.00958296936005354, + 0.01654692180454731, + 0.006042598746716976, + 0.006957091856747866, + 0.009775733575224876, + 0.0025459397584199905, + 0.0030623602215200663, + 0.004188214428722858, + 0.002557516098022461, + 0.003831056412309408, + 0.0070084030739963055, + 0.004817618057131767, + 0.006117047742009163, + 0.0041346475481987, + 0.005877429619431496, + 0.003852825378999114, + 0.0037063935305923223, + 0.0039069331251084805, + 0.00675489054992795, + 0.004701503086835146, + 0.006693102419376373, + 0.0066055553033947945, + 0.004913562908768654, + 0.00555225508287549, + 0.00936646107584238, + 0.005614789202809334, + 0.04761531576514244, + 0.03462861478328705, + 0.04584331065416336, + 0.03281218186020851, + 0.018959397450089455, + 0.025111038237810135, + 0.021149588748812675, + 0.056816134601831436, + 0.002946277614682913, + 0.003203590400516987, + 0.0026908786967396736, + 0.0035251874942332506, + 0.0018893553642556071, + 0.0016676144441589713, + 0.0024014629889279604, + 0.0012474453542381525, + 0.013406547717750072, + 0.008650967851281166, + 0.010890739969909191, + 0.009236541576683521, + 0.006653667893260717, + 0.0054574343375861645, + 0.009172814898192883, + 0.006455954164266586, + 0.0019507947145029902, + 0.0018829968757927418, + 0.0019076247699558735, + 0.0015697740018367767, + 0.0017228563083335757, + 0.0022432918194681406, + 0.0018308371072635055, + 0.0013475432060658932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.009791914373636246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638862796127796 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.004638862796127796, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638862796127796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13597162067890167, + 0.10793211311101913, + 0.2590909004211426, + 0.05024154856801033, + 0.08754924684762955, + 0.05668313428759575, + 0.113987036049366, + 0.08518153429031372, + 0.07307478040456772, + 0.030288074165582657 + ], + "credit_map": [ + 0.012993047013878822, + 0.018889237195253372, + 0.021485455334186554, + 0.012065587565302849, + 0.01222511287778616, + 0.030299222096800804, + 0.01729222759604454, + 0.010721743106842041, + 0.0231634434312582, + 0.011597453616559505, + 0.02859414927661419, + 0.006751247216016054, + 0.0069669270887970924, + 0.009231639094650745, + 0.008159114979207516, + 0.013468132354319096, + 0.04191719368100166, + 0.04699714481830597, + 0.04831889271736145, + 0.01622193492949009, + 0.02675403468310833, + 0.037494808435440063, + 0.014693566597998142, + 0.02669329196214676, + 0.008014388382434845, + 0.008834310807287693, + 0.010575564578175545, + 0.0033291885629296303, + 0.005449657328426838, + 0.006199460010975599, + 0.002897333586588502, + 0.004941642750054598, + 0.015666890889406204, + 0.00928664579987526, + 0.011328422464430332, + 0.010156681761145592, + 0.010594303719699383, + 0.011950378306210041, + 0.008837444707751274, + 0.009728481993079185, + 0.007853541523218155, + 0.006236528046429157, + 0.005472672171890736, + 0.007359338458627462, + 0.006443931721150875, + 0.008622951805591583, + 0.007129016797989607, + 0.007565155625343323, + 0.01938379555940628, + 0.015375960618257523, + 0.017843179404735565, + 0.013424145057797432, + 0.00717007415369153, + 0.009288830682635307, + 0.009072493761777878, + 0.02242855913937092, + 0.013215040788054466, + 0.015938706696033478, + 0.018206894397735596, + 0.012341569177806377, + 0.0055021983571350574, + 0.006722926627844572, + 0.006498591508716345, + 0.00675560487434268, + 0.013564291410148144, + 0.008843396790325642, + 0.012117755599319935, + 0.012067845091223717, + 0.006697485689073801, + 0.006258679088205099, + 0.007470784243196249, + 0.006054544355720282, + 0.0038648692425340414, + 0.0045180050656199455, + 0.0043887062929570675, + 0.0033054021187126637, + 0.0026732287369668484, + 0.004144731909036636, + 0.004191610496491194, + 0.0032015210017561913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.004638862796127796 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Kamov Ka-52 attack helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the helicopter flying sideways just meters above an arid, sun-scorched ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the helicopter display a muted green and tan military camouflage pattern on its metal fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the distinctive coaxial propellers visible as a blur of counter-rotating motion kicking up a vast cloud of brown dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly and deliberately from a low angle while maintaining tight focus on the cockpit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798689365386963 + ], + "opposite_logprob_tokens": [ + -1.8048689365386963 + ], + "logprob_reward": -0.1798689365386963, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798689365386963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16293442249298096, + 0.07804454118013382, + 0.09148900955915451, + 0.129373237490654, + 0.037161968648433685, + 0.04277805984020233, + 0.17989106476306915, + 0.07354194670915604, + 0.19135670363903046, + 0.013429018668830395 + ], + "credit_map": [ + 0.01721104606986046, + 0.021354151889681816, + 0.026154562830924988, + 0.012813453562557697, + 0.012804586440324783, + 0.03896665573120117, + 0.016099832952022552, + 0.017530124634504318, + 0.015324149280786514, + 0.008313760161399841, + 0.01799880526959896, + 0.004935241304337978, + 0.0071970331482589245, + 0.010862940922379494, + 0.00550017599016428, + 0.007912441156804562, + 0.014923850074410439, + 0.015375617891550064, + 0.018564622849225998, + 0.006955853197723627, + 0.008510266430675983, + 0.01223683636635542, + 0.0057343123480677605, + 0.009187650866806507, + 0.02489614486694336, + 0.02228747121989727, + 0.03302397578954697, + 0.006305759772658348, + 0.012106155976653099, + 0.016946852207183838, + 0.006493881344795227, + 0.0073129963129758835, + 0.008245142176747322, + 0.004781659226864576, + 0.0051296669989824295, + 0.004691353999078274, + 0.0034404878970235586, + 0.00339419674128294, + 0.003790052607655525, + 0.0036894071381539106, + 0.005341507960110903, + 0.0059563880786299706, + 0.0036697268951684237, + 0.004670176655054092, + 0.006369120441377163, + 0.006187243852764368, + 0.00625254213809967, + 0.004331355448812246, + 0.03474435582756996, + 0.02613922581076622, + 0.03660653904080391, + 0.017998075112700462, + 0.011356834322214127, + 0.015202592127025127, + 0.009714031592011452, + 0.02812940999865532, + 0.009717771783471107, + 0.01512912567704916, + 0.009938571602106094, + 0.008960097096860409, + 0.008579099550843239, + 0.007632243912667036, + 0.006505527999252081, + 0.007079510949552059, + 0.040107667446136475, + 0.024518044665455818, + 0.03685738146305084, + 0.027364542707800865, + 0.01751547120511532, + 0.014700036495923996, + 0.01687747985124588, + 0.013416068628430367, + 0.0016410077223554254, + 0.002163950353860855, + 0.0021318309009075165, + 0.0013963341480121017, + 0.0016272622160613537, + 0.001663786475546658, + 0.0015584196662530303, + 0.0012464281171560287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.1798689365386963 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_15/index_tensor([9191, 9191])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_15/index_tensor([9191, 9191])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..132e3f5d3165d95ba2aabce7f9922e1d67087e21 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_15/index_tensor([9191, 9191])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This low-angle, tight medium shot, captured with documentary-style realism and a slight handheld shake, immerses the viewer in the waters of a clear, shallow river. The central focus is two majestic sockeye salmon, their bodies a brilliant, pristine crimson, their heads a contrasting metallic emerald, as they fiercely thrash against the current. Their powerful bodies collide and twist, churning the water into a white foam and sending arcs of shimmering droplets into the air. Their hooked jaws open and close with aggressive speed, revealing the white interiors of their mouths in a desperate fight for spawning rights. The entire scene is rendered in a stark, desaturated color grading, muting the greens and browns of the out-of-focus riverbank and the multicolored pebbles visible through the clear water. This deliberate artistic choice allows the visceral, saturated crimson of the salmon's bodies to stand out with a shocking intensity against the muted surroundings, creating an atmosphere of tension and drama focused purely on the brutal beauty of their fierce, instinctive struggle.", + "seed": 28618, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_15/index_9191_seed_28618_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12660327786579728, + "vlm_logprob": -0.12660327786579728, + "vlm_yes_no_margin": 3.299999952316284, + "vlm_token_credit": -0.12660327786579728, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013906371779739857 + ], + "opposite_logprob_tokens": [ + -4.638906478881836 + ], + "logprob_reward": -0.013906371779739857, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013906371779739857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12930305302143097, + 0.09616022557020187, + 0.05573742091655731, + 0.03771053999662399, + 0.18660181760787964, + 0.2507694661617279, + 0.02218133583664894, + 0.07390515506267548, + 0.08876802772283554, + 0.058863040059804916 + ], + "credit_map": [ + 0.012305096723139286, + 0.02396535687148571, + 0.016342615708708763, + 0.014757618308067322, + 0.012011476792395115, + 0.01605760119855404, + 0.018398422747850418, + 0.015464858151972294, + 0.008645249530673027, + 0.013322107493877411, + 0.010764113627374172, + 0.009910529479384422, + 0.015999769791960716, + 0.00899224542081356, + 0.015037862583994865, + 0.0134883439168334, + 0.006137578748166561, + 0.006270777899771929, + 0.007484240457415581, + 0.00507748918607831, + 0.003541218116879463, + 0.01342606171965599, + 0.005842271260917187, + 0.007957777939736843, + 0.004229298792779446, + 0.0044748964719474316, + 0.004157195799052715, + 0.005609527230262756, + 0.004725626204162836, + 0.0038626811001449823, + 0.005101970862597227, + 0.0055493442341685295, + 0.015669086948037148, + 0.019800228998064995, + 0.023094985634088516, + 0.018509460613131523, + 0.025483882054686546, + 0.04597892984747887, + 0.015001854859292507, + 0.023063378408551216, + 0.02020782232284546, + 0.016288479790091515, + 0.0690544843673706, + 0.01717882975935936, + 0.014504792168736458, + 0.016909539699554443, + 0.01678933948278427, + 0.0798361599445343, + 0.002664607483893633, + 0.0023664834443479776, + 0.002673364244401455, + 0.0018119504675269127, + 0.003260199446231127, + 0.003943862859159708, + 0.0025214431807398796, + 0.002939424477517605, + 0.004662571009248495, + 0.009573515504598618, + 0.003373408690094948, + 0.006271228194236755, + 0.018561309203505516, + 0.005639995448291302, + 0.006051331292837858, + 0.019771797582507133, + 0.006863363087177277, + 0.010005129501223564, + 0.010971623472869396, + 0.008323094807565212, + 0.017905814573168755, + 0.01225086860358715, + 0.01150074414908886, + 0.010947390459477901, + 0.005445315968245268, + 0.00852214265614748, + 0.006285405717790127, + 0.00889354944229126, + 0.0077982270158827305, + 0.007085664197802544, + 0.007152288220822811, + 0.007680445909500122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.013906371779739857 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700201466679573 + ], + "opposite_logprob_tokens": [ + -2.6950201988220215 + ], + "logprob_reward": -0.0700201466679573, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700201466679573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09186957776546478, + 0.17425212264060974, + 0.2342357039451599, + 0.09329099208116531, + 0.04835418611764908, + 0.17419806122779846, + 0.021044250577688217, + 0.0959923192858696, + 0.042678579688072205, + 0.024084217846393585 + ], + "credit_map": [ + 0.009343380108475685, + 0.012368924915790558, + 0.01084135752171278, + 0.010080994106829166, + 0.01294123474508524, + 0.008472924120724201, + 0.01806553453207016, + 0.009755225852131844, + 0.01607779785990715, + 0.023442553356289864, + 0.027656257152557373, + 0.017366686835885048, + 0.0405716747045517, + 0.017295481637120247, + 0.015757650136947632, + 0.016084013506770134, + 0.024456415325403214, + 0.02003541588783264, + 0.041762519627809525, + 0.02242772839963436, + 0.016057848930358887, + 0.05108628422021866, + 0.026118380948901176, + 0.03229111433029175, + 0.009194323793053627, + 0.008453994058072567, + 0.008867193944752216, + 0.01138402335345745, + 0.011044730432331562, + 0.017229631543159485, + 0.011426959186792374, + 0.015690134838223457, + 0.005694567691534758, + 0.0047529046423733234, + 0.0057331593707203865, + 0.005530322436243296, + 0.005572014022618532, + 0.005323647987097502, + 0.0059154764749109745, + 0.009832093492150307, + 0.01547119952738285, + 0.011411960236728191, + 0.047069016844034195, + 0.011701952666044235, + 0.010468271560966969, + 0.012429322116076946, + 0.011581688188016415, + 0.054064664989709854, + 0.0019397581927478313, + 0.0023356883320957422, + 0.0016208161832764745, + 0.0034789012279361486, + 0.003931362647563219, + 0.003418661653995514, + 0.002109651919454336, + 0.002209410071372986, + 0.005963762756437063, + 0.00819556973874569, + 0.0040851766243577, + 0.008018692955374718, + 0.029391471296548843, + 0.008111861534416676, + 0.004434344824403524, + 0.02779144048690796, + 0.0033261298667639494, + 0.004773628897964954, + 0.004786536563187838, + 0.004776055924594402, + 0.00824738759547472, + 0.010026372037827969, + 0.003335594432428479, + 0.0034068755339831114, + 0.0024798084050416946, + 0.004247536417096853, + 0.0030470779165625572, + 0.0034503531642258167, + 0.0030494232196360826, + 0.002429995918646455, + 0.0027148150838911533, + 0.00266520818695426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.0700201466679573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625873666256666 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007625873666256666, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625873666256666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12691950798034668, + 0.29003480076789856, + 0.07447005808353424, + 0.0903121754527092, + 0.04361753165721893, + 0.14436933398246765, + 0.03390403464436531, + 0.09958082437515259, + 0.04950113222002983, + 0.047290608286857605 + ], + "credit_map": [ + 0.011556324549019337, + 0.020976781845092773, + 0.016661327332258224, + 0.013918104581534863, + 0.01321196835488081, + 0.014110418036580086, + 0.02235923521220684, + 0.014125337824225426, + 0.02325384132564068, + 0.040191974490880966, + 0.037582382559776306, + 0.02884315326809883, + 0.06043117865920067, + 0.02903798408806324, + 0.0376911498606205, + 0.03300315514206886, + 0.01293894276022911, + 0.007255619391798973, + 0.011796724051237106, + 0.007296337280422449, + 0.007723876740783453, + 0.014119270257651806, + 0.0061220512725412846, + 0.007217239122837782, + 0.009222712367773056, + 0.009407708421349525, + 0.009376383386552334, + 0.012389273382723331, + 0.011366152204573154, + 0.012955337762832642, + 0.011520233005285263, + 0.014074371196329594, + 0.005407959222793579, + 0.005843775346875191, + 0.007173899561166763, + 0.0043740165419876575, + 0.0039617689326405525, + 0.010359074920415878, + 0.0028420963790267706, + 0.0036549416836351156, + 0.014629298821091652, + 0.009427804499864578, + 0.03761305287480354, + 0.011024802923202515, + 0.009972486644983292, + 0.01015047263354063, + 0.00869024358689785, + 0.042861174792051315, + 0.0031539686024188995, + 0.0031395151745527983, + 0.0033653383143246174, + 0.006206783466041088, + 0.00582940224558115, + 0.004329203628003597, + 0.0037818311247974634, + 0.004097987897694111, + 0.008234445936977863, + 0.019738705828785896, + 0.009250715374946594, + 0.009350094012916088, + 0.017994733527302742, + 0.013918240554630756, + 0.009186445735394955, + 0.011907440610229969, + 0.005106421187520027, + 0.004977711476385593, + 0.007432225160300732, + 0.005083131138235331, + 0.006493245251476765, + 0.009540088474750519, + 0.005881939548999071, + 0.004986371845006943, + 0.004617589525878429, + 0.005778163205832243, + 0.006798737216740847, + 0.00636284239590168, + 0.005294428206980228, + 0.005205458030104637, + 0.004827817901968956, + 0.00840557087212801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.007625873666256666 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5232834815979004 + ], + "opposite_logprob_tokens": [ + -0.8982834815979004 + ], + "logprob_reward": -0.5232834815979004, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5232834815979004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10946818441152573, + 0.20590393245220184, + 0.033604998141527176, + 0.07487817853689194, + 0.23181471228599548, + 0.060876354575157166, + 0.025336774066090584, + 0.0681362897157669, + 0.044964537024497986, + 0.14501601457595825 + ], + "credit_map": [ + 0.010560588911175728, + 0.015568885020911694, + 0.016561061143875122, + 0.011928525753319263, + 0.015181680209934711, + 0.0135993966832757, + 0.013640434481203556, + 0.012427611276507378, + 0.0178968645632267, + 0.027545690536499023, + 0.029203876852989197, + 0.021830148994922638, + 0.034491028636693954, + 0.018216688185930252, + 0.030519217252731323, + 0.02620043233036995, + 0.00588828232139349, + 0.004724035505205393, + 0.004962571430951357, + 0.0038561502005904913, + 0.0027120569720864296, + 0.005928280297666788, + 0.0026162192225456238, + 0.0029174056835472584, + 0.009242774918675423, + 0.010954274795949459, + 0.006789252627640963, + 0.011851313523948193, + 0.007358924485743046, + 0.009315339848399162, + 0.009662763215601444, + 0.009703543037176132, + 0.031251680105924606, + 0.023736560717225075, + 0.0408988893032074, + 0.021860182285308838, + 0.02254549041390419, + 0.04828476905822754, + 0.0251981932669878, + 0.018038980662822723, + 0.005206060130149126, + 0.003931726329028606, + 0.015498383902013302, + 0.004109818022698164, + 0.003721232060343027, + 0.005084584467113018, + 0.00474165752530098, + 0.018582895398139954, + 0.0025816941633820534, + 0.004163793288171291, + 0.0018272784072905779, + 0.005388778634369373, + 0.003231002716347575, + 0.0025971964932978153, + 0.0025031911209225655, + 0.0030438415706157684, + 0.004742699675261974, + 0.006325836759060621, + 0.00500113470479846, + 0.006748056039214134, + 0.017195040360093117, + 0.00782362837344408, + 0.005278652999550104, + 0.015021249651908875, + 0.004610611591488123, + 0.006826083641499281, + 0.005787963047623634, + 0.005810567643493414, + 0.0046283709816634655, + 0.006352642085403204, + 0.005623518489301205, + 0.005324782337993383, + 0.012442674487829208, + 0.017123540863394737, + 0.018023060634732246, + 0.022017760202288628, + 0.01628423109650612, + 0.01779019646346569, + 0.020971355959773064, + 0.020363206043839455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.5232834815979004 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180515617132187 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180515617132187, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018180515617132187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09206226468086243, + 0.28257524967193604, + 0.03370284661650658, + 0.09693329781293869, + 0.08785498887300491, + 0.13569381833076477, + 0.061069607734680176, + 0.04575761780142784, + 0.029331233352422714, + 0.13501903414726257 + ], + "credit_map": [ + 0.010551560670137405, + 0.011581534519791603, + 0.016568506136536598, + 0.010975655168294907, + 0.007885947823524475, + 0.013450795784592628, + 0.01225039642304182, + 0.008797872811555862, + 0.02198064886033535, + 0.036703988909721375, + 0.0381443053483963, + 0.02797509916126728, + 0.05682196468114853, + 0.02741422690451145, + 0.04126351699233055, + 0.03227151557803154, + 0.005085407756268978, + 0.0032621591817587614, + 0.005005988292396069, + 0.003004379104822874, + 0.002881451277062297, + 0.008065257221460342, + 0.0032093969639390707, + 0.0031888096127659082, + 0.01104075089097023, + 0.009964201599359512, + 0.008539276197552681, + 0.011591391637921333, + 0.013399733230471611, + 0.015016471035778522, + 0.012833498418331146, + 0.014547989703714848, + 0.011267322115600109, + 0.008488606661558151, + 0.014745539985597134, + 0.009103097952902317, + 0.010517312213778496, + 0.017916133627295494, + 0.009274516254663467, + 0.006542470306158066, + 0.013055781833827496, + 0.00899733230471611, + 0.036989450454711914, + 0.008486361242830753, + 0.007481585256755352, + 0.009687086567282677, + 0.009546751156449318, + 0.04144948720932007, + 0.006821274291723967, + 0.010664782486855984, + 0.0065808468498289585, + 0.011248884722590446, + 0.007875208742916584, + 0.007299861405044794, + 0.0064189741387963295, + 0.004159779287874699, + 0.003516831435263157, + 0.005173295736312866, + 0.004928156267851591, + 0.0052439384162425995, + 0.007765751797705889, + 0.006556229200214148, + 0.005294166970998049, + 0.007279250305145979, + 0.0031759322155267, + 0.0033411451149731874, + 0.0020732099656015635, + 0.002750137122347951, + 0.005865080747753382, + 0.004691012669354677, + 0.0037686380092054605, + 0.0036660826299339533, + 0.011653042398393154, + 0.016487477347254753, + 0.016159744933247566, + 0.019626008346676826, + 0.01594509929418564, + 0.016625825315713882, + 0.018758263438940048, + 0.019763585180044174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.018180515617132187 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This low-angle, tight medium shot, captured with documentary-style realism and a slight handheld shake, immerses the viewer in the waters of a clear, shallow river. The central focus is two majestic sockeye salmon, their bodies a brilliant, pristine crimson, their heads a contrasting metallic emerald, as they fiercely thrash against the current. Their powerful bodies collide and twist, churning the water into a white foam and sending arcs of shimmering droplets into the air. Their hooked jaws open and close with aggressive speed, revealing the white interiors of their mouths in a desperate fight for spawning rights. The entire scene is rendered in a stark, desaturated color grading, muting the greens and browns of the out-of-focus riverbank and the multicolored pebbles visible through the clear water. This deliberate artistic choice allows the visceral, saturated crimson of the salmon's bodies to stand out with a shocking intensity against the muted surroundings, creating an atmosphere of tension and drama focused purely on the brutal beauty of their fierce, instinctive struggle.", + "seed": 28619, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_15/index_9191_seed_28619_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2645286167040467, + "vlm_logprob": -0.2645286167040467, + "vlm_yes_no_margin": 2.749999988079071, + "vlm_token_credit": -0.2645286167040467, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022875003516674042 + ], + "opposite_logprob_tokens": [ + -4.27287483215332 + ], + "logprob_reward": -0.022875003516674042, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.022875003516674042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05094147101044655, + 0.020007919520139694, + 0.029117008671164513, + 0.023530973121523857, + 0.031823765486478806, + 0.15196435153484344, + 0.12145544588565826, + 0.3130980134010315, + 0.20967985689640045, + 0.04838120564818382 + ], + "credit_map": [ + 0.007340059615671635, + 0.005963602568954229, + 0.007913210429251194, + 0.005577528849244118, + 0.00749543122947216, + 0.0065947650000452995, + 0.004564169328659773, + 0.005492710042744875, + 0.0021877733524888754, + 0.001482205232605338, + 0.002833029255270958, + 0.002463772427290678, + 0.00218744995072484, + 0.0024975179694592953, + 0.0031513150315731764, + 0.0032048586290329695, + 0.0021551521494984627, + 0.002885689027607441, + 0.0021762975957244635, + 0.0026758755557239056, + 0.0043861777521669865, + 0.00586185185238719, + 0.0040923962369561195, + 0.004883572459220886, + 0.0032021235674619675, + 0.004177355673164129, + 0.0028325302992016077, + 0.0019925879314541817, + 0.0025807323399931192, + 0.0035561874974519014, + 0.002557122614234686, + 0.00263233738951385, + 0.0048319087363779545, + 0.00517340237274766, + 0.005108190234750509, + 0.0027993330731987953, + 0.0033908102195709944, + 0.0032325375359505415, + 0.0034399917349219322, + 0.0038475943729281425, + 0.011651537381112576, + 0.01242644153535366, + 0.016519254073500633, + 0.025839686393737793, + 0.026600142940878868, + 0.0159663837403059, + 0.019616976380348206, + 0.023343950510025024, + 0.008614350110292435, + 0.004060705192387104, + 0.010508605279028416, + 0.006793396081775427, + 0.005024193786084652, + 0.04020606726408005, + 0.009310056455433369, + 0.03693808242678642, + 0.01947803422808647, + 0.023281438276171684, + 0.02700023539364338, + 0.03328362852334976, + 0.10010730475187302, + 0.03668168559670448, + 0.021197576075792313, + 0.052068136632442474, + 0.02014591544866562, + 0.021920187398791313, + 0.030066851526498795, + 0.03780625760555267, + 0.0238199345767498, + 0.023234622552990913, + 0.029155351221561432, + 0.023530760779976845, + 0.0038524491246789694, + 0.005900057032704353, + 0.00885832030326128, + 0.005559252109378576, + 0.00729965977370739, + 0.004411324393004179, + 0.006356884725391865, + 0.006143262609839439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.022875003516674042 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294341832399368 + ], + "opposite_logprob_tokens": [ + -2.237943410873413 + ], + "logprob_reward": -0.11294341832399368, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294341832399368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06053721904754639, + 0.033386651426553726, + 0.13223978877067566, + 0.02148321084678173, + 0.06180959194898605, + 0.06932219117879868, + 0.11425315588712692, + 0.22485941648483276, + 0.2227294147014618, + 0.059379346668720245 + ], + "credit_map": [ + 0.009891529567539692, + 0.007396320346742868, + 0.0088516129180789, + 0.0048220232129096985, + 0.006830403581261635, + 0.007939938455820084, + 0.0070704510435462, + 0.007734933868050575, + 0.004135047551244497, + 0.004250972997397184, + 0.004013095051050186, + 0.0035784034989774227, + 0.004590858239680529, + 0.003708799369633198, + 0.004002423956990242, + 0.005107048898935318, + 0.008040403947234154, + 0.010310175828635693, + 0.00438030157238245, + 0.007877982221543789, + 0.013821154832839966, + 0.049001600593328476, + 0.008723516948521137, + 0.030084647238254547, + 0.002971670823171735, + 0.0021112821996212006, + 0.002532693324610591, + 0.0021816997323185205, + 0.0028656248468905687, + 0.003053000196814537, + 0.002662739483639598, + 0.0031044979114085436, + 0.009364844299852848, + 0.007462512701749802, + 0.006302389316260815, + 0.0072195627726614475, + 0.009096033871173859, + 0.00541651202365756, + 0.010217889212071896, + 0.006729842163622379, + 0.0067417980171740055, + 0.004636991769075394, + 0.0070850360207259655, + 0.007773950230330229, + 0.013665670529007912, + 0.006508458871394396, + 0.007300263270735741, + 0.01561002153903246, + 0.008249777369201183, + 0.005035333801060915, + 0.010756710544228554, + 0.007180800195783377, + 0.004201246425509453, + 0.03543562814593315, + 0.007778288330882788, + 0.03561536222696304, + 0.018801715224981308, + 0.017703194171190262, + 0.02354205586016178, + 0.02942429482936859, + 0.059152230620384216, + 0.023242942988872528, + 0.019192123785614967, + 0.03380085155367851, + 0.022526001557707787, + 0.02099197544157505, + 0.03079705871641636, + 0.039438799023628235, + 0.02606760524213314, + 0.028532521799206734, + 0.030308660119771957, + 0.02406679280102253, + 0.004658050369471312, + 0.005775324534624815, + 0.009027903899550438, + 0.007621329743415117, + 0.008938700892031193, + 0.005962407682090998, + 0.010203039273619652, + 0.0071925874799489975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.11294341832399368 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738794036209583 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738794036209583, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738794036209583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11205438524484634, + 0.07089202105998993, + 0.13621638715267181, + 0.0923902690410614, + 0.06285961717367172, + 0.18985356390476227, + 0.10255284607410431, + 0.08808328956365585, + 0.0674566850066185, + 0.07764095067977905 + ], + "credit_map": [ + 0.01297832652926445, + 0.011972174979746342, + 0.01563209854066372, + 0.010219786316156387, + 0.014772972092032433, + 0.01541891973465681, + 0.013560858555138111, + 0.017499245703220367, + 0.00703828502446413, + 0.007609330117702484, + 0.00866628810763359, + 0.007929190993309021, + 0.008568402379751205, + 0.007862974889576435, + 0.011211381293833256, + 0.012006159871816635, + 0.00909392163157463, + 0.008265357464551926, + 0.007528285495936871, + 0.012032546103000641, + 0.015853162854909897, + 0.04415108263492584, + 0.009458384476602077, + 0.02983364649116993, + 0.010846258141100407, + 0.011224564164876938, + 0.014903348870575428, + 0.010958154685795307, + 0.010220283642411232, + 0.011914149858057499, + 0.009727125987410545, + 0.012596379965543747, + 0.010312565602362156, + 0.007270910311490297, + 0.008075390942394733, + 0.007900355383753777, + 0.009221337735652924, + 0.005683351773768663, + 0.006679270416498184, + 0.0077164312824606895, + 0.015049219131469727, + 0.015249188058078289, + 0.016424495726823807, + 0.02855309098958969, + 0.03967767953872681, + 0.018358787521719933, + 0.020657854154706, + 0.035883236676454544, + 0.011684075929224491, + 0.005450431257486343, + 0.016830597072839737, + 0.011561215855181217, + 0.005633039865642786, + 0.02221338078379631, + 0.010194974020123482, + 0.018985116854310036, + 0.00549854664131999, + 0.0066472808830440044, + 0.007562373764812946, + 0.008485034108161926, + 0.025593571364879608, + 0.011539858765900135, + 0.008816452696919441, + 0.013940166682004929, + 0.010121584869921207, + 0.007105882279574871, + 0.01033305749297142, + 0.009674486704170704, + 0.00782802328467369, + 0.007171055767685175, + 0.007603120058774948, + 0.007619471754878759, + 0.006071221083402634, + 0.008331396616995335, + 0.011160430498421192, + 0.008745690807700157, + 0.014987434260547161, + 0.007108971942216158, + 0.012589351274073124, + 0.008646449074149132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.006738794036209583 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.137019157409668 + ], + "opposite_logprob_tokens": [ + -0.3870190978050232 + ], + "logprob_reward": -1.137019157409668, + "yes_no_margin_reward": -0.7500000596046448, + "token_credit_reward": -1.137019157409668, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10780405253171921, + 0.027273915708065033, + 0.23262959718704224, + 0.08674688637256622, + 0.031746845692396164, + 0.12953530251979828, + 0.045378178358078, + 0.1393759548664093, + 0.05935736000537872, + 0.1401519477367401 + ], + "credit_map": [ + 0.014346880838274956, + 0.009352386929094791, + 0.010636985301971436, + 0.008309725672006607, + 0.011939730495214462, + 0.01166459172964096, + 0.019400855526328087, + 0.022152891382575035, + 0.002663800958544016, + 0.0032501970417797565, + 0.002314962213858962, + 0.00315547245554626, + 0.003940968308597803, + 0.0028198272921144962, + 0.004260191228240728, + 0.0048684971407055855, + 0.013692216947674751, + 0.0199210736900568, + 0.01317574642598629, + 0.023138443008065224, + 0.023167308419942856, + 0.07786009460687637, + 0.012800616212189198, + 0.04887410253286362, + 0.013066597282886505, + 0.00922370795160532, + 0.013386300764977932, + 0.008750826120376587, + 0.007205683272331953, + 0.014041310176253319, + 0.010865015909075737, + 0.010207444429397583, + 0.004333344753831625, + 0.00449597742408514, + 0.005250392481684685, + 0.003635231638327241, + 0.004101426340639591, + 0.0032899449579417706, + 0.0037912188563495874, + 0.0028493080753833055, + 0.01762886717915535, + 0.008763084188103676, + 0.018775269389152527, + 0.012239106930792332, + 0.013918214477598667, + 0.016489634290337563, + 0.012866154313087463, + 0.0288549717515707, + 0.004035390447825193, + 0.0028410758823156357, + 0.004405147861689329, + 0.003688013879582286, + 0.002945826854556799, + 0.01086506713181734, + 0.004766357596963644, + 0.011831296607851982, + 0.015437831170856953, + 0.009783525951206684, + 0.01527805533260107, + 0.015132037922739983, + 0.033001262694597244, + 0.01719343289732933, + 0.01179481390863657, + 0.021754994988441467, + 0.00677088275551796, + 0.005766641814261675, + 0.00839854497462511, + 0.009956267662346363, + 0.008280267007648945, + 0.006938685197383165, + 0.00720730796456337, + 0.006038764491677284, + 0.011477767489850521, + 0.014697406440973282, + 0.019625570625066757, + 0.021706536412239075, + 0.017201673239469528, + 0.015207363292574883, + 0.027250919491052628, + 0.012984715402126312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -1.137019157409668 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043066710233688354 + ], + "opposite_logprob_tokens": [ + -3.1680667400360107 + ], + "logprob_reward": -0.043066710233688354, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043066710233688354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07370992004871368, + 0.029593495652079582, + 0.17229971289634705, + 0.03420791029930115, + 0.05219529941678047, + 0.10031380504369736, + 0.07562672346830368, + 0.22767797112464905, + 0.15155892074108124, + 0.08281618356704712 + ], + "credit_map": [ + 0.008580256253480911, + 0.008299091830849648, + 0.009170752950012684, + 0.005683063063770533, + 0.008355481550097466, + 0.00919139664620161, + 0.010842286050319672, + 0.013587593100965023, + 0.002949354238808155, + 0.0032217882107943296, + 0.003151256823912263, + 0.0029137858655303717, + 0.0032450570724904537, + 0.0035150530748069286, + 0.004274771083146334, + 0.006322428118437529, + 0.00991563219577074, + 0.012545757927000523, + 0.009045028127729893, + 0.011732755228877068, + 0.012520093470811844, + 0.06603413820266724, + 0.008952328003942966, + 0.04155397415161133, + 0.00685421098023653, + 0.0037833352107554674, + 0.00397343747317791, + 0.003545399522408843, + 0.003261582227423787, + 0.004285239614546299, + 0.004777745343744755, + 0.0037269603926688433, + 0.008342219516634941, + 0.004611024167388678, + 0.010316401720046997, + 0.005882627330720425, + 0.006501935888081789, + 0.004684237763285637, + 0.006238851230591536, + 0.00561800180003047, + 0.008898102678358555, + 0.010966789908707142, + 0.013553871773183346, + 0.01575721614062786, + 0.012191289104521275, + 0.012999218888580799, + 0.014995790086686611, + 0.01095152273774147, + 0.006881088018417358, + 0.004062814638018608, + 0.007445890456438065, + 0.007342966739088297, + 0.0032444573007524014, + 0.02068094164133072, + 0.0052835759706795216, + 0.020684989169239998, + 0.01381597574800253, + 0.01775331236422062, + 0.02100880816578865, + 0.025669576600193977, + 0.06902759522199631, + 0.026721635833382607, + 0.0189643744379282, + 0.03471669182181358, + 0.014190029352903366, + 0.01494213379919529, + 0.021516630426049232, + 0.02896655723452568, + 0.017681919038295746, + 0.017833586782217026, + 0.020399095490574837, + 0.01602896675467491, + 0.007589856628328562, + 0.00915090274065733, + 0.011913258582353592, + 0.013294369913637638, + 0.00911287683993578, + 0.008950719609856606, + 0.014624572359025478, + 0.008179628290235996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.043066710233688354 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "This low-angle, tight medium shot, captured with documentary-style realism and a slight handheld shake, immerses the viewer in the waters of a clear, shallow river. The central focus is two majestic sockeye salmon, their bodies a brilliant, pristine crimson, their heads a contrasting metallic emerald, as they fiercely thrash against the current. Their powerful bodies collide and twist, churning the water into a white foam and sending arcs of shimmering droplets into the air. Their hooked jaws open and close with aggressive speed, revealing the white interiors of their mouths in a desperate fight for spawning rights. The entire scene is rendered in a stark, desaturated color grading, muting the greens and browns of the out-of-focus riverbank and the multicolored pebbles visible through the clear water. This deliberate artistic choice allows the visceral, saturated crimson of the salmon's bodies to stand out with a shocking intensity against the muted surroundings, creating an atmosphere of tension and drama focused purely on the brutal beauty of their fierce, instinctive struggle.", + "seed": 28620, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_15/index_9191_seed_28620_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10525983031839133, + "vlm_logprob": -0.10525983031839133, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.10525983031839133, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01392306573688984 + ], + "opposite_logprob_tokens": [ + -4.763923168182373 + ], + "logprob_reward": -0.01392306573688984, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01392306573688984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3170953691005707, + 0.1012328490614891, + 0.07238312810659409, + 0.05157647654414177, + 0.15607266128063202, + 0.06695584207773209, + 0.04436054453253746, + 0.04294177517294884, + 0.0493231900036335, + 0.09805820137262344 + ], + "credit_map": [ + 0.026082249358296394, + 0.04492194950580597, + 0.04725589603185654, + 0.037878647446632385, + 0.05869065970182419, + 0.04116461053490639, + 0.03458866477012634, + 0.026512669399380684, + 0.0136664267629385, + 0.007260180544108152, + 0.02259352058172226, + 0.0139531921595335, + 0.005547805223613977, + 0.006348593160510063, + 0.01830265112221241, + 0.013560484163463116, + 0.007677391637116671, + 0.007694849744439125, + 0.007331309840083122, + 0.013569037429988384, + 0.009449982084333897, + 0.011414730921387672, + 0.009022671729326248, + 0.006223154719918966, + 0.0054701860062778, + 0.005289844702929258, + 0.007257533259689808, + 0.005861747544258833, + 0.01003117486834526, + 0.007206013426184654, + 0.00483650341629982, + 0.005623471457511187, + 0.01267168764024973, + 0.01297564897686243, + 0.02630689926445484, + 0.014443228021264076, + 0.03442145138978958, + 0.019281618297100067, + 0.013872746378183365, + 0.022099371999502182, + 0.007673092186450958, + 0.007231342140585184, + 0.0075565786100924015, + 0.008275668136775494, + 0.007684274110943079, + 0.009134133346378803, + 0.010644366964697838, + 0.008756386116147041, + 0.004479421302676201, + 0.004551451187580824, + 0.0068336790427565575, + 0.007188483607023954, + 0.003195948200300336, + 0.005153993144631386, + 0.007274718023836613, + 0.005682848859578371, + 0.007426166441291571, + 0.0027058294508606195, + 0.004664204083383083, + 0.007268531713634729, + 0.00400192104279995, + 0.0036267442628741264, + 0.007095692213624716, + 0.006152682937681675, + 0.005761377979069948, + 0.004531765356659889, + 0.008419386111199856, + 0.007914816029369831, + 0.003238057717680931, + 0.006222674623131752, + 0.006924847606569529, + 0.006310266442596912, + 0.01006089337170124, + 0.022560495883226395, + 0.008152458816766739, + 0.007995592430233955, + 0.01326720230281353, + 0.012516174465417862, + 0.01546507142484188, + 0.008040321059525013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.01392306573688984 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054969415068626404 + ], + "opposite_logprob_tokens": [ + -2.929969310760498 + ], + "logprob_reward": -0.054969415068626404, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054969415068626404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36907142400741577, + 0.10300958901643753, + 0.054875489324331284, + 0.02845904976129532, + 0.03701182082295418, + 0.09890391677618027, + 0.05447642132639885, + 0.15062250196933746, + 0.03485282137989998, + 0.06871697306632996 + ], + "credit_map": [ + 0.02613367885351181, + 0.06415637582540512, + 0.04492070525884628, + 0.04749108478426933, + 0.06798989325761795, + 0.043435223400592804, + 0.04159048572182655, + 0.033353980630636215, + 0.01058812253177166, + 0.006150693632662296, + 0.012441056780517101, + 0.025522608309984207, + 0.006440655793994665, + 0.007946979254484177, + 0.017717542126774788, + 0.01620193012058735, + 0.00749244587495923, + 0.005272629205137491, + 0.005930345505475998, + 0.009145661257207394, + 0.008191676810383797, + 0.007210811134427786, + 0.00727654155343771, + 0.004355379845947027, + 0.003958297893404961, + 0.0033822674304246902, + 0.004161030054092407, + 0.0031285618897527456, + 0.004705378785729408, + 0.0034889322705566883, + 0.002300349297001958, + 0.0033342326059937477, + 0.004179227165877819, + 0.0028695736546069384, + 0.005193839780986309, + 0.003442426910623908, + 0.004015799146145582, + 0.00782267190515995, + 0.0037643059622496367, + 0.005723979789763689, + 0.010604985989630222, + 0.011282597668468952, + 0.011170366778969765, + 0.011546245776116848, + 0.0116050336509943, + 0.011442002840340137, + 0.01697472482919693, + 0.014277957379817963, + 0.007276540156453848, + 0.005253651645034552, + 0.007890559732913971, + 0.009106524288654327, + 0.004300104919821024, + 0.004973193164914846, + 0.009604761376976967, + 0.006071085575968027, + 0.013808497227728367, + 0.01166904903948307, + 0.040469586849212646, + 0.009262599982321262, + 0.007344118785113096, + 0.010971459560096264, + 0.011558913625776768, + 0.0455382838845253, + 0.0049157701432704926, + 0.0038046680856496096, + 0.004620382096618414, + 0.005182843655347824, + 0.003998258151113987, + 0.003107985947281122, + 0.0052291410975158215, + 0.0039937738329172134, + 0.006464673671871424, + 0.012125968933105469, + 0.006576495245099068, + 0.005997163709253073, + 0.008868932723999023, + 0.00899195671081543, + 0.014193026348948479, + 0.005498755723237991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.054969415068626404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253794603049755 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.005253794603049755, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253794603049755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30010858178138733, + 0.11472707986831665, + 0.07749813795089722, + 0.04971140995621681, + 0.06638974696397781, + 0.10672785341739655, + 0.05950045958161354, + 0.023314867168664932, + 0.0848732739686966, + 0.11714861541986465 + ], + "credit_map": [ + 0.025448095053434372, + 0.04905208200216293, + 0.037612274289131165, + 0.03167765587568283, + 0.0630718544125557, + 0.033692099153995514, + 0.030538776889443398, + 0.029015732929110527, + 0.012660342268645763, + 0.007085898891091347, + 0.020067915320396423, + 0.01898086816072464, + 0.007339044474065304, + 0.00910411961376667, + 0.01427715364843607, + 0.025211727246642113, + 0.013740291818976402, + 0.009002000093460083, + 0.007707259617745876, + 0.009923536330461502, + 0.010818345472216606, + 0.010774809867143631, + 0.007661939132958651, + 0.007869955152273178, + 0.00680833775550127, + 0.0056838239543139935, + 0.007645016070455313, + 0.0074169146828353405, + 0.0058996113948524, + 0.006505217868834734, + 0.0044031087309122086, + 0.005349375773221254, + 0.005865368526428938, + 0.004754901863634586, + 0.009677513502538204, + 0.006792940199375153, + 0.007675085216760635, + 0.012611505575478077, + 0.007677124347537756, + 0.011335302144289017, + 0.012359611690044403, + 0.012079623527824879, + 0.01221739873290062, + 0.013193270191550255, + 0.012770396657288074, + 0.012984927743673325, + 0.015931595116853714, + 0.015191021375358105, + 0.01017762441188097, + 0.006049857474863529, + 0.008906710892915726, + 0.008233053609728813, + 0.004644004628062248, + 0.007184735499322414, + 0.006972800474613905, + 0.00733167165890336, + 0.0039023300632834435, + 0.0026926309801638126, + 0.002392763737589121, + 0.0028632916510105133, + 0.0024437836837023497, + 0.0018975504208356142, + 0.003568601794540882, + 0.003553912043571472, + 0.01351089682430029, + 0.007958673872053623, + 0.018260061740875244, + 0.007617177441716194, + 0.0052880337461829185, + 0.009248790331184864, + 0.017601557075977325, + 0.005388083402067423, + 0.01112560834735632, + 0.0298237893730402, + 0.007609517779201269, + 0.010260599665343761, + 0.015750257298350334, + 0.013162199407815933, + 0.018258705735206604, + 0.011157937347888947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.005253794603049755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288671314716339 + ], + "opposite_logprob_tokens": [ + -1.0538671016693115 + ], + "logprob_reward": -0.4288671314716339, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288671314716339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0759531557559967, + 0.11745849996805191, + 0.039957717061042786, + 0.08680643141269684, + 0.026433011516928673, + 0.04579402506351471, + 0.1466793715953827, + 0.3252374529838562, + 0.025395020842552185, + 0.11028527468442917 + ], + "credit_map": [ + 0.00565064512193203, + 0.011297566816210747, + 0.008066704496741295, + 0.0115268649533391, + 0.011972740292549133, + 0.007442744914442301, + 0.012557859532535076, + 0.007438033819198608, + 0.010958924889564514, + 0.007511531468480825, + 0.020270463079214096, + 0.01577005907893181, + 0.004675588104873896, + 0.00915138516575098, + 0.01598825864493847, + 0.033132292330265045, + 0.005268912296742201, + 0.00415800092741847, + 0.004255327861756086, + 0.007658130954951048, + 0.004935920238494873, + 0.0042598070576786995, + 0.004988613538444042, + 0.004433008376508951, + 0.008823658339679241, + 0.010292613878846169, + 0.012227597646415234, + 0.010011689737439156, + 0.016359003260731697, + 0.009393792599439621, + 0.007050958462059498, + 0.012647118419408798, + 0.0028667552396655083, + 0.002740795724093914, + 0.006036543752998114, + 0.00211272481828928, + 0.0038032231386750937, + 0.003395988140255213, + 0.0028857518918812275, + 0.002591231372207403, + 0.005757973995059729, + 0.00507954228669405, + 0.007139645051211119, + 0.010026361793279648, + 0.003714810824021697, + 0.0033220683690160513, + 0.006209015380591154, + 0.004544614348560572, + 0.017890123650431633, + 0.01157974824309349, + 0.02463448792695999, + 0.026281189173460007, + 0.010713860392570496, + 0.01714460179209709, + 0.020767977461218834, + 0.01766739971935749, + 0.03891259804368019, + 0.020123714581131935, + 0.06977462768554688, + 0.030489536002278328, + 0.02955971285700798, + 0.020076433196663857, + 0.030149061232805252, + 0.08615179359912872, + 0.0035016650799661875, + 0.0021874296944588423, + 0.0036104382015764713, + 0.004939727485179901, + 0.0012745002750307322, + 0.003180071013048291, + 0.003779900958761573, + 0.0029212883673608303, + 0.010980555787682533, + 0.025712378323078156, + 0.008240398950874805, + 0.009890326298773289, + 0.014264336787164211, + 0.01305400487035513, + 0.01705111190676689, + 0.011092170141637325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.4288671314716339 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023285744711756706 + ], + "opposite_logprob_tokens": [ + -3.7732856273651123 + ], + "logprob_reward": -0.023285744711756706, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023285744711756706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40206339955329895, + 0.07332557439804077, + 0.043603792786598206, + 0.01580223999917507, + 0.04862232133746147, + 0.04337514191865921, + 0.07359984517097473, + 0.057268381118774414, + 0.1151285320520401, + 0.1272108405828476 + ], + "credit_map": [ + 0.03304475173354149, + 0.062296122312545776, + 0.051915813237428665, + 0.04692364111542702, + 0.08240210264921188, + 0.047208793461322784, + 0.0446951724588871, + 0.033576950430870056, + 0.006325689610093832, + 0.004592263605445623, + 0.010145384818315506, + 0.01244322583079338, + 0.004465845413506031, + 0.005720439366996288, + 0.011461044661700726, + 0.01817167177796364, + 0.007196780759841204, + 0.00476149981841445, + 0.006249779835343361, + 0.006236198358237743, + 0.004903953522443771, + 0.005067463032901287, + 0.004709798377007246, + 0.004478314891457558, + 0.0015565267531201243, + 0.002021154621616006, + 0.0024806384462863207, + 0.002047954825684428, + 0.0024609528481960297, + 0.001611623796634376, + 0.002084017964079976, + 0.0015393702778965235, + 0.005553094204515219, + 0.004000493790954351, + 0.005917458329349756, + 0.004311435855925083, + 0.009749958291649818, + 0.008260509930551052, + 0.0052648503333330154, + 0.005564517807215452, + 0.004998141899704933, + 0.0049804700538516045, + 0.005798969883471727, + 0.007734205573797226, + 0.005538243800401688, + 0.004598380997776985, + 0.004400692880153656, + 0.00532603170722723, + 0.008988280780613422, + 0.007226680871099234, + 0.011574014090001583, + 0.011889181099832058, + 0.005894229281693697, + 0.010143334046006203, + 0.00987572968006134, + 0.008008395321667194, + 0.006468543782830238, + 0.003888193052262068, + 0.013766949065029621, + 0.003340276191011071, + 0.004547901451587677, + 0.0029408081900328398, + 0.005491313058882952, + 0.01682439260184765, + 0.0271415114402771, + 0.013255860656499863, + 0.014675669372081757, + 0.010005283169448376, + 0.008078315295279026, + 0.01416704896837473, + 0.018672451376914978, + 0.009132384322583675, + 0.012332619167864323, + 0.03132067248225212, + 0.008622799068689346, + 0.010888342745602131, + 0.01712953671813011, + 0.014581154100596905, + 0.021244140341877937, + 0.011091564781963825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.023285744711756706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This low-angle, tight medium shot, captured with documentary-style realism and a slight handheld shake, immerses the viewer in the waters of a clear, shallow river. The central focus is two majestic sockeye salmon, their bodies a brilliant, pristine crimson, their heads a contrasting metallic emerald, as they fiercely thrash against the current. Their powerful bodies collide and twist, churning the water into a white foam and sending arcs of shimmering droplets into the air. Their hooked jaws open and close with aggressive speed, revealing the white interiors of their mouths in a desperate fight for spawning rights. The entire scene is rendered in a stark, desaturated color grading, muting the greens and browns of the out-of-focus riverbank and the multicolored pebbles visible through the clear water. This deliberate artistic choice allows the visceral, saturated crimson of the salmon's bodies to stand out with a shocking intensity against the muted surroundings, creating an atmosphere of tension and drama focused purely on the brutal beauty of their fierce, instinctive struggle.", + "seed": 28621, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_15/index_9191_seed_28621_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15143038593232633, + "vlm_logprob": -0.15143038593232633, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.15143038593232633, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014448895119130611 + ], + "opposite_logprob_tokens": [ + -4.639449119567871 + ], + "logprob_reward": -0.014448895119130611, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.014448895119130611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11883250623941422, + 0.0785922110080719, + 0.0651489868760109, + 0.030612407252192497, + 0.08144661784172058, + 0.20558501780033112, + 0.055040448904037476, + 0.15637041628360748, + 0.07875571399927139, + 0.1296156346797943 + ], + "credit_map": [ + 0.014754334464669228, + 0.006236327346414328, + 0.023019468411803246, + 0.007123473100364208, + 0.014002710580825806, + 0.012289082631468773, + 0.015293272212147713, + 0.02611383982002735, + 0.007095694076269865, + 0.004260010551661253, + 0.007525162305682898, + 0.005472780670970678, + 0.009310939349234104, + 0.012842793948948383, + 0.012386247515678406, + 0.01969858445227146, + 0.0053849187679588795, + 0.005364453885704279, + 0.007470368407666683, + 0.017722200602293015, + 0.005869411397725344, + 0.0072064693085849285, + 0.010135007090866566, + 0.005996158812195063, + 0.0026535706128925085, + 0.004290077835321426, + 0.0032151350751519203, + 0.004739878699183464, + 0.004377431236207485, + 0.004282061010599136, + 0.003479191567748785, + 0.0035750584211200476, + 0.004924186505377293, + 0.003927652724087238, + 0.01125177089124918, + 0.029903888702392578, + 0.009217769838869572, + 0.013922427780926228, + 0.0040551479905843735, + 0.004243770614266396, + 0.020124228671193123, + 0.005380652844905853, + 0.07376879453659058, + 0.02251792885363102, + 0.010064306668937206, + 0.06000930070877075, + 0.005736778490245342, + 0.007983037270605564, + 0.004458138253539801, + 0.004510919563472271, + 0.006886118091642857, + 0.00700240908190608, + 0.008202283643186092, + 0.006331183481961489, + 0.010217641480267048, + 0.007431753445416689, + 0.014027891680598259, + 0.01102521177381277, + 0.022567927837371826, + 0.0235271193087101, + 0.0196651853621006, + 0.023381179198622704, + 0.01769358664751053, + 0.024482322856783867, + 0.007124859374016523, + 0.010374236851930618, + 0.013097901828587055, + 0.010742475278675556, + 0.009923002682626247, + 0.010258003138005733, + 0.010485708713531494, + 0.0067495242692530155, + 0.007117057219147682, + 0.006576563697308302, + 0.02508729137480259, + 0.043145839124917984, + 0.008197513408958912, + 0.018832892179489136, + 0.012451237067580223, + 0.008207234553992748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.014448895119130611 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204963102936745 + ], + "opposite_logprob_tokens": [ + -2.812049627304077 + ], + "logprob_reward": -0.06204963102936745, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204963102936745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1650286316871643, + 0.16644686460494995, + 0.1312083899974823, + 0.05553044006228447, + 0.07012731581926346, + 0.2531982362270355, + 0.03650892525911331, + 0.03324243053793907, + 0.033150382339954376, + 0.055558327585458755 + ], + "credit_map": [ + 0.020748533308506012, + 0.01791343092918396, + 0.018280021846294403, + 0.021044112741947174, + 0.01728026196360588, + 0.017362462356686592, + 0.020338892936706543, + 0.03206094726920128, + 0.014430281706154346, + 0.008560314774513245, + 0.011021282523870468, + 0.011454298160970211, + 0.02812814712524414, + 0.033356159925460815, + 0.01950559765100479, + 0.03999080881476402, + 0.009945420548319817, + 0.009648524224758148, + 0.011510515585541725, + 0.018751895055174828, + 0.014521793462336063, + 0.01598614640533924, + 0.028232520446181297, + 0.0226115845143795, + 0.007645081263035536, + 0.008394081145524979, + 0.007188622839748859, + 0.00745546817779541, + 0.005522033199667931, + 0.007847592234611511, + 0.005898054223507643, + 0.0055795107036828995, + 0.004485911689698696, + 0.004562876652926207, + 0.009983993135392666, + 0.020880648866295815, + 0.008817882277071476, + 0.011447870172560215, + 0.003880544099956751, + 0.006067594978958368, + 0.025036275386810303, + 0.006550353951752186, + 0.0910714790225029, + 0.026940563693642616, + 0.012617960572242737, + 0.0751064196228981, + 0.006282974500209093, + 0.009592208079993725, + 0.004053726326674223, + 0.003132304409518838, + 0.007215001154690981, + 0.005083299707621336, + 0.004073597025126219, + 0.00518580200150609, + 0.0038525667041540146, + 0.003912630956619978, + 0.003285854822024703, + 0.0029323033522814512, + 0.003374444553628564, + 0.004576782695949078, + 0.004726245068013668, + 0.005014821887016296, + 0.003362029790878296, + 0.005969948600977659, + 0.004926508292555809, + 0.0032283952459692955, + 0.007830843329429626, + 0.0049735731445252895, + 0.0026504432316869497, + 0.0031142153311520815, + 0.003362567164003849, + 0.0030638400930911303, + 0.0041741859167814255, + 0.0029741351027041674, + 0.016598019748926163, + 0.014058888889849186, + 0.003214640310034156, + 0.005762000102549791, + 0.005298715550452471, + 0.003477748716250062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.06204963102936745 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192714355885983 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192714355885983, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192714355885983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1249130517244339, + 0.06970430165529251, + 0.0916837602853775, + 0.038652434945106506, + 0.05604713410139084, + 0.1229783147573471, + 0.076994888484478, + 0.11883612722158432, + 0.05168069526553154, + 0.24850931763648987 + ], + "credit_map": [ + 0.014578702859580517, + 0.008953015320003033, + 0.02082253433763981, + 0.009509327821433544, + 0.014141402207314968, + 0.014142154715955257, + 0.014462854713201523, + 0.028303047642111778, + 0.008368601091206074, + 0.004805485252290964, + 0.006001630332320929, + 0.00623079901561141, + 0.01094598788768053, + 0.0125917699187994, + 0.008017673157155514, + 0.012742352671921253, + 0.009129033423960209, + 0.011874285526573658, + 0.01290496438741684, + 0.011063873767852783, + 0.013410920277237892, + 0.010453570634126663, + 0.01136330422013998, + 0.011483801528811455, + 0.004220173228532076, + 0.006844500545412302, + 0.006118841003626585, + 0.006238764617592096, + 0.004113222938030958, + 0.0040220157243311405, + 0.0030600624158978462, + 0.004034852143377066, + 0.004686845000833273, + 0.003420367604121566, + 0.012918424792587757, + 0.01674201525747776, + 0.00515703484416008, + 0.005318700801581144, + 0.0032680814620107412, + 0.00453566387295723, + 0.010015764273703098, + 0.004295541904866695, + 0.03715279698371887, + 0.0134867699816823, + 0.007571352645754814, + 0.03455951437354088, + 0.00885138101875782, + 0.007045190315693617, + 0.007114811800420284, + 0.0051132384687662125, + 0.011263564229011536, + 0.011219516396522522, + 0.00980850588530302, + 0.007711084559559822, + 0.01742343232035637, + 0.007340726908296347, + 0.01585952378809452, + 0.014290086925029755, + 0.014488349668681622, + 0.017258821055293083, + 0.011714575812220573, + 0.014377374202013016, + 0.012217619456350803, + 0.01862976886332035, + 0.005091099999845028, + 0.004377976059913635, + 0.009926361963152885, + 0.011664195917546749, + 0.005084330216050148, + 0.006270667538046837, + 0.005111561622470617, + 0.0041544996201992035, + 0.021651361137628555, + 0.011991079896688461, + 0.05699513107538223, + 0.02336117811501026, + 0.017263201996684074, + 0.07383564114570618, + 0.025737151503562927, + 0.01767454855144024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.014192714355885983 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327689290046692 + ], + "opposite_logprob_tokens": [ + -0.7577689290046692 + ], + "logprob_reward": -0.6327689290046692, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327689290046692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07009921222925186, + 0.10091570764780045, + 0.08419475704431534, + 0.05533600226044655, + 0.042419422417879105, + 0.2269875556230545, + 0.020732609555125237, + 0.016906853765249252, + 0.03449329361319542, + 0.34791457653045654 + ], + "credit_map": [ + 0.010096454992890358, + 0.005508834030479193, + 0.010695499368011951, + 0.006744163576513529, + 0.008367435075342655, + 0.007295746356248856, + 0.008687039837241173, + 0.012704038992524147, + 0.008746964856982231, + 0.004715483170002699, + 0.0053751179948449135, + 0.007450954057276249, + 0.02099587209522724, + 0.022196415811777115, + 0.010751069523394108, + 0.020683830603957176, + 0.014428787864744663, + 0.006445399485528469, + 0.012288491241633892, + 0.014347061514854431, + 0.00719319935888052, + 0.0070180743932724, + 0.01281615998595953, + 0.009657574817538261, + 0.005498558282852173, + 0.004765693098306656, + 0.006972221191972494, + 0.009108031168580055, + 0.005748833995312452, + 0.009869011119008064, + 0.0071123819798231125, + 0.006261272821575403, + 0.003867879044264555, + 0.0032942034304142, + 0.007296708878129721, + 0.010576101019978523, + 0.00465266453102231, + 0.005500224884599447, + 0.0028583535458892584, + 0.004373287316411734, + 0.023599086329340935, + 0.006090350914746523, + 0.0781758651137352, + 0.02647210843861103, + 0.009927001781761646, + 0.06503992527723312, + 0.0066178832203149796, + 0.011065339669585228, + 0.002895382698625326, + 0.0015531819080933928, + 0.0026064568664878607, + 0.0035578575916588306, + 0.0027219513431191444, + 0.0025072582066059113, + 0.0022097998298704624, + 0.002680719830095768, + 0.002456798916682601, + 0.0015449600759893656, + 0.002242530696094036, + 0.0032836091704666615, + 0.0017241507302969694, + 0.001973863923922181, + 0.002050698036327958, + 0.0016302429139614105, + 0.0059528411366045475, + 0.004096963442862034, + 0.00510060228407383, + 0.004653514362871647, + 0.002870454452931881, + 0.00383228762075305, + 0.0034276375081390142, + 0.004558991640806198, + 0.029243817552924156, + 0.010429078713059425, + 0.11446770280599594, + 0.05531204119324684, + 0.020769383758306503, + 0.0756680816411972, + 0.02195107191801071, + 0.020073389634490013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.6327689290046692 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two sockeye salmon visible as the central focus of the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the salmon have bodies colored brilliant crimson and heads colored metallic emerald?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the salmon thrashing, colliding, and twisting in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered with a stark, desaturated color grading that mutes the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle, tight medium shot with a slight handheld shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033691760152578354 + ], + "opposite_logprob_tokens": [ + -3.408691644668579 + ], + "logprob_reward": -0.033691760152578354, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033691760152578354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08487686514854431, + 0.06040436774492264, + 0.18080578744411469, + 0.0826517641544342, + 0.10217884182929993, + 0.18042972683906555, + 0.05906045064330101, + 0.05041629076004028, + 0.06961418688297272, + 0.129561647772789 + ], + "credit_map": [ + 0.012476163916289806, + 0.0073689054697752, + 0.017041046172380447, + 0.008821663446724415, + 0.008353369310498238, + 0.007767257280647755, + 0.007949902676045895, + 0.015098572708666325, + 0.004784933757036924, + 0.004096939228475094, + 0.006325600203126669, + 0.006206280551850796, + 0.007920465432107449, + 0.01074788998812437, + 0.008391053415834904, + 0.011931215412914753, + 0.022293400019407272, + 0.01609581708908081, + 0.013846902176737785, + 0.018115030601620674, + 0.019591011106967926, + 0.02278798446059227, + 0.034288499504327774, + 0.033787183463573456, + 0.007052879314869642, + 0.00864019151777029, + 0.009879905730485916, + 0.013802463188767433, + 0.006085660308599472, + 0.018199007958173752, + 0.009905464015901089, + 0.009086206555366516, + 0.006385388318449259, + 0.006172823254019022, + 0.014679556712508202, + 0.030893757939338684, + 0.01190818753093481, + 0.01647217944264412, + 0.005287428852170706, + 0.01037953794002533, + 0.019238410517573357, + 0.005716234445571899, + 0.06000658869743347, + 0.020908070728182793, + 0.00822662003338337, + 0.048788297921419144, + 0.00748484255746007, + 0.010060686618089676, + 0.005338997580111027, + 0.005291203502565622, + 0.00814969465136528, + 0.008026687428355217, + 0.007243372034281492, + 0.005843709222972393, + 0.012555993162095547, + 0.006610802840441465, + 0.006051917094737291, + 0.0036698386538773775, + 0.006612833589315414, + 0.010306844487786293, + 0.005722661502659321, + 0.006141036283224821, + 0.004548093304038048, + 0.007363075390458107, + 0.008637542836368084, + 0.007116853725165129, + 0.014945224858820438, + 0.007876615971326828, + 0.005906106438487768, + 0.01003782544285059, + 0.008886558935046196, + 0.0062074726447463036, + 0.00789138674736023, + 0.008126452565193176, + 0.019481373950839043, + 0.03759948909282684, + 0.009739630855619907, + 0.024121152237057686, + 0.013055222108960152, + 0.009546958841383457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 9 + ], + "selected_reward": -0.033691760152578354 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_3/index_tensor([1965, 1965])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_3/index_tensor([1965, 1965])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8099a5b8890a00568fbe8d5af79eb80e74cf3636 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_3/index_tensor([1965, 1965])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a shot steeped in the spirit of a competitive sports documentary, captured from a canted overhead angle, the camera trembles slightly with the force of the impact, capturing the decisive moment of two muscular, sweat-drenched opponents clashing fiercely on the green artificial turf of an indoor tumble court. Clad in red and blue sleeveless uniforms, their colors darkened by sweat, their taut muscles visible under the bright stadium lights. The giant transparent tumble ball encasing their bodies is covered in scratches from the intense movement, its surface reflecting glaring light. In this moment of battle for the league championship, they lower their center of gravity and violently collide with each other. The two massive plastic spheres collide with a resounding thud in the center of the frame, instantly deforming and compressing under the immense force of the impact, emitting a dull yet powerful sound. The energy of the collision shook the sweat off their tense cheeks and tangled hair, drawing tiny arcs in the air. Their clenched teeth expressions were a mixture of pain and determination, and their eyes were burning with an unyielding fighting spirit. The whole picture was full of primitive power and the tense atmosphere of competitive sports, witnessing this explosive peak showdown.", + "seed": 13484, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_3/index_1965_seed_13484_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01995312049984932, + "vlm_logprob": -0.01995312049984932, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.01995312049984932, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010601157322525978 + ], + "opposite_logprob_tokens": [ + -4.760601043701172 + ], + "logprob_reward": -0.010601157322525978, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010601157322525978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18203993141651154, + 0.11068225651979446, + 0.05527671054005623, + 0.12140853703022003, + 0.211001917719841, + 0.037091467529535294, + 0.08018482476472855, + 0.08411049097776413, + 0.03737681359052658, + 0.0808270126581192 + ], + "credit_map": [ + 0.015255973674356937, + 0.018419712781906128, + 0.011256199330091476, + 0.04545833542943001, + 0.014704924076795578, + 0.012772633694112301, + 0.009406685829162598, + 0.0547654815018177, + 0.011447321623563766, + 0.009959418326616287, + 0.008652003481984138, + 0.020042840391397476, + 0.019096631556749344, + 0.016764871776103973, + 0.0119835389778018, + 0.012735635973513126, + 0.005129652097821236, + 0.006753831170499325, + 0.005087462719529867, + 0.009750979021191597, + 0.006794375367462635, + 0.007759186904877424, + 0.0061999219469726086, + 0.007801305502653122, + 0.008328696712851524, + 0.011288406327366829, + 0.012335172854363918, + 0.005999570246785879, + 0.008360465988516808, + 0.01496868021786213, + 0.05037674307823181, + 0.009750797413289547, + 0.012394345365464687, + 0.024135569110512733, + 0.011580833233892918, + 0.024463294073939323, + 0.029368558898568153, + 0.02430260181427002, + 0.05594344437122345, + 0.02881328947842121, + 0.0044767712242901325, + 0.0041531044989824295, + 0.004404018167406321, + 0.005952654406428337, + 0.007527959533035755, + 0.003780883038416505, + 0.0038020031061023474, + 0.0029940742533653975, + 0.007314058020710945, + 0.008902136236429214, + 0.006238531321287155, + 0.011029877699911594, + 0.011370568536221981, + 0.0109616843983531, + 0.01295787189155817, + 0.011410095728933811, + 0.010754263959825039, + 0.010232370346784592, + 0.005825836677104235, + 0.013816273771226406, + 0.00694003002718091, + 0.015657693147659302, + 0.01099640317261219, + 0.009887628257274628, + 0.0033879680559039116, + 0.006305038463324308, + 0.004688379820436239, + 0.006355905439704657, + 0.005133455619215965, + 0.004314830992370844, + 0.004460815340280533, + 0.0027304256800562143, + 0.008714377880096436, + 0.010697889141738415, + 0.008598522283136845, + 0.01580396108329296, + 0.009056244045495987, + 0.007134740240871906, + 0.015319308266043663, + 0.005501977633684874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.010601157322525978 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020597022026777267 + ], + "opposite_logprob_tokens": [ + -3.895596981048584 + ], + "logprob_reward": -0.020597022026777267, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020597022026777267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.519013524055481, + 0.046159226447343826, + 0.042869918048381805, + 0.06966252624988556, + 0.050876155495643616, + 0.04014851152896881, + 0.05413318797945976, + 0.049977611750364304, + 0.03674466907978058, + 0.0904146358370781 + ], + "credit_map": [ + 0.03419074788689613, + 0.11559576541185379, + 0.023114629089832306, + 0.0768677294254303, + 0.04454577714204788, + 0.04872008040547371, + 0.024351364001631737, + 0.15162746608257294, + 0.00607183575630188, + 0.008418156765401363, + 0.004361685831099749, + 0.006649651564657688, + 0.008174888789653778, + 0.004382377956062555, + 0.003300236538052559, + 0.004800400696694851, + 0.006260753143578768, + 0.00424503767862916, + 0.004044109955430031, + 0.00882797408849001, + 0.004533383063971996, + 0.005645631812512875, + 0.005066827405244112, + 0.004246206022799015, + 0.008537954650819302, + 0.005669581703841686, + 0.00852241925895214, + 0.004780993796885014, + 0.005520674865692854, + 0.007980176247656345, + 0.021735792979598045, + 0.006914936937391758, + 0.003792751580476761, + 0.007125820033252239, + 0.00416450435295701, + 0.00821116752922535, + 0.007890450768172741, + 0.004393137525767088, + 0.007846405729651451, + 0.007451921701431274, + 0.004262477625161409, + 0.006974054034799337, + 0.00487202825024724, + 0.006252789869904518, + 0.007625976111739874, + 0.00335218315012753, + 0.003164484165608883, + 0.003644522512331605, + 0.006267473567277193, + 0.00929961260408163, + 0.004347067791968584, + 0.006388240959495306, + 0.007968530990183353, + 0.006283584516495466, + 0.007872750982642174, + 0.005705930758267641, + 0.0041748229414224625, + 0.004963665269315243, + 0.0038736455608159304, + 0.010948770679533482, + 0.0036352460738271475, + 0.008959731087088585, + 0.007525988854467869, + 0.005895746871829033, + 0.0031403852626681328, + 0.00644583348184824, + 0.004580419510602951, + 0.006606326904147863, + 0.004358822945505381, + 0.004384240601211786, + 0.0046533639542758465, + 0.002575281308963895, + 0.008718003518879414, + 0.00972004234790802, + 0.00807894766330719, + 0.017773956060409546, + 0.012896295636892319, + 0.00938720814883709, + 0.017145201563835144, + 0.0066949836909770966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.020597022026777267 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193509429693222 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193509429693222, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193509429693222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19689230620861053, + 0.04313437268137932, + 0.11345555633306503, + 0.1297096461057663, + 0.09968189150094986, + 0.04286548122763634, + 0.06402183324098587, + 0.11636943370103836, + 0.046793390065431595, + 0.14707602560520172 + ], + "credit_map": [ + 0.017422061413526535, + 0.050512783229351044, + 0.012174074538052082, + 0.044121529906988144, + 0.023322174325585365, + 0.019378192722797394, + 0.01302014384418726, + 0.016941383481025696, + 0.0050867171958088875, + 0.004394329618662596, + 0.004146963823586702, + 0.009269108064472675, + 0.006701344158500433, + 0.005386310629546642, + 0.00404046755284071, + 0.00410913722589612, + 0.013674154877662659, + 0.011827684938907623, + 0.008223491720855236, + 0.017072414979338646, + 0.01397030521184206, + 0.02192782424390316, + 0.011666499078273773, + 0.015093198046088219, + 0.009109735488891602, + 0.01073641236871481, + 0.01346918661147356, + 0.0068325623869895935, + 0.008085575886070728, + 0.014953053556382656, + 0.04956522956490517, + 0.016957910731434822, + 0.006090947426855564, + 0.014588605612516403, + 0.006510948296636343, + 0.010450715199112892, + 0.016744131222367287, + 0.011627311818301678, + 0.02168874815106392, + 0.011980497278273106, + 0.005307723768055439, + 0.006403315346688032, + 0.0045524174347519875, + 0.005849019158631563, + 0.009326196275651455, + 0.005348912440240383, + 0.0026480169035494328, + 0.0034298815298825502, + 0.0040681506507098675, + 0.005402383394539356, + 0.00454828841611743, + 0.011364146135747433, + 0.007788359187543392, + 0.006359932012856007, + 0.015100483782589436, + 0.009390098974108696, + 0.015332666225731373, + 0.012922213412821293, + 0.009219429455697536, + 0.013390209525823593, + 0.00949727464467287, + 0.02516009472310543, + 0.017583796754479408, + 0.013263758271932602, + 0.004711909685283899, + 0.006599580403417349, + 0.004873823840171099, + 0.014765024185180664, + 0.0029972149059176445, + 0.0053613437339663506, + 0.00503492821007967, + 0.002449572319164872, + 0.015643704682588577, + 0.015250731259584427, + 0.009733219631016254, + 0.028601255267858505, + 0.024943167343735695, + 0.014876254834234715, + 0.025535432621836662, + 0.012492286041378975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.018193509429693222 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581956952810287 + ], + "opposite_logprob_tokens": [ + -3.8955819606781006 + ], + "logprob_reward": -0.020581956952810287, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581956952810287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26631370186805725, + 0.0700172483921051, + 0.09621605277061462, + 0.08129838854074478, + 0.07312291860580444, + 0.08243463933467865, + 0.0658729299902916, + 0.10385019332170486, + 0.05223023146390915, + 0.10864366590976715 + ], + "credit_map": [ + 0.02710966393351555, + 0.04292411729693413, + 0.020306207239627838, + 0.07681037485599518, + 0.027774840593338013, + 0.02175281010568142, + 0.02022280916571617, + 0.0294128879904747, + 0.009607097133994102, + 0.01100218202918768, + 0.0083006601780653, + 0.011210153810679913, + 0.007849717512726784, + 0.007800890598446131, + 0.005506972316652536, + 0.00873958133161068, + 0.007582715712487698, + 0.010819089598953724, + 0.008338495157659054, + 0.017237376421689987, + 0.011534231714904308, + 0.017417486757040024, + 0.009706653654575348, + 0.01358000561594963, + 0.006512846332043409, + 0.006869778968393803, + 0.013982178643345833, + 0.007570542860776186, + 0.006627343595027924, + 0.011271325871348381, + 0.021872643381357193, + 0.006591738201677799, + 0.007462626788765192, + 0.008602692745625973, + 0.0068595013581216335, + 0.012405319139361382, + 0.009656504727900028, + 0.007872196845710278, + 0.008182445541024208, + 0.01208164170384407, + 0.010909895412623882, + 0.009238858707249165, + 0.013874425552785397, + 0.018289675936102867, + 0.010741498321294785, + 0.0063570342026650906, + 0.006096869707107544, + 0.006926385220140219, + 0.006000556517392397, + 0.0057786148972809315, + 0.003951883874833584, + 0.006700980011373758, + 0.007723396178334951, + 0.014688797295093536, + 0.013626311905682087, + 0.007402391172945499, + 0.0097018051892519, + 0.010781463235616684, + 0.007605595514178276, + 0.023590587079524994, + 0.008806916885077953, + 0.02013656124472618, + 0.012902461923658848, + 0.01032480038702488, + 0.004617879632860422, + 0.011515729129314423, + 0.00494755245745182, + 0.010366784408688545, + 0.005327852908521891, + 0.0064333351328969, + 0.0057084038853645325, + 0.0033127020578831434, + 0.012777752242982388, + 0.012935719452798367, + 0.011494224891066551, + 0.016970621421933174, + 0.016491759568452835, + 0.011243176646530628, + 0.016468508169054985, + 0.010261905379593372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.020581956952810287 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029791956767439842 + ], + "opposite_logprob_tokens": [ + -3.529792070388794 + ], + "logprob_reward": -0.029791956767439842, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029791956767439842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25286102294921875, + 0.06070884317159653, + 0.06416063755750656, + 0.0918799340724945, + 0.11035424470901489, + 0.04905213415622711, + 0.05964258685708046, + 0.1391390711069107, + 0.02988564409315586, + 0.14231592416763306 + ], + "credit_map": [ + 0.017482656985521317, + 0.07980997860431671, + 0.01507993321865797, + 0.05409691110253334, + 0.013057038187980652, + 0.022696509957313538, + 0.015194687061011791, + 0.03544329106807709, + 0.0062514725141227245, + 0.006871243007481098, + 0.006841940805315971, + 0.011932551860809326, + 0.009156554006040096, + 0.006569221615791321, + 0.006185987498611212, + 0.006899866741150618, + 0.0072438521310687065, + 0.007785320281982422, + 0.005961444694548845, + 0.012598776258528233, + 0.006422955077141523, + 0.010427558794617653, + 0.00532741192728281, + 0.008393317461013794, + 0.009295686148107052, + 0.0062841991893947124, + 0.01449503656476736, + 0.007327293045818806, + 0.007489068899303675, + 0.011282680556178093, + 0.02577243186533451, + 0.009933529421687126, + 0.006272138096392155, + 0.015446550212800503, + 0.006054086145013571, + 0.012166686356067657, + 0.018344834446907043, + 0.013216007500886917, + 0.023493900895118713, + 0.015360033139586449, + 0.0074478196911513805, + 0.007451780140399933, + 0.005610537249594927, + 0.007101575843989849, + 0.00835843849927187, + 0.00601375475525856, + 0.003226433414965868, + 0.0038417878095060587, + 0.004718163516372442, + 0.005283907987177372, + 0.004476488567888737, + 0.007926232181489468, + 0.007402503862977028, + 0.007817383855581284, + 0.01110791601240635, + 0.01090998575091362, + 0.015254036523401737, + 0.015831662341952324, + 0.009666578844189644, + 0.024013139307498932, + 0.012542420998215675, + 0.027555353939533234, + 0.01981780119240284, + 0.014458073303103447, + 0.0023929260205477476, + 0.004216092638671398, + 0.0030624684877693653, + 0.0049537490122020245, + 0.004125663079321384, + 0.003364650532603264, + 0.0054683685302734375, + 0.0023017218336462975, + 0.01638919673860073, + 0.0156551580876112, + 0.010436199605464935, + 0.022503450512886047, + 0.032220933586359024, + 0.01781633496284485, + 0.018884902819991112, + 0.008409732021391392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.029791956767439842 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a shot steeped in the spirit of a competitive sports documentary, captured from a canted overhead angle, the camera trembles slightly with the force of the impact, capturing the decisive moment of two muscular, sweat-drenched opponents clashing fiercely on the green artificial turf of an indoor tumble court. Clad in red and blue sleeveless uniforms, their colors darkened by sweat, their taut muscles visible under the bright stadium lights. The giant transparent tumble ball encasing their bodies is covered in scratches from the intense movement, its surface reflecting glaring light. In this moment of battle for the league championship, they lower their center of gravity and violently collide with each other. The two massive plastic spheres collide with a resounding thud in the center of the frame, instantly deforming and compressing under the immense force of the impact, emitting a dull yet powerful sound. The energy of the collision shook the sweat off their tense cheeks and tangled hair, drawing tiny arcs in the air. Their clenched teeth expressions were a mixture of pain and determination, and their eyes were burning with an unyielding fighting spirit. The whole picture was full of primitive power and the tense atmosphere of competitive sports, witnessing this explosive peak showdown.", + "seed": 13485, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_3/index_1965_seed_13485_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028170623630285264, + "vlm_logprob": -0.028170623630285264, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.028170623630285264, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022115280851721764 + ], + "opposite_logprob_tokens": [ + -3.8971152305603027 + ], + "logprob_reward": -0.022115280851721764, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.022115280851721764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14814136922359467, + 0.021962467581033707, + 0.08395592123270035, + 0.12303745746612549, + 0.036320723593235016, + 0.12615974247455597, + 0.04024825245141983, + 0.06943034380674362, + 0.24175366759300232, + 0.10899003595113754 + ], + "credit_map": [ + 0.018552865833044052, + 0.021905919536948204, + 0.010587901808321476, + 0.019938480108976364, + 0.013885699212551117, + 0.025284217670559883, + 0.01765563152730465, + 0.020330648869276047, + 0.002990857232362032, + 0.00311406422406435, + 0.0019980587530881166, + 0.003017704701051116, + 0.00264544109813869, + 0.0033597543369978666, + 0.0023555634543299675, + 0.0024810221511870623, + 0.010164856910705566, + 0.007418714463710785, + 0.009736618027091026, + 0.008926107548177242, + 0.010205451399087906, + 0.014577867463231087, + 0.011673365719616413, + 0.01125293131917715, + 0.028238270431756973, + 0.01393308024853468, + 0.013178330846130848, + 0.010722744278609753, + 0.017239604145288467, + 0.014555838890373707, + 0.012586038559675217, + 0.012583533301949501, + 0.0027688422705978155, + 0.004344494082033634, + 0.006025148555636406, + 0.004537608940154314, + 0.004836181178689003, + 0.0036015661899000406, + 0.006266283802688122, + 0.003940596245229244, + 0.024437149986624718, + 0.012501641176640987, + 0.02259952388703823, + 0.020456310361623764, + 0.01664889045059681, + 0.009723679162561893, + 0.011704208329319954, + 0.008088331669569016, + 0.0032630902715027332, + 0.004091382492333651, + 0.004993950016796589, + 0.005428334232419729, + 0.006546431221067905, + 0.0064633553847670555, + 0.005308705847710371, + 0.004152999259531498, + 0.005939384922385216, + 0.010677251033484936, + 0.009972416795790195, + 0.007690832018852234, + 0.0065877847373485565, + 0.010158591903746128, + 0.009742921218276024, + 0.008661152794957161, + 0.018432864919304848, + 0.024750640615820885, + 0.025114504620432854, + 0.03399602696299553, + 0.0287200678139925, + 0.040990374982357025, + 0.029060447588562965, + 0.04068871960043907, + 0.010819051414728165, + 0.01629810407757759, + 0.012579482980072498, + 0.016552329063415527, + 0.015909967944025993, + 0.010307523421943188, + 0.013774077408015728, + 0.012749486602842808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.022115280851721764 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811566159129143 + ], + "opposite_logprob_tokens": [ + -3.2881157398223877 + ], + "logprob_reward": -0.03811566159129143, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811566159129143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13753043115139008, + 0.04316532239317894, + 0.2126518189907074, + 0.3091888129711151, + 0.06143622100353241, + 0.04868898540735245, + 0.026800500229001045, + 0.01858201064169407, + 0.08251018822193146, + 0.05944573134183884 + ], + "credit_map": [ + 0.016621660441160202, + 0.021145541220903397, + 0.0094370748847723, + 0.01740032434463501, + 0.01581188477575779, + 0.02372574433684349, + 0.016110878437757492, + 0.017277320846915245, + 0.007752046920359135, + 0.004949228838086128, + 0.00516957463696599, + 0.006285136565566063, + 0.00466753076761961, + 0.005019941832870245, + 0.003996210638433695, + 0.005325651727616787, + 0.026476820930838585, + 0.018468763679265976, + 0.022517012432217598, + 0.014080232940614223, + 0.028980139642953873, + 0.038657721132040024, + 0.03180084377527237, + 0.03167028725147247, + 0.07507055997848511, + 0.03364581614732742, + 0.03232811018824577, + 0.027838587760925293, + 0.03949551284313202, + 0.03371347859501839, + 0.03502252697944641, + 0.0320742167532444, + 0.0051559144631028175, + 0.0077944565564394, + 0.01084174495190382, + 0.0058518387377262115, + 0.00872285757213831, + 0.006648282054811716, + 0.008428620174527168, + 0.00799250602722168, + 0.009460380300879478, + 0.007032013963907957, + 0.007282787933945656, + 0.0059355199337005615, + 0.005324495956301689, + 0.003574684029445052, + 0.005410702433437109, + 0.004668399691581726, + 0.002485620090737939, + 0.0027323307003825903, + 0.003639284521341324, + 0.004265099298208952, + 0.004820424132049084, + 0.004382219631224871, + 0.0024181348271667957, + 0.0020573867950588465, + 0.002125651342794299, + 0.002825361443683505, + 0.0024845239240676165, + 0.0025397588033229113, + 0.0017860359512269497, + 0.002363219391554594, + 0.0023199727293103933, + 0.0021374870557338, + 0.0074695716612041, + 0.01095073577016592, + 0.017043787986040115, + 0.009400551207363605, + 0.01101688016206026, + 0.010263622738420963, + 0.007879585959017277, + 0.008485455997288227, + 0.0043924832716584206, + 0.005869519896805286, + 0.006117156241089106, + 0.005348603706806898, + 0.00839686207473278, + 0.008438020944595337, + 0.009921841323375702, + 0.010961242951452732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.03811566159129143 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058732882142067 + ], + "opposite_logprob_tokens": [ + -3.89558744430542 + ], + "logprob_reward": -0.02058732882142067, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058732882142067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3303523659706116, + 0.03393396735191345, + 0.08196157962083817, + 0.061390504240989685, + 0.0810798704624176, + 0.05924058333039284, + 0.04103878512978554, + 0.058654285967350006, + 0.11978346854448318, + 0.13256460428237915 + ], + "credit_map": [ + 0.044039126485586166, + 0.04352483153343201, + 0.024538088589906693, + 0.04612359777092934, + 0.030930479988455772, + 0.05956457182765007, + 0.037722811102867126, + 0.043908875435590744, + 0.003989080898463726, + 0.00549877667799592, + 0.0032372395507991314, + 0.0049802870489656925, + 0.0038858260959386826, + 0.003549582324922085, + 0.005833516828715801, + 0.0029596586246043444, + 0.010741219855844975, + 0.006662593688815832, + 0.009786022827029228, + 0.00861282180994749, + 0.01077517494559288, + 0.012524641118943691, + 0.012541812844574451, + 0.010317289270460606, + 0.008169530890882015, + 0.006507884711027145, + 0.005384290590882301, + 0.006194214336574078, + 0.011480957269668579, + 0.007525257300585508, + 0.009123883210122585, + 0.007004487328231335, + 0.012633858248591423, + 0.012539812363684177, + 0.013601219281554222, + 0.006368356756865978, + 0.012267275713384151, + 0.007179096341133118, + 0.00972837395966053, + 0.00676187826320529, + 0.008069033734500408, + 0.00604998180642724, + 0.008650077506899834, + 0.008729602210223675, + 0.007424827665090561, + 0.005760022439062595, + 0.008326960727572441, + 0.006230077240616083, + 0.00388109078630805, + 0.00485088350251317, + 0.004655675962567329, + 0.00435365317389369, + 0.007407134398818016, + 0.00731236208230257, + 0.0046295905485749245, + 0.0039483942091465, + 0.006315157283097506, + 0.00769518269225955, + 0.007943219505250454, + 0.0064447373151779175, + 0.005689783953130245, + 0.00881822220981121, + 0.007581120822578669, + 0.008166860789060593, + 0.008891698904335499, + 0.016560036689043045, + 0.020067282021045685, + 0.015683872625231743, + 0.014534072950482368, + 0.01776500977575779, + 0.00957919005304575, + 0.016702299937605858, + 0.010620162822306156, + 0.015064735896885395, + 0.017344694584608078, + 0.017201559618115425, + 0.021930010989308357, + 0.016764890402555466, + 0.012896752916276455, + 0.020741784945130348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.02058732882142067 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346659287810326 + ], + "opposite_logprob_tokens": [ + -3.6513466835021973 + ], + "logprob_reward": -0.026346659287810326, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346659287810326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15948358178138733, + 0.05907663702964783, + 0.0952741876244545, + 0.09881926327943802, + 0.08874955773353577, + 0.11243082582950592, + 0.05792878568172455, + 0.08465573936700821, + 0.16873309016227722, + 0.07484836131334305 + ], + "credit_map": [ + 0.018415916711091995, + 0.023690277710556984, + 0.010835775174200535, + 0.020351339131593704, + 0.02099533937871456, + 0.027681557461619377, + 0.01738838478922844, + 0.02012498490512371, + 0.006995086558163166, + 0.006578236352652311, + 0.005916557274758816, + 0.009063837118446827, + 0.007666273042559624, + 0.005690468475222588, + 0.0071396129205822945, + 0.01002656389027834, + 0.014461425133049488, + 0.007725434377789497, + 0.007701822090893984, + 0.008311175741255283, + 0.01806744560599327, + 0.015231102705001831, + 0.011161664500832558, + 0.012614114210009575, + 0.020163746550679207, + 0.009455958381295204, + 0.01142791472375393, + 0.010013503022491932, + 0.015262789092957973, + 0.009517326019704342, + 0.01001481432467699, + 0.012963205575942993, + 0.0088358698412776, + 0.01102416031062603, + 0.011418076232075691, + 0.01083874050527811, + 0.01909688673913479, + 0.011099816299974918, + 0.007917981594800949, + 0.008518017828464508, + 0.02192082814872265, + 0.012274768203496933, + 0.01364974770694971, + 0.020048394799232483, + 0.016836173832416534, + 0.010401764884591103, + 0.01086881197988987, + 0.0064303199760615826, + 0.005020670127123594, + 0.004868595860898495, + 0.005197061691433191, + 0.00866408459842205, + 0.007943786680698395, + 0.011561496183276176, + 0.008814294822514057, + 0.00585878686979413, + 0.009943219833076, + 0.025610467419028282, + 0.010947092436254025, + 0.0074026696383953094, + 0.008953629992902279, + 0.006053471937775612, + 0.00827049557119608, + 0.007474682293832302, + 0.012635939754545689, + 0.021384797990322113, + 0.030946293845772743, + 0.026449790224432945, + 0.02407977171242237, + 0.019442584365606308, + 0.01795164681971073, + 0.015842251479625702, + 0.007917366921901703, + 0.009353585541248322, + 0.009892505593597889, + 0.006471886765211821, + 0.014120272360742092, + 0.008770109154284, + 0.008720321580767632, + 0.00960230827331543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.026346659287810326 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368818759918213 + ], + "opposite_logprob_tokens": [ + -3.4086880683898926 + ], + "logprob_reward": -0.03368818759918213, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368818759918213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28186365962028503, + 0.0669298991560936, + 0.08405353128910065, + 0.056753210723400116, + 0.05625429004430771, + 0.09077873080968857, + 0.06530774384737015, + 0.05215049162507057, + 0.12889137864112854, + 0.11701704561710358 + ], + "credit_map": [ + 0.03240836411714554, + 0.04245443269610405, + 0.020331881940364838, + 0.03326185420155525, + 0.034049440175294876, + 0.053398702293634415, + 0.029267095029354095, + 0.03669189289212227, + 0.010544677264988422, + 0.00968150608241558, + 0.006526880897581577, + 0.010340609587728977, + 0.009181280620396137, + 0.007258516736328602, + 0.006534811109304428, + 0.006861617788672447, + 0.008591406047344208, + 0.006713289767503738, + 0.011720306240022182, + 0.01097393874078989, + 0.008664642460644245, + 0.01377285085618496, + 0.011525589041411877, + 0.012091509066522121, + 0.00666958699002862, + 0.006674740929156542, + 0.004985395818948746, + 0.005874562077224255, + 0.011548555456101894, + 0.0058600748889148235, + 0.008572405204176903, + 0.006567888427525759, + 0.007049307227134705, + 0.005454824771732092, + 0.00901044998317957, + 0.004747095517814159, + 0.010541684925556183, + 0.005572578404098749, + 0.008379059843719006, + 0.005499287974089384, + 0.01798468455672264, + 0.011739060282707214, + 0.014444130472838879, + 0.010728586465120316, + 0.010760968551039696, + 0.00640270859003067, + 0.010199886746704578, + 0.008518705144524574, + 0.00609807251021266, + 0.005740390624850988, + 0.006939735729247332, + 0.010072681121528149, + 0.013177408836781979, + 0.009022804908454418, + 0.008134279400110245, + 0.00612237211316824, + 0.006912080571055412, + 0.007352733984589577, + 0.005886690691113472, + 0.005227549932897091, + 0.005064540542662144, + 0.007052937988191843, + 0.006971156224608421, + 0.007682800758630037, + 0.01019959431141615, + 0.018856069073081017, + 0.022098006680607796, + 0.014451839961111546, + 0.017093585804104805, + 0.018146764487028122, + 0.010669219307601452, + 0.017376292496919632, + 0.010121305473148823, + 0.016043785959482193, + 0.01615278050303459, + 0.01456309761852026, + 0.02205684222280979, + 0.011923350393772125, + 0.01363535039126873, + 0.012520527467131615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.03368818759918213 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a shot steeped in the spirit of a competitive sports documentary, captured from a canted overhead angle, the camera trembles slightly with the force of the impact, capturing the decisive moment of two muscular, sweat-drenched opponents clashing fiercely on the green artificial turf of an indoor tumble court. Clad in red and blue sleeveless uniforms, their colors darkened by sweat, their taut muscles visible under the bright stadium lights. The giant transparent tumble ball encasing their bodies is covered in scratches from the intense movement, its surface reflecting glaring light. In this moment of battle for the league championship, they lower their center of gravity and violently collide with each other. The two massive plastic spheres collide with a resounding thud in the center of the frame, instantly deforming and compressing under the immense force of the impact, emitting a dull yet powerful sound. The energy of the collision shook the sweat off their tense cheeks and tangled hair, drawing tiny arcs in the air. Their clenched teeth expressions were a mixture of pain and determination, and their eyes were burning with an unyielding fighting spirit. The whole picture was full of primitive power and the tense atmosphere of competitive sports, witnessing this explosive peak showdown.", + "seed": 13486, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_3/index_1965_seed_13486_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.029865378513932228, + "vlm_logprob": -0.029865378513932228, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.029865378513932228, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03134011849761009 + ], + "opposite_logprob_tokens": [ + -3.5313401222229004 + ], + "logprob_reward": -0.03134011849761009, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03134011849761009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23398761451244354, + 0.03244252875447273, + 0.10159171372652054, + 0.012078506872057915, + 0.046731628477573395, + 0.014286844059824944, + 0.060810476541519165, + 0.05129950866103172, + 0.40510690212249756, + 0.041664283722639084 + ], + "credit_map": [ + 0.023192226886749268, + 0.02557218074798584, + 0.04733886569738388, + 0.04253396391868591, + 0.026723714545369148, + 0.021984735503792763, + 0.022439436987042427, + 0.024202486500144005, + 0.002506406744942069, + 0.006174413021653891, + 0.002872997894883156, + 0.004468739032745361, + 0.007764809299260378, + 0.00356528302654624, + 0.002162823686376214, + 0.0029270555824041367, + 0.011613396927714348, + 0.014766966924071312, + 0.010238087736070156, + 0.02358202263712883, + 0.00949614867568016, + 0.012537690810859203, + 0.01221233606338501, + 0.007145062554627657, + 0.0016123688546940684, + 0.0015339544042944908, + 0.0018454157980158925, + 0.0019254556391388178, + 0.0013386833015829325, + 0.0014230752130970359, + 0.0014305731747299433, + 0.0009689805447123945, + 0.0023558372631669044, + 0.006467437371611595, + 0.004744763486087322, + 0.0033134964760392904, + 0.007203207816928625, + 0.01452304795384407, + 0.00404398376122117, + 0.004079856909811497, + 0.0017573428340256214, + 0.0016882496420294046, + 0.0013894448056817055, + 0.0015727314166724682, + 0.0018801606493070722, + 0.0012362675042822957, + 0.0016261483542621136, + 0.0031364995520561934, + 0.007308581843972206, + 0.005514791700989008, + 0.008821757510304451, + 0.008090094663202763, + 0.0097592081874609, + 0.004732740111649036, + 0.00914859026670456, + 0.007434715982526541, + 0.005657175090163946, + 0.004958914127200842, + 0.005432874895632267, + 0.007039804942905903, + 0.011956114321947098, + 0.005669513717293739, + 0.005134996026754379, + 0.005450120195746422, + 0.07563334703445435, + 0.13336186110973358, + 0.03106563910841942, + 0.021642595529556274, + 0.04713838919997215, + 0.04269885644316673, + 0.03360471501946449, + 0.01996150054037571, + 0.004395836498588324, + 0.004362207371741533, + 0.005664496682584286, + 0.0089564248919487, + 0.003963020164519548, + 0.005334834568202496, + 0.0054236180149018764, + 0.0035638483241200447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.03134011849761009 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054981376975774765 + ], + "opposite_logprob_tokens": [ + -2.9299814701080322 + ], + "logprob_reward": -0.054981376975774765, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054981376975774765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4245067238807678, + 0.03506552428007126, + 0.048142142593860626, + 0.02748960442841053, + 0.043576423078775406, + 0.024064205586910248, + 0.04044448584318161, + 0.08631118386983871, + 0.21312327682971954, + 0.057276442646980286 + ], + "credit_map": [ + 0.041207004338502884, + 0.025056099519133568, + 0.08407064527273178, + 0.06973613053560257, + 0.04497562721371651, + 0.044856395572423935, + 0.05800451710820198, + 0.056600287556648254, + 0.0026197077240794897, + 0.006610632408410311, + 0.0022087080869823694, + 0.005560070741921663, + 0.009263502433896065, + 0.003427932970225811, + 0.002297172788530588, + 0.003077796893194318, + 0.006330353673547506, + 0.0073741283267736435, + 0.006830364465713501, + 0.00778395077213645, + 0.004730353131890297, + 0.006016268394887447, + 0.004920273553580046, + 0.004156450275331736, + 0.004279915243387222, + 0.004907538183033466, + 0.003174716839566827, + 0.004201812203973532, + 0.0027210828848183155, + 0.002935395808890462, + 0.00290456204675138, + 0.002364582847803831, + 0.001975136809051037, + 0.010775445960462093, + 0.0037259499076753855, + 0.004726540297269821, + 0.007292544469237328, + 0.007565079256892204, + 0.003914821892976761, + 0.0036009035538882017, + 0.004307668190449476, + 0.0033954507671296597, + 0.0032425401732325554, + 0.0020631810184568167, + 0.0032450647559016943, + 0.0019612915348261595, + 0.0022604181431233883, + 0.0035885917022824287, + 0.006117445882409811, + 0.005213368684053421, + 0.006622973829507828, + 0.00455505819991231, + 0.005233624018728733, + 0.0033868197351694107, + 0.004856147337704897, + 0.0044590504840016365, + 0.00882402528077364, + 0.008426319807767868, + 0.00859859585762024, + 0.011262092739343643, + 0.01636331155896187, + 0.010554032400250435, + 0.014699413441121578, + 0.007583393715322018, + 0.03361152857542038, + 0.05908871442079544, + 0.01835860311985016, + 0.01249639131128788, + 0.027991903945803642, + 0.02440820261836052, + 0.024374425411224365, + 0.012793506495654583, + 0.005614630877971649, + 0.004889273550361395, + 0.008936953730881214, + 0.006062912754714489, + 0.012828408740460873, + 0.006339693907648325, + 0.006800557021051645, + 0.005804012529551983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.054981376975774765 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016077518463134766 + ], + "opposite_logprob_tokens": [ + -4.141077518463135 + ], + "logprob_reward": -0.016077518463134766, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016077518463134766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5138859748840332, + 0.057150427252054214, + 0.045144230127334595, + 0.018304038792848587, + 0.13016892969608307, + 0.025658050552010536, + 0.037262190133333206, + 0.07269137352705002, + 0.03590068966150284, + 0.06383409351110458 + ], + "credit_map": [ + 0.04836144298315048, + 0.03233910724520683, + 0.06558971852064133, + 0.06615151464939117, + 0.09002532809972763, + 0.06749211251735687, + 0.0638740286231041, + 0.0800526961684227, + 0.005867179948836565, + 0.008655095472931862, + 0.005151593126356602, + 0.008100928738713264, + 0.014241469092667103, + 0.006049917545169592, + 0.003939188085496426, + 0.005145057570189238, + 0.005822788458317518, + 0.006509387865662575, + 0.007454464677721262, + 0.008091243915259838, + 0.004181989468634129, + 0.0034932189155369997, + 0.005285922903567553, + 0.004305215086787939, + 0.002259632805362344, + 0.002749891020357609, + 0.0029517931398004293, + 0.0028758624102920294, + 0.002017364837229252, + 0.0018593072891235352, + 0.0018335479544475675, + 0.0017566407332196832, + 0.006537226494401693, + 0.03693900257349014, + 0.008403538726270199, + 0.008220167830586433, + 0.04168197512626648, + 0.010792122222483158, + 0.010112190619111061, + 0.007482700049877167, + 0.004382031038403511, + 0.003600028110668063, + 0.0029321748297661543, + 0.0026712235994637012, + 0.003920161165297031, + 0.0029987450689077377, + 0.002411541296169162, + 0.002742145676165819, + 0.006443087477236986, + 0.004279290325939655, + 0.004941944032907486, + 0.003573161317035556, + 0.003818430006504059, + 0.003635497996583581, + 0.00524924136698246, + 0.005321540404111147, + 0.007102327886968851, + 0.006888416595757008, + 0.007807263173162937, + 0.0071802218444645405, + 0.01847565919160843, + 0.007933206856250763, + 0.009640628471970558, + 0.00766365323215723, + 0.0042996178381145, + 0.005395802669227123, + 0.0036120114382356405, + 0.003214410739019513, + 0.006622267886996269, + 0.006530378013849258, + 0.0036078246776014566, + 0.0026183747686445713, + 0.004690553527325392, + 0.007342251483350992, + 0.006279631517827511, + 0.005706300493329763, + 0.017787931486964226, + 0.010395358316600323, + 0.006758286152034998, + 0.004873781464993954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.016077518463134766 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026352928951382637 + ], + "opposite_logprob_tokens": [ + -3.651352882385254 + ], + "logprob_reward": -0.026352928951382637, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026352928951382637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26706042885780334, + 0.0711873322725296, + 0.061617787927389145, + 0.13408443331718445, + 0.1535780131816864, + 0.07936185598373413, + 0.040576446801424026, + 0.05949756130576134, + 0.08215919137001038, + 0.05087701976299286 + ], + "credit_map": [ + 0.02513381466269493, + 0.017933327704668045, + 0.03736761212348938, + 0.04286446422338486, + 0.060959283262491226, + 0.027567142620682716, + 0.026370037347078323, + 0.028864707797765732, + 0.006510946899652481, + 0.011109554208815098, + 0.00870379339903593, + 0.012835921719670296, + 0.011808126233518124, + 0.007311370223760605, + 0.006353907752782106, + 0.006553709041327238, + 0.007794529665261507, + 0.01044594869017601, + 0.006524130702018738, + 0.011049080640077591, + 0.005426910240203142, + 0.0054555111564695835, + 0.010009896010160446, + 0.004911778029054403, + 0.016312571242451668, + 0.02135009877383709, + 0.017580047249794006, + 0.021602265536785126, + 0.013605955988168716, + 0.013336540199816227, + 0.017386849969625473, + 0.012910088524222374, + 0.008562872186303139, + 0.03979221731424332, + 0.01516928244382143, + 0.01536373421549797, + 0.04744807258248329, + 0.012707797810435295, + 0.008341864682734013, + 0.0061921472661197186, + 0.012941373512148857, + 0.011462117545306683, + 0.008475844748318195, + 0.006338865030556917, + 0.014456239528954029, + 0.008561449125409126, + 0.00892175454646349, + 0.0082042021676898, + 0.008319341577589512, + 0.004084320273250341, + 0.006803265772759914, + 0.004082008730620146, + 0.004730219021439552, + 0.0046007647179067135, + 0.0028724779840558767, + 0.00508404616266489, + 0.007117010187357664, + 0.005427342373877764, + 0.007483671419322491, + 0.0066320449113845825, + 0.014305385760962963, + 0.005306526552885771, + 0.007958807982504368, + 0.005266763735562563, + 0.01393961813300848, + 0.02915971726179123, + 0.007355059497058392, + 0.00399744464084506, + 0.008437245152890682, + 0.008590246550738811, + 0.006128896959125996, + 0.00455095898360014, + 0.004494227934628725, + 0.006610686890780926, + 0.007709469646215439, + 0.007302646059542894, + 0.006310580763965845, + 0.006724060047417879, + 0.005516715347766876, + 0.00620862515643239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.026352928951382637 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057494968175888 + ], + "opposite_logprob_tokens": [ + -3.8955750465393066 + ], + "logprob_reward": -0.02057494968175888, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057494968175888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44555312395095825, + 0.03611317649483681, + 0.04298041760921478, + 0.026058591902256012, + 0.13019149005413055, + 0.05262714996933937, + 0.037935297936201096, + 0.07639557123184204, + 0.12065906077623367, + 0.03148616477847099 + ], + "credit_map": [ + 0.04756173864006996, + 0.0355277881026268, + 0.04716271162033081, + 0.05349033325910568, + 0.07482493668794632, + 0.04860466346144676, + 0.05250432342290878, + 0.08587661385536194, + 0.004400145728141069, + 0.004666867665946484, + 0.00400296738371253, + 0.008630075491964817, + 0.004077108111232519, + 0.004077104385942221, + 0.0030054207891225815, + 0.003253484610468149, + 0.005092784762382507, + 0.008085086941719055, + 0.005343408323824406, + 0.006584152113646269, + 0.004273956175893545, + 0.003790448419749737, + 0.005146111827343702, + 0.004664470907300711, + 0.003263747552409768, + 0.004980361554771662, + 0.0026442089583724737, + 0.003343665972352028, + 0.0028308772016316652, + 0.0029462215024977922, + 0.003417408559471369, + 0.002632100135087967, + 0.007380537688732147, + 0.012235835194587708, + 0.013385633938014507, + 0.013701662421226501, + 0.021942786872386932, + 0.036431457847356796, + 0.011976197361946106, + 0.013137378729879856, + 0.007051306776702404, + 0.007871572859585285, + 0.00556817464530468, + 0.005308673717081547, + 0.009210166521370411, + 0.005479217041283846, + 0.0059725260362029076, + 0.006165511906147003, + 0.00633997330442071, + 0.004623988177627325, + 0.004432348068803549, + 0.0037070997059345245, + 0.004576988518238068, + 0.004004479385912418, + 0.0045472243800759315, + 0.005703199654817581, + 0.008372673764824867, + 0.007030734792351723, + 0.008147351443767548, + 0.008396713063120842, + 0.01851164549589157, + 0.007851812057197094, + 0.01018451526761055, + 0.007900121621787548, + 0.02092180959880352, + 0.04025895148515701, + 0.011216232553124428, + 0.00680098682641983, + 0.014134781435132027, + 0.012385138310492039, + 0.008764821104705334, + 0.006176340859383345, + 0.0034091973211616278, + 0.00405963696539402, + 0.005341877695173025, + 0.003925527445971966, + 0.002850464778020978, + 0.0045006973668932915, + 0.004026663023978472, + 0.003372100181877613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.02057494968175888 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a shot steeped in the spirit of a competitive sports documentary, captured from a canted overhead angle, the camera trembles slightly with the force of the impact, capturing the decisive moment of two muscular, sweat-drenched opponents clashing fiercely on the green artificial turf of an indoor tumble court. Clad in red and blue sleeveless uniforms, their colors darkened by sweat, their taut muscles visible under the bright stadium lights. The giant transparent tumble ball encasing their bodies is covered in scratches from the intense movement, its surface reflecting glaring light. In this moment of battle for the league championship, they lower their center of gravity and violently collide with each other. The two massive plastic spheres collide with a resounding thud in the center of the frame, instantly deforming and compressing under the immense force of the impact, emitting a dull yet powerful sound. The energy of the collision shook the sweat off their tense cheeks and tangled hair, drawing tiny arcs in the air. Their clenched teeth expressions were a mixture of pain and determination, and their eyes were burning with an unyielding fighting spirit. The whole picture was full of primitive power and the tense atmosphere of competitive sports, witnessing this explosive peak showdown.", + "seed": 13487, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_3/index_1965_seed_13487_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03511698748916388, + "vlm_logprob": -0.03511698748916388, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.03511698748916388, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010983367450535297 + ], + "opposite_logprob_tokens": [ + -4.635983467102051 + ], + "logprob_reward": -0.010983367450535297, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010983367450535297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09276802092790604, + 0.03449276462197304, + 0.07767321169376373, + 0.0609746128320694, + 0.1421428918838501, + 0.29078802466392517, + 0.06474515795707703, + 0.054312944412231445, + 0.1467774510383606, + 0.03532491996884346 + ], + "credit_map": [ + 0.00876007042825222, + 0.01182042807340622, + 0.017918288707733154, + 0.00992958340793848, + 0.008731735870242119, + 0.007998252287507057, + 0.009399532340466976, + 0.01821012608706951, + 0.005496676545590162, + 0.0033920658752322197, + 0.0038551895413547754, + 0.004318431485444307, + 0.00593955535441637, + 0.003767609130591154, + 0.004517488181591034, + 0.0032057466451078653, + 0.005895128007978201, + 0.010824433527886868, + 0.007700099144130945, + 0.008940167725086212, + 0.014368337579071522, + 0.008219883777201176, + 0.009656449779868126, + 0.01206870935857296, + 0.005657732021063566, + 0.011200330220162868, + 0.005974769592285156, + 0.0063683995977044106, + 0.007823447696864605, + 0.006922339554876089, + 0.008822436444461346, + 0.008205153048038483, + 0.012186198495328426, + 0.022373540326952934, + 0.0266820527613163, + 0.02568131498992443, + 0.019703643396496773, + 0.014666206203401089, + 0.011239724233746529, + 0.009610201232135296, + 0.021567784249782562, + 0.02106897160410881, + 0.015986012294888496, + 0.01878558099269867, + 0.07660111784934998, + 0.022391684353351593, + 0.02186272107064724, + 0.09252413362264633, + 0.006118710618466139, + 0.006849432829767466, + 0.0067103151232004166, + 0.00749678211286664, + 0.011402621865272522, + 0.006894343066960573, + 0.011733200401067734, + 0.0075397491455078125, + 0.004668161738663912, + 0.0030570444650948048, + 0.004456663504242897, + 0.011026189662516117, + 0.004450948443263769, + 0.005000667180866003, + 0.005804845131933689, + 0.015848420560359955, + 0.016838790848851204, + 0.015012121759355068, + 0.02770315296947956, + 0.02263270877301693, + 0.014106259681284428, + 0.018979918211698532, + 0.014958643354475498, + 0.0165458582341671, + 0.002999149030074477, + 0.0031970213167369366, + 0.0031014487612992525, + 0.0036207903176546097, + 0.00491740508005023, + 0.005125831812620163, + 0.005066010635346174, + 0.007297258824110031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.010983367450535297 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811784088611603 + ], + "opposite_logprob_tokens": [ + -3.2881178855895996 + ], + "logprob_reward": -0.03811784088611603, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811784088611603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2782832980155945, + 0.027101824060082436, + 0.07354867458343506, + 0.20384986698627472, + 0.13479562103748322, + 0.05640469491481781, + 0.060791920870542526, + 0.08074648678302765, + 0.03771420195698738, + 0.046763353049755096 + ], + "credit_map": [ + 0.031020408496260643, + 0.01547927875071764, + 0.10133057087659836, + 0.017371956259012222, + 0.014115710742771626, + 0.01781829074025154, + 0.015943236649036407, + 0.06520383059978485, + 0.003342932555824518, + 0.0031672415789216757, + 0.002799524459987879, + 0.003885024692863226, + 0.0038331809919327497, + 0.002976366551592946, + 0.00441315583884716, + 0.0026843969244509935, + 0.009263617917895317, + 0.00930241122841835, + 0.011868026107549667, + 0.007348848972469568, + 0.008371986448764801, + 0.009101664647459984, + 0.011023078113794327, + 0.007269044406712055, + 0.024508774280548096, + 0.019936393946409225, + 0.01957896165549755, + 0.02386552095413208, + 0.02567962557077408, + 0.033606357872486115, + 0.02908504009246826, + 0.027589181438088417, + 0.008506398648023605, + 0.045027658343315125, + 0.04148479551076889, + 0.010845049284398556, + 0.0102538438513875, + 0.0059255254454910755, + 0.003708893433213234, + 0.00904345978051424, + 0.0040983157232403755, + 0.005836444906890392, + 0.004255781415849924, + 0.005610451567918062, + 0.01105474028736353, + 0.007048182189464569, + 0.0068815057165920734, + 0.011619272641837597, + 0.006671880837529898, + 0.006421681959182024, + 0.005972657352685928, + 0.006137266289442778, + 0.01103883981704712, + 0.007081144023686647, + 0.011354177258908749, + 0.006114274263381958, + 0.00946125853806734, + 0.005243635270744562, + 0.005532660987228155, + 0.012309938669204712, + 0.0062292288057506084, + 0.00645463028922677, + 0.006936362944543362, + 0.028578773140907288, + 0.002756060566753149, + 0.0032380479387938976, + 0.005004170350730419, + 0.0045962161384522915, + 0.0059952870942652225, + 0.00474513741210103, + 0.005501832813024521, + 0.005877448711544275, + 0.004583118017762899, + 0.0047238664701581, + 0.005683181341737509, + 0.005648865830153227, + 0.005831949412822723, + 0.00562977883964777, + 0.006998772732913494, + 0.007663820404559374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.03811784088611603 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038104645907878876 + ], + "opposite_logprob_tokens": [ + -3.28810453414917 + ], + "logprob_reward": -0.038104645907878876, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038104645907878876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27020809054374695, + 0.06730201840400696, + 0.16027584671974182, + 0.05551949515938759, + 0.1556931585073471, + 0.0783734992146492, + 0.04984727129340172, + 0.05726237595081329, + 0.06344455480575562, + 0.04207373782992363 + ], + "credit_map": [ + 0.02417016215622425, + 0.03425014764070511, + 0.07379329204559326, + 0.029834147542715073, + 0.024384066462516785, + 0.017257122322916985, + 0.026249082759022713, + 0.04027007520198822, + 0.008258501067757607, + 0.006183126475661993, + 0.0087595134973526, + 0.011873090639710426, + 0.00913841463625431, + 0.0058326381258666515, + 0.011466301046311855, + 0.005790421739220619, + 0.013451559469103813, + 0.018072927370667458, + 0.025970902293920517, + 0.02687462791800499, + 0.022222956642508507, + 0.01502961665391922, + 0.024028724059462547, + 0.014624524861574173, + 0.003806379623711109, + 0.009287822991609573, + 0.00652313232421875, + 0.006643905304372311, + 0.010902341455221176, + 0.006339884363114834, + 0.0060864840634167194, + 0.005929544568061829, + 0.008390596136450768, + 0.05238133296370506, + 0.041132114827632904, + 0.01886959932744503, + 0.011145771481096745, + 0.008096350356936455, + 0.0045286002568900585, + 0.011148780584335327, + 0.004791258834302425, + 0.008052237331867218, + 0.00766341295093298, + 0.009223777800798416, + 0.013387213461101055, + 0.006525793578475714, + 0.021010521799325943, + 0.007719275541603565, + 0.0034903562627732754, + 0.004042115993797779, + 0.004061287734657526, + 0.007160694804042578, + 0.008324164897203445, + 0.005125859752297401, + 0.011180696077644825, + 0.00646209018304944, + 0.0043195690959692, + 0.004420658107846975, + 0.004335439298301935, + 0.013051378540694714, + 0.006340597756206989, + 0.006360903382301331, + 0.006360465195029974, + 0.012073355726897717, + 0.00588561687618494, + 0.006377906072884798, + 0.009834952652454376, + 0.008679937571287155, + 0.009536546654999256, + 0.007713869214057922, + 0.008018478751182556, + 0.007397246081382036, + 0.0025874366983771324, + 0.004274618811905384, + 0.0038907628040760756, + 0.007445263676345348, + 0.007645999547094107, + 0.007791402284055948, + 0.004500017035752535, + 0.0039382330141961575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.038104645907878876 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634805254638195 + ], + "opposite_logprob_tokens": [ + -3.651348114013672 + ], + "logprob_reward": -0.02634805254638195, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634805254638195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23523077368736267, + 0.035293616354465485, + 0.04539496824145317, + 0.036564912647008896, + 0.19442228972911835, + 0.11320513486862183, + 0.08528320491313934, + 0.1435503363609314, + 0.05996677652001381, + 0.051087938249111176 + ], + "credit_map": [ + 0.01969711109995842, + 0.025026433169841766, + 0.051613397896289825, + 0.02539009600877762, + 0.016466518864035606, + 0.015926698222756386, + 0.04708898067474365, + 0.03402154520153999, + 0.004588523879647255, + 0.004192329477518797, + 0.004977297969162464, + 0.005877898074686527, + 0.005537820979952812, + 0.002690706169232726, + 0.004235847853124142, + 0.003193190088495612, + 0.004180106334388256, + 0.004613490309566259, + 0.0068567791022360325, + 0.006669826805591583, + 0.006230168044567108, + 0.004697526805102825, + 0.005028229672461748, + 0.007118841167539358, + 0.0029590879566967487, + 0.00423699663951993, + 0.0049748667515814304, + 0.003931059502065182, + 0.008427610620856285, + 0.004347481764853001, + 0.004119720309972763, + 0.003568087238818407, + 0.01387436967343092, + 0.05734119191765785, + 0.06516590714454651, + 0.013470468111336231, + 0.015434646047651768, + 0.011100102216005325, + 0.006249739322811365, + 0.01178587693721056, + 0.00883518811315298, + 0.008033178746700287, + 0.010960202664136887, + 0.007321412209421396, + 0.030228475108742714, + 0.0118704279884696, + 0.010778293013572693, + 0.025177953764796257, + 0.0077715530060231686, + 0.00876371469348669, + 0.0082982387393713, + 0.011467350646853447, + 0.015372471883893013, + 0.010429359041154385, + 0.01526123471558094, + 0.007919288240373135, + 0.02167351171374321, + 0.007890199311077595, + 0.009006801061332226, + 0.017281675711274147, + 0.011244256980717182, + 0.012328620068728924, + 0.009692632593214512, + 0.05443266034126282, + 0.005279687233269215, + 0.005965429358184338, + 0.005438958294689655, + 0.006369693670421839, + 0.010842152871191502, + 0.007161625660955906, + 0.009951086714863777, + 0.008958142250776291, + 0.0036761148367077112, + 0.0062345582991838455, + 0.005956477019935846, + 0.006925660651177168, + 0.00736123276874423, + 0.00594295933842659, + 0.008136664517223835, + 0.006854269653558731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.02634805254638195 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two opponents wearing red and blue sleeveless uniforms visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on green artificial turf inside an indoor tumble court?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a giant transparent tumble ball encasing the two opponents visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a canted overhead angle and tremble slightly during the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two massive plastic spheres collide violently in the center of the frame, deforming and compressing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203103065490723 + ], + "opposite_logprob_tokens": [ + -2.8120310306549072 + ], + "logprob_reward": -0.06203103065490723, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203103065490723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.125385582447052, + 0.039351917803287506, + 0.11321543157100677, + 0.08377455174922943, + 0.15332403779029846, + 0.2089492380619049, + 0.055277928709983826, + 0.08510541915893555, + 0.05952421575784683, + 0.07609163224697113 + ], + "credit_map": [ + 0.011486775241792202, + 0.011346620507538319, + 0.020910492166876793, + 0.009197711013257504, + 0.01450872328132391, + 0.011912389658391476, + 0.028734780848026276, + 0.017288096249103546, + 0.004761500284075737, + 0.0041569918394088745, + 0.004625135567039251, + 0.004387334920465946, + 0.005811966955661774, + 0.004921916406601667, + 0.006439915392547846, + 0.00424715643748641, + 0.008640289306640625, + 0.011500157415866852, + 0.016173899173736572, + 0.02851009927690029, + 0.012741224840283394, + 0.009098184294998646, + 0.018454302102327347, + 0.008097268640995026, + 0.00806457083672285, + 0.015925418585538864, + 0.009535768069326878, + 0.008559510111808777, + 0.013658523559570312, + 0.008602488785982132, + 0.009155319072306156, + 0.010272961109876633, + 0.008382130414247513, + 0.051429588347673416, + 0.04905545711517334, + 0.012322074733674526, + 0.010854654014110565, + 0.006172205787152052, + 0.004421651363372803, + 0.010686282999813557, + 0.016917478293180466, + 0.01613384671509266, + 0.016698362305760384, + 0.019571011886000633, + 0.040461216121912, + 0.014780866913497448, + 0.02950260415673256, + 0.05488385632634163, + 0.004574751015752554, + 0.005254280753433704, + 0.004342143423855305, + 0.005981795489788055, + 0.013574556447565556, + 0.005456536076962948, + 0.0068618133664131165, + 0.009232052601873875, + 0.009626861661672592, + 0.009499476291239262, + 0.007164597976952791, + 0.011930220760405064, + 0.014781969599425793, + 0.011291926726698875, + 0.012282715179026127, + 0.008527655154466629, + 0.005204382818192244, + 0.007048081606626511, + 0.005969167221337557, + 0.009068900719285011, + 0.01045161858201027, + 0.007532997988164425, + 0.006685551255941391, + 0.007563516963273287, + 0.004514091648161411, + 0.006397672463208437, + 0.006038829684257507, + 0.010918905027210712, + 0.012753932736814022, + 0.012892905622720718, + 0.009496183134615421, + 0.013079117983579636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.06203103065490723 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_4/index_tensor([2141, 2141])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_4/index_tensor([2141, 2141])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..35ae6f83cecd0e6885b03c8a39e26d246c83ed2e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_4/index_tensor([2141, 2141])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion shot, shot from an upward tilt, captured in a striking, high-contrast black-and-white aesthetic, captures a strikingly muscular female athlete at the peak of her leap during a beach volleyball match. Her body, tied in a tight ponytail, arches in the air, the tense muscles of her back, shoulders, and arms clearly visible in the strong sunlight. Clad in a dark sports bikini, her facial expression is one of intense concentration, eyes narrowed, jaw set, her complete focus focused on the volleyball hovering directly in front of her. Her right arm is fully extended back, palm open, poised for a thunderous smash. The intense, direct sunlight streaming from above and side sculpts her form into a dynamic composition of deep shadows and dazzling whites, while each drop of sweat dripping from her skin shimmers with a shimmering light. Below her, fine grains of sand, sent flying by her explosive leap, freeze in mid-air, while the upper edge of the volleyball net forms the foreground. The entire scene is imbued with a raw sense of power and frozen dynamic beauty, conveying a tense, focused, and explosive competitive atmosphere.", + "seed": 11930, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_4/index_2141_seed_11930_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.6406133342534304, + "vlm_logprob": -0.6406133342534304, + "vlm_yes_no_margin": 1.0500000357627868, + "vlm_token_credit": -0.6406133342534304, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4531049132347107 + ], + "opposite_logprob_tokens": [ + -1.2031049728393555 + ], + "logprob_reward": -0.4531049132347107, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.4531049132347107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12394126504659653, + 0.19646324217319489, + 0.05455430969595909, + 0.12201378494501114, + 0.05355066806077957, + 0.09901493042707443, + 0.05880385637283325, + 0.1472589373588562, + 0.09253748506307602, + 0.05186154320836067 + ], + "credit_map": [ + 0.018320519477128983, + 0.02479742094874382, + 0.01898830570280552, + 0.015786074101924896, + 0.012715848162770271, + 0.01054997369647026, + 0.010181635618209839, + 0.01260148361325264, + 0.025133175775408745, + 0.014880886301398277, + 0.03620964288711548, + 0.022114446386694908, + 0.028096506372094154, + 0.022084694355726242, + 0.028786085546016693, + 0.019157804548740387, + 0.008694215677678585, + 0.0047117541544139385, + 0.0075604841113090515, + 0.006499539595097303, + 0.00563596049323678, + 0.00458049401640892, + 0.006907049100846052, + 0.009964811615645885, + 0.01508760079741478, + 0.02369983121752739, + 0.015027775429189205, + 0.022870900109410286, + 0.019045235589146614, + 0.008528145961463451, + 0.007624490186572075, + 0.01012980006635189, + 0.00935275387018919, + 0.008429021574556828, + 0.006083392538130283, + 0.00935366190969944, + 0.004030271898955107, + 0.0053467885591089725, + 0.005443718284368515, + 0.0055110570974648, + 0.012495264410972595, + 0.024271462112665176, + 0.010739964433014393, + 0.011179538443684578, + 0.011370131745934486, + 0.01209949143230915, + 0.007494695018976927, + 0.009364383295178413, + 0.009456288069486618, + 0.012217270210385323, + 0.005038229748606682, + 0.004792234860360622, + 0.005170359741896391, + 0.01056220568716526, + 0.0058272420428693295, + 0.005740028340369463, + 0.03748779743909836, + 0.017590148374438286, + 0.01297284197062254, + 0.00992276705801487, + 0.01114039309322834, + 0.041377656161785126, + 0.008110260590910912, + 0.008657067082822323, + 0.01137461792677641, + 0.011189364828169346, + 0.014239962212741375, + 0.012271752581000328, + 0.008479740470647812, + 0.019608283415436745, + 0.007002993952482939, + 0.008370770141482353, + 0.012722577899694443, + 0.007750030141323805, + 0.0068651833571493626, + 0.004190270323306322, + 0.004456056747585535, + 0.006503733806312084, + 0.003945779986679554, + 0.00542790861800313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.4531049132347107 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34866011142730713 + ], + "opposite_logprob_tokens": [ + -1.2236601114273071 + ], + "logprob_reward": -0.34866011142730713, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34866011142730713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12498348206281662, + 0.0665198564529419, + 0.09481554478406906, + 0.1333596408367157, + 0.050846364349126816, + 0.21703095734119415, + 0.08698702603578568, + 0.09510106593370438, + 0.05336892604827881, + 0.07698714733123779 + ], + "credit_map": [ + 0.0209127739071846, + 0.020938411355018616, + 0.02745712921023369, + 0.013702237978577614, + 0.01357590314000845, + 0.009268330410122871, + 0.008730767294764519, + 0.010397924110293388, + 0.008633011020720005, + 0.009048933163285255, + 0.0065526110120117664, + 0.006405276712030172, + 0.009248214773833752, + 0.010736534371972084, + 0.008595643565058708, + 0.007299636956304312, + 0.012235570698976517, + 0.009789244271814823, + 0.010962734930217266, + 0.013773507438600063, + 0.01217956654727459, + 0.011233310215175152, + 0.011819028295576572, + 0.012822581455111504, + 0.019233455881476402, + 0.032003067433834076, + 0.020071279257535934, + 0.029096418991684914, + 0.010339380241930485, + 0.00893204566091299, + 0.007000510115176439, + 0.006683481857180595, + 0.008983774110674858, + 0.00962737388908863, + 0.005858915857970715, + 0.0067880600690841675, + 0.0046098241582512856, + 0.0040966239757835865, + 0.006504174321889877, + 0.0043776207603514194, + 0.027169646695256233, + 0.04799819365143776, + 0.02632407285273075, + 0.02195826731622219, + 0.02397952787578106, + 0.029277263209223747, + 0.020869627594947815, + 0.019454367458820343, + 0.01423993892967701, + 0.014814887195825577, + 0.009394349530339241, + 0.008360086008906364, + 0.008335684426128864, + 0.01789543777704239, + 0.007872501388192177, + 0.006074137054383755, + 0.022537153214216232, + 0.011348099447786808, + 0.012954533100128174, + 0.006334051489830017, + 0.0066540418192744255, + 0.026269780471920967, + 0.0036732791922986507, + 0.0053301360458135605, + 0.004986563231796026, + 0.005428125150501728, + 0.006144311744719744, + 0.007207910530269146, + 0.0064819990657269955, + 0.0084309633821249, + 0.00641412241384387, + 0.0082749268040061, + 0.013598635792732239, + 0.013470440171658993, + 0.0077467868104577065, + 0.008549041114747524, + 0.007736049126833677, + 0.01139146275818348, + 0.006425673607736826, + 0.008069059811532497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.34866011142730713 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.2379066944122314 + ], + "opposite_logprob_tokens": [ + -0.1129067912697792 + ], + "logprob_reward": -2.2379066944122314, + "yes_no_margin_reward": -2.125, + "token_credit_reward": -2.2379066944122314, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21969984471797943, + 0.08152589946985245, + 0.036887411028146744, + 0.06431344896554947, + 0.04045940935611725, + 0.052982915192842484, + 0.05435210466384888, + 0.34034430980682373, + 0.07423833012580872, + 0.03519625961780548 + ], + "credit_map": [ + 0.048030219972133636, + 0.02018222026526928, + 0.05567340552806854, + 0.02291249670088291, + 0.018685318529605865, + 0.01992337964475155, + 0.01585659198462963, + 0.018436195328831673, + 0.009710144251585007, + 0.006742505822330713, + 0.014036271721124649, + 0.010295422747731209, + 0.011123809032142162, + 0.008086624555289745, + 0.012686059810221195, + 0.008845058269798756, + 0.007828117348253727, + 0.003512121271342039, + 0.004796308930963278, + 0.003393429797142744, + 0.004070753697305918, + 0.0039787376299500465, + 0.003769109956920147, + 0.005538831930607557, + 0.004863443784415722, + 0.008418416604399681, + 0.008356970734894276, + 0.007714065723121166, + 0.00894249975681305, + 0.00443935114890337, + 0.01268098596483469, + 0.008897720836102962, + 0.007757723797112703, + 0.005200323183089495, + 0.006477152928709984, + 0.005459911655634642, + 0.0030663078650832176, + 0.003563164733350277, + 0.005032307934015989, + 0.003902516094967723, + 0.007065810728818178, + 0.010163995437324047, + 0.007718083914369345, + 0.0066040582023561, + 0.005152638535946608, + 0.007312963251024485, + 0.003924823831766844, + 0.005040542688220739, + 0.012842478230595589, + 0.007930411025881767, + 0.006189604755491018, + 0.0029649215284734964, + 0.005095972679555416, + 0.009723903611302376, + 0.005034881178289652, + 0.004569934215396643, + 0.06236926466226578, + 0.047192372381687164, + 0.04804697632789612, + 0.03158802539110184, + 0.02465151995420456, + 0.08948295563459396, + 0.018492618575692177, + 0.018520575016736984, + 0.009419357404112816, + 0.008962497115135193, + 0.0114038847386837, + 0.006481529679149389, + 0.007681143470108509, + 0.015185490250587463, + 0.007562241517007351, + 0.007542187348008156, + 0.004589209798723459, + 0.0045423610135912895, + 0.004540763795375824, + 0.0032991624902933836, + 0.004223753232508898, + 0.005725480616092682, + 0.004223489202558994, + 0.004052040167152882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -2.2379066944122314 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060566283762455 + ], + "opposite_logprob_tokens": [ + -3.8956055641174316 + ], + "logprob_reward": -0.02060566283762455, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060566283762455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1523301899433136, + 0.1340121179819107, + 0.05441335588693619, + 0.08777813613414764, + 0.05807947367429733, + 0.1049274280667305, + 0.051258888095617294, + 0.2619245648384094, + 0.058406706899404526, + 0.03686906769871712 + ], + "credit_map": [ + 0.0415237694978714, + 0.01419296208769083, + 0.05018008127808571, + 0.014175588265061378, + 0.008463075384497643, + 0.006906862836331129, + 0.007797676138579845, + 0.00909016840159893, + 0.01529412716627121, + 0.012154340744018555, + 0.024275749921798706, + 0.01522835623472929, + 0.017892926931381226, + 0.015215563587844372, + 0.01911173202097416, + 0.01483931764960289, + 0.010476397350430489, + 0.0045606475323438644, + 0.00799645110964775, + 0.006060475949198008, + 0.005839647259563208, + 0.0051249354146420956, + 0.00559252267703414, + 0.008762281388044357, + 0.010771454311907291, + 0.016108451411128044, + 0.010266073979437351, + 0.013486511074006557, + 0.013791050761938095, + 0.0064746360294520855, + 0.00822360534220934, + 0.008656356483697891, + 0.00991586409509182, + 0.007531074341386557, + 0.007464557886123657, + 0.008867345750331879, + 0.0053924028761684895, + 0.0057515669614076614, + 0.007298044860363007, + 0.005858615972101688, + 0.014140141196548939, + 0.019926071166992188, + 0.014710542745888233, + 0.015311416238546371, + 0.011835390701889992, + 0.011135446839034557, + 0.007047486025840044, + 0.010820931755006313, + 0.009666452184319496, + 0.008609897457063198, + 0.006181872449815273, + 0.003500835271552205, + 0.005505204666405916, + 0.008217927999794483, + 0.004893348552286625, + 0.004683353006839752, + 0.048471514135599136, + 0.032442037016153336, + 0.03866498917341232, + 0.02345738559961319, + 0.01952943205833435, + 0.07163526117801666, + 0.012951257638633251, + 0.014772702939808369, + 0.008603697642683983, + 0.007566260173916817, + 0.010101278312504292, + 0.005320366006344557, + 0.0066142030991613865, + 0.010946212336421013, + 0.004459319170564413, + 0.004795372951775789, + 0.0068006087094545364, + 0.005543872714042664, + 0.004040522966533899, + 0.004298675339668989, + 0.004193123430013657, + 0.0043683238327503204, + 0.004203370772302151, + 0.0034205697011202574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.02060566283762455 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278928935527802 + ], + "opposite_logprob_tokens": [ + -2.017789363861084 + ], + "logprob_reward": -0.14278928935527802, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14278928935527802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10911595076322556, + 0.036016084253787994, + 0.04643620178103447, + 0.0248199962079525, + 0.010807568207383156, + 0.030338464304804802, + 0.06852930039167404, + 0.36077991127967834, + 0.16460008919239044, + 0.1485564112663269 + ], + "credit_map": [ + 0.013314579613506794, + 0.009189270436763763, + 0.016683844849467278, + 0.011534948833286762, + 0.019438382238149643, + 0.012711528688669205, + 0.012919992208480835, + 0.013323407620191574, + 0.0044716158881783485, + 0.003523746272549033, + 0.0048177363350987434, + 0.003680181223899126, + 0.005207113455981016, + 0.004776152782142162, + 0.005473348777741194, + 0.004066190682351589, + 0.011245914734899998, + 0.005092431791126728, + 0.006621288601309061, + 0.004323425702750683, + 0.0042238617315888405, + 0.004752460867166519, + 0.003786814399063587, + 0.006390003487467766, + 0.0033216727897524834, + 0.004088304471224546, + 0.0030704904347658157, + 0.0037814397364854813, + 0.002213793806731701, + 0.0026839266065508127, + 0.0034391386434435844, + 0.0022212287876755, + 0.002368619665503502, + 0.0014860874507576227, + 0.0010165793355554342, + 0.0014213380636647344, + 0.0009849050547927618, + 0.0010387781076133251, + 0.0016060047782957554, + 0.0008852552273310721, + 0.004064156673848629, + 0.006376704666763544, + 0.003679996822029352, + 0.002744003664702177, + 0.003229975700378418, + 0.00532535882666707, + 0.0024546259082853794, + 0.0024636411108076572, + 0.013486944139003754, + 0.012826981954276562, + 0.009706894867122173, + 0.0053777387365698814, + 0.00614692410454154, + 0.009035778231918812, + 0.006483906880021095, + 0.005464131478220224, + 0.10572792589664459, + 0.04273330420255661, + 0.031664811074733734, + 0.013359282165765762, + 0.016335569322109222, + 0.1245288997888565, + 0.011041032150387764, + 0.015389066189527512, + 0.02449563518166542, + 0.021127857267856598, + 0.023968011140823364, + 0.013631935231387615, + 0.018006015568971634, + 0.03181367367506027, + 0.014928076416254044, + 0.016628887504339218, + 0.031036045402288437, + 0.026898905634880066, + 0.015390577726066113, + 0.01581438072025776, + 0.015577594749629498, + 0.019005296751856804, + 0.01148728933185339, + 0.013346309773623943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.14278928935527802 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion shot, shot from an upward tilt, captured in a striking, high-contrast black-and-white aesthetic, captures a strikingly muscular female athlete at the peak of her leap during a beach volleyball match. Her body, tied in a tight ponytail, arches in the air, the tense muscles of her back, shoulders, and arms clearly visible in the strong sunlight. Clad in a dark sports bikini, her facial expression is one of intense concentration, eyes narrowed, jaw set, her complete focus focused on the volleyball hovering directly in front of her. Her right arm is fully extended back, palm open, poised for a thunderous smash. The intense, direct sunlight streaming from above and side sculpts her form into a dynamic composition of deep shadows and dazzling whites, while each drop of sweat dripping from her skin shimmers with a shimmering light. Below her, fine grains of sand, sent flying by her explosive leap, freeze in mid-air, while the upper edge of the volleyball net forms the foreground. The entire scene is imbued with a raw sense of power and frozen dynamic beauty, conveying a tense, focused, and explosive competitive atmosphere.", + "seed": 11931, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_4/index_2141_seed_11931_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.41193240620195865, + "vlm_logprob": -0.41193240620195865, + "vlm_yes_no_margin": 1.375, + "vlm_token_credit": -0.41193240620195865, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32213065028190613 + ], + "opposite_logprob_tokens": [ + -1.5721306800842285 + ], + "logprob_reward": -0.32213065028190613, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.32213065028190613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1980700045824051, + 0.10021626949310303, + 0.0963529646396637, + 0.056344132870435715, + 0.1043403223156929, + 0.06925131380558014, + 0.17246277630329132, + 0.09246687591075897, + 0.07681204378604889, + 0.03368326649069786 + ], + "credit_map": [ + 0.019164560362696648, + 0.02315996214747429, + 0.024909501895308495, + 0.032413285225629807, + 0.023564612492918968, + 0.029437100514769554, + 0.01206991821527481, + 0.03335106000304222, + 0.009602148085832596, + 0.007710624020546675, + 0.013877980411052704, + 0.009471731260418892, + 0.011434139683842659, + 0.012725230306386948, + 0.01233850046992302, + 0.02305591106414795, + 0.009060808457434177, + 0.012490345165133476, + 0.010000399313867092, + 0.01250535249710083, + 0.01460039522498846, + 0.01221281848847866, + 0.015835899859666824, + 0.009646948426961899, + 0.007759708445519209, + 0.008090919815003872, + 0.005560980178415775, + 0.0074434783309698105, + 0.005719633307307959, + 0.007050239015370607, + 0.007060754578560591, + 0.007658416870981455, + 0.019424516707658768, + 0.008445791900157928, + 0.012081841006875038, + 0.010900872759521008, + 0.007467115297913551, + 0.0222616083920002, + 0.013973388820886612, + 0.009785184636712074, + 0.0087346863001585, + 0.010385238565504551, + 0.006846574135124683, + 0.008805928751826286, + 0.007916425354778767, + 0.009211408905684948, + 0.006248171441257, + 0.011102876625955105, + 0.051074542105197906, + 0.01450328528881073, + 0.008082610554993153, + 0.010752789676189423, + 0.012452144175767899, + 0.05290122702717781, + 0.012743862345814705, + 0.009952327236533165, + 0.008920294232666492, + 0.01562302466481924, + 0.010647564195096493, + 0.010034628212451935, + 0.009896044619381428, + 0.015939680859446526, + 0.00867551751434803, + 0.012730119749903679, + 0.0065229302272200584, + 0.00807109009474516, + 0.01090044341981411, + 0.01885414868593216, + 0.005247923545539379, + 0.011704355478286743, + 0.006400002166628838, + 0.009111152961850166, + 0.001874220441095531, + 0.002038415055721998, + 0.0055286637507379055, + 0.004130563698709011, + 0.0020335924345999956, + 0.004778842907398939, + 0.007290450390428305, + 0.006008516997098923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.32213065028190613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31346240639686584 + ], + "opposite_logprob_tokens": [ + -1.3134623765945435 + ], + "logprob_reward": -0.31346240639686584, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31346240639686584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06867919862270355, + 0.1563137173652649, + 0.11187820136547089, + 0.054534588009119034, + 0.11167719215154648, + 0.06668794900178909, + 0.09369904547929764, + 0.10391390323638916, + 0.19761107861995697, + 0.0350051186978817 + ], + "credit_map": [ + 0.010480058379471302, + 0.00901530310511589, + 0.006204934790730476, + 0.012613333761692047, + 0.007949435152113438, + 0.010695770382881165, + 0.004650362301617861, + 0.00707000307738781, + 0.013940036296844482, + 0.01106176245957613, + 0.022060709074139595, + 0.012329705990850925, + 0.0174150038510561, + 0.025973578914999962, + 0.014078157022595406, + 0.03945476934313774, + 0.01340967882424593, + 0.008949721232056618, + 0.010693621821701527, + 0.01048621442168951, + 0.014495383016765118, + 0.019071482121944427, + 0.022012747824192047, + 0.01275935210287571, + 0.00783641915768385, + 0.006262802053242922, + 0.005218347068876028, + 0.007789529860019684, + 0.005197602789849043, + 0.0069649312645196915, + 0.007999079301953316, + 0.0072658793069422245, + 0.021491356194019318, + 0.00940596591681242, + 0.01666233316063881, + 0.017998380586504936, + 0.007915526628494263, + 0.017700644209980965, + 0.010935243219137192, + 0.009567749686539173, + 0.009704524651169777, + 0.01141485758125782, + 0.010987973771989346, + 0.007228020578622818, + 0.008052329532802105, + 0.007543790154159069, + 0.005981716327369213, + 0.005774739198386669, + 0.030958635732531548, + 0.00521528534591198, + 0.004455585032701492, + 0.004971795715391636, + 0.005991396494209766, + 0.0336461216211319, + 0.004087524488568306, + 0.004372705239802599, + 0.011534548364579678, + 0.032980792224407196, + 0.005920120980590582, + 0.00633536372333765, + 0.0070986212231218815, + 0.02373378351330757, + 0.0081235496327281, + 0.008187134750187397, + 0.018730109557509422, + 0.02071104571223259, + 0.02869074046611786, + 0.033029038459062576, + 0.014667999930679798, + 0.04229237884283066, + 0.015351014211773872, + 0.024138767272233963, + 0.001997585641220212, + 0.0027055225800722837, + 0.003122687805444002, + 0.004964996594935656, + 0.0023849615827202797, + 0.007105072494596243, + 0.0073562003672122955, + 0.0053680953569710255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.31346240639686584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236204147338867 + ], + "opposite_logprob_tokens": [ + -0.3486204147338867 + ], + "logprob_reward": -1.2236204147338867, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236204147338867, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07778932899236679, + 0.0828753262758255, + 0.04035159945487976, + 0.056391775608062744, + 0.05405816063284874, + 0.20094764232635498, + 0.2646113932132721, + 0.12547676265239716, + 0.061266716569662094, + 0.036231230944395065 + ], + "credit_map": [ + 0.011515283957123756, + 0.007267667446285486, + 0.01064196228981018, + 0.009530600160360336, + 0.008456927724182606, + 0.009911829605698586, + 0.007795115001499653, + 0.012669945135712624, + 0.007397853769361973, + 0.005306439008563757, + 0.012786856852471828, + 0.004899892024695873, + 0.007264182902872562, + 0.014346393756568432, + 0.008016412146389484, + 0.02285730093717575, + 0.005483821500092745, + 0.004005260765552521, + 0.004617195576429367, + 0.005106941796839237, + 0.004497419577091932, + 0.004973788280040026, + 0.007150481920689344, + 0.0045166886411607265, + 0.007607804145663977, + 0.005959589034318924, + 0.005498612765222788, + 0.007291992660611868, + 0.004691208712756634, + 0.007655064109712839, + 0.007417966146022081, + 0.010269535705447197, + 0.010329731740057468, + 0.004371690098196268, + 0.007126207463443279, + 0.0049856253899633884, + 0.003539110068231821, + 0.012684590183198452, + 0.0059143370017409325, + 0.005106867756694555, + 0.04510412737727165, + 0.015731530264019966, + 0.0661827027797699, + 0.02292756922543049, + 0.007500531617552042, + 0.017661476507782936, + 0.009217825718224049, + 0.01662188582122326, + 0.061513807624578476, + 0.01529761590063572, + 0.021586507558822632, + 0.016482923179864883, + 0.0197006668895483, + 0.06788497418165207, + 0.02562786266207695, + 0.036517027765512466, + 0.018324464559555054, + 0.027292240411043167, + 0.018365586176514626, + 0.009847075678408146, + 0.010892371647059917, + 0.020354554057121277, + 0.008101149462163448, + 0.012299319729208946, + 0.0072327894158661366, + 0.008478010073304176, + 0.008736606687307358, + 0.014783134683966637, + 0.0034269411116838455, + 0.005940865725278854, + 0.004104221239686012, + 0.008564149960875511, + 0.0023300033062696457, + 0.00429893471300602, + 0.005754091776907444, + 0.004135994240641594, + 0.002342009684070945, + 0.006546972319483757, + 0.005521907936781645, + 0.0053013162687420845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -1.2236204147338867 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592467859387398 + ], + "opposite_logprob_tokens": [ + -3.895592451095581 + ], + "logprob_reward": -0.020592467859387398, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592467859387398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07549574226140976, + 0.08258146792650223, + 0.0939117968082428, + 0.056185416877269745, + 0.10658100247383118, + 0.23663152754306793, + 0.1530941277742386, + 0.06568463891744614, + 0.09413618594408035, + 0.035698048770427704 + ], + "credit_map": [ + 0.007322160992771387, + 0.008556216955184937, + 0.007959880866110325, + 0.01358617004007101, + 0.011549276299774647, + 0.0098471250385046, + 0.0057749818079173565, + 0.01089993491768837, + 0.00781719759106636, + 0.007434293162077665, + 0.009503980167210102, + 0.009279798716306686, + 0.010939632542431355, + 0.01222194079309702, + 0.010024089366197586, + 0.015360541641712189, + 0.013374865986406803, + 0.00826701894402504, + 0.010611793026328087, + 0.010640314780175686, + 0.007597572635859251, + 0.02036910690367222, + 0.011937204748392105, + 0.01111391931772232, + 0.007157245650887489, + 0.007109536323696375, + 0.0057567935436964035, + 0.007292183581739664, + 0.005471008364111185, + 0.0090295085683465, + 0.007879120297729969, + 0.006490022409707308, + 0.015727804973721504, + 0.011608737520873547, + 0.01995307020843029, + 0.01510465145111084, + 0.00756981922313571, + 0.014486344531178474, + 0.01090841181576252, + 0.011222167871892452, + 0.041656456887722015, + 0.05605441704392433, + 0.03303146734833717, + 0.024582520127296448, + 0.021749690175056458, + 0.023276764899492264, + 0.02360987663269043, + 0.012670360505580902, + 0.048341356217861176, + 0.008094098418951035, + 0.006831706501543522, + 0.011863463558256626, + 0.010546760633587837, + 0.0498950332403183, + 0.009538061916828156, + 0.007983646355569363, + 0.008560337126255035, + 0.0154088931158185, + 0.008126620203256607, + 0.006788082420825958, + 0.005794293712824583, + 0.010021349415183067, + 0.004805553238838911, + 0.006179518532007933, + 0.006695961579680443, + 0.014099745079874992, + 0.016345972195267677, + 0.025340337306261063, + 0.005143802613019943, + 0.008357442915439606, + 0.006223262753337622, + 0.011929663829505444, + 0.0020862827077507973, + 0.0027970413211733103, + 0.006581416819244623, + 0.004589952994138002, + 0.003133447840809822, + 0.005146989598870277, + 0.006064692046493292, + 0.005298227071762085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.020592467859387398 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798560917377472 + ], + "opposite_logprob_tokens": [ + -1.8048560619354248 + ], + "logprob_reward": -0.1798560917377472, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798560917377472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060663193464279175, + 0.08388669788837433, + 0.06482908874750137, + 0.08321920782327652, + 0.19769705832004547, + 0.2531701922416687, + 0.05432302504777908, + 0.05207644775509834, + 0.09800669550895691, + 0.05212836340069771 + ], + "credit_map": [ + 0.007105346769094467, + 0.008129307068884373, + 0.0076826331205666065, + 0.0068445391952991486, + 0.006286602467298508, + 0.008374744094908237, + 0.004617859143763781, + 0.011622157879173756, + 0.012034197337925434, + 0.006899724714457989, + 0.009167376905679703, + 0.005633162800222635, + 0.011788036674261093, + 0.017944645136594772, + 0.005586747545748949, + 0.014832800254225731, + 0.0073247128166258335, + 0.004826402757316828, + 0.007248187903314829, + 0.007914770394563675, + 0.0072318254970014095, + 0.009313014335930347, + 0.012250163592398167, + 0.008720003999769688, + 0.012135030701756477, + 0.011381380259990692, + 0.007137124426662922, + 0.010145258158445358, + 0.0069790007546544075, + 0.015258584171533585, + 0.009971324354410172, + 0.01021149568259716, + 0.02758973278105259, + 0.016704708337783813, + 0.03054818883538246, + 0.02965552918612957, + 0.01698126643896103, + 0.024283478036522865, + 0.025276966392993927, + 0.02665717713534832, + 0.04884761571884155, + 0.03177576884627342, + 0.07279154658317566, + 0.030573217198252678, + 0.01810256950557232, + 0.02085857093334198, + 0.012567522935569286, + 0.017653392627835274, + 0.012970083393156528, + 0.003962114918977022, + 0.00399350468069315, + 0.004202537704259157, + 0.0043144687078893185, + 0.013834841549396515, + 0.0038501450326293707, + 0.007195326033979654, + 0.005330532789230347, + 0.017796536907553673, + 0.004590287804603577, + 0.0026199831627309322, + 0.002445286139845848, + 0.012676384299993515, + 0.003127763047814369, + 0.003489673137664795, + 0.008872170932590961, + 0.011334492824971676, + 0.011402704752981663, + 0.019652463495731354, + 0.006791246123611927, + 0.012113252654671669, + 0.012061492539942265, + 0.015778863802552223, + 0.00366894225589931, + 0.0038737209979444742, + 0.006515114102512598, + 0.007353279739618301, + 0.00303662265650928, + 0.009964999742805958, + 0.007256136741489172, + 0.010459545068442822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.1798560917377472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion shot, shot from an upward tilt, captured in a striking, high-contrast black-and-white aesthetic, captures a strikingly muscular female athlete at the peak of her leap during a beach volleyball match. Her body, tied in a tight ponytail, arches in the air, the tense muscles of her back, shoulders, and arms clearly visible in the strong sunlight. Clad in a dark sports bikini, her facial expression is one of intense concentration, eyes narrowed, jaw set, her complete focus focused on the volleyball hovering directly in front of her. Her right arm is fully extended back, palm open, poised for a thunderous smash. The intense, direct sunlight streaming from above and side sculpts her form into a dynamic composition of deep shadows and dazzling whites, while each drop of sweat dripping from her skin shimmers with a shimmering light. Below her, fine grains of sand, sent flying by her explosive leap, freeze in mid-air, while the upper edge of the volleyball net forms the foreground. The entire scene is imbued with a raw sense of power and frozen dynamic beauty, conveying a tense, focused, and explosive competitive atmosphere.", + "seed": 11932, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_4/index_2141_seed_11932_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.46455565430223944, + "vlm_logprob": -0.46455565430223944, + "vlm_yes_no_margin": 1.475, + "vlm_token_credit": -0.46455565430223944, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1977413296699524 + ], + "opposite_logprob_tokens": [ + -2.3227412700653076 + ], + "logprob_reward": -0.1977413296699524, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1977413296699524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030479663982987404, + 0.015116102062165737, + 0.04047132283449173, + 0.028639156371355057, + 0.6283281445503235, + 0.05154792591929436, + 0.029857845976948738, + 0.060028742998838425, + 0.027528222650289536, + 0.08800286054611206 + ], + "credit_map": [ + 0.0033201032783836126, + 0.0028019193559885025, + 0.007439304143190384, + 0.0031692327465862036, + 0.001755348639562726, + 0.003549521556124091, + 0.0019246001029387116, + 0.006519637536257505, + 0.0022385281044989824, + 0.0019706599414348602, + 0.001915498054586351, + 0.0018519422737881541, + 0.0010536483023315668, + 0.002804056741297245, + 0.0009798652026802301, + 0.0023019055370241404, + 0.0036768473219126463, + 0.002690010005608201, + 0.007551460526883602, + 0.0054087331518530846, + 0.0028350052889436483, + 0.005045641679316759, + 0.004400154110044241, + 0.008863473311066628, + 0.004410736262798309, + 0.005912092514336109, + 0.0024212822318077087, + 0.0034858458675444126, + 0.004308653529733419, + 0.0029668444767594337, + 0.001986522227525711, + 0.003147182520478964, + 0.1227545216679573, + 0.018317967653274536, + 0.1670449823141098, + 0.026549693197011948, + 0.019620176404714584, + 0.23357513546943665, + 0.013620705343782902, + 0.026844976469874382, + 0.00972721353173256, + 0.005906489212065935, + 0.006291884928941727, + 0.0035410395357757807, + 0.004591081291437149, + 0.009569116868078709, + 0.003857983974739909, + 0.008063122630119324, + 0.008074729703366756, + 0.0025733644142746925, + 0.004786018747836351, + 0.0018415000522509217, + 0.002376353368163109, + 0.005233714357018471, + 0.001562230521813035, + 0.0034099407494068146, + 0.02092483453452587, + 0.003234485164284706, + 0.01275202352553606, + 0.0026905506383627653, + 0.0027722734957933426, + 0.010111195035278797, + 0.002754779765382409, + 0.004788605961948633, + 0.006255703512579203, + 0.0016489995177835226, + 0.0033555852714926004, + 0.0020666494965553284, + 0.0027498770505189896, + 0.003894323715940118, + 0.0014717340236529708, + 0.006085352972149849, + 0.005646029952913523, + 0.0028107506223022938, + 0.0028177788481116295, + 0.002133988542482257, + 0.035271238535642624, + 0.03267725557088852, + 0.0016290495404973626, + 0.005016776733100414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.1977413296699524 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289000332355499 + ], + "opposite_logprob_tokens": [ + -1.0539000034332275 + ], + "logprob_reward": -0.4289000332355499, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289000332355499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0166894793510437, + 0.012226849794387817, + 0.019657865166664124, + 0.014210651628673077, + 0.73209547996521, + 0.12929394841194153, + 0.013797024264931679, + 0.027337918058037758, + 0.022836903110146523, + 0.011853852309286594 + ], + "credit_map": [ + 0.0019439567113295197, + 0.0013550086878240108, + 0.003147246316075325, + 0.0018030105857178569, + 0.0007583261467516422, + 0.0025956411845982075, + 0.001003300305455923, + 0.004082989413291216, + 0.0017423807876184583, + 0.0011639707954600453, + 0.0014389610150828958, + 0.0011319246841594577, + 0.0007779033039696515, + 0.002456609858199954, + 0.0007922965451143682, + 0.0027228030376136303, + 0.0019424649653956294, + 0.0012988913804292679, + 0.004485893528908491, + 0.0024378644302487373, + 0.0014288933016359806, + 0.0027609714306890965, + 0.001261755358427763, + 0.004041130188852549, + 0.00220743240788579, + 0.002734849927946925, + 0.0015212403377518058, + 0.0012789624743163586, + 0.0016293285880237818, + 0.0016664356226101518, + 0.001285279169678688, + 0.001887122867628932, + 0.15371446311473846, + 0.021246442571282387, + 0.18438047170639038, + 0.02985933981835842, + 0.022772930562496185, + 0.27454498410224915, + 0.015324562788009644, + 0.03025227226316929, + 0.036575719714164734, + 0.007698231842368841, + 0.02770380862057209, + 0.005045630969107151, + 0.010096285492181778, + 0.021451398730278015, + 0.0058282287791371346, + 0.0148946363478899, + 0.0035034380853176117, + 0.000618527818005532, + 0.0024393729399889708, + 0.0007862265920266509, + 0.0013128237333148718, + 0.0025764750316739082, + 0.0007446612580679357, + 0.0018154988065361977, + 0.00998231302946806, + 0.00080898217856884, + 0.007355296518653631, + 0.0010505558457225561, + 0.0010026897070929408, + 0.004157651215791702, + 0.0009332121699117124, + 0.0020472186151891947, + 0.005704397335648537, + 0.0019760136492550373, + 0.0021249635610729456, + 0.0016845061909407377, + 0.0015627987449988723, + 0.0044201514683663845, + 0.0012460147263482213, + 0.004118059296160936, + 0.0018852086504921317, + 0.0010851913830265403, + 0.0012022763257846236, + 0.0007797323632985353, + 0.00208672764711082, + 0.0034102797508239746, + 0.0005161476437933743, + 0.0008882885449565947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.4289000332355499 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020678043365479 + ], + "opposite_logprob_tokens": [ + -0.2520677447319031 + ], + "logprob_reward": -1.5020678043365479, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020678043365479, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.02178538590669632, + 0.011364432983100414, + 0.0168585404753685, + 0.019805805757641792, + 0.16315895318984985, + 0.6763474345207214, + 0.016853097826242447, + 0.0227553378790617, + 0.009817437268793583, + 0.04125354811549187 + ], + "credit_map": [ + 0.002750591840595007, + 0.002165035577490926, + 0.004100373480468988, + 0.0025097241159528494, + 0.0019624580163508654, + 0.0022397127468138933, + 0.0022320759017020464, + 0.0038254123646765947, + 0.0015594519209116697, + 0.0012471372028812766, + 0.0014647928765043616, + 0.001515616662800312, + 0.0008903450798243284, + 0.0016637699445709586, + 0.0007068304694257677, + 0.002316487720236182, + 0.0023178791161626577, + 0.0010057342005893588, + 0.0028235982172191143, + 0.002174373483285308, + 0.0012349612079560757, + 0.002245082287117839, + 0.0010849806712940335, + 0.003971931058913469, + 0.002717249095439911, + 0.00488338153809309, + 0.0016834646230563521, + 0.0013545920373871922, + 0.003519105026498437, + 0.0017861485248431563, + 0.001546272891573608, + 0.0023155920207500458, + 0.0354907251894474, + 0.005306941922754049, + 0.039408713579177856, + 0.006783826742321253, + 0.005146462935954332, + 0.060652244836091995, + 0.0037065898068249226, + 0.006663437932729721, + 0.2638639807701111, + 0.01886100508272648, + 0.20842188596725464, + 0.01965760625898838, + 0.0199107863008976, + 0.09261936694383621, + 0.012087004259228706, + 0.040925782173871994, + 0.0036272837314754725, + 0.0021511628292500973, + 0.0023009246215224266, + 0.0017817830666899681, + 0.0018096553394570947, + 0.002059939783066511, + 0.0013981660595163703, + 0.0017241798341274261, + 0.004147141706198454, + 0.001087200129404664, + 0.004360914696007967, + 0.0011712461709976196, + 0.0014280901523306966, + 0.004502561874687672, + 0.001982073998078704, + 0.004076107405126095, + 0.0023255913984030485, + 0.0008839837973937392, + 0.0010640070540830493, + 0.0007383169722743332, + 0.0007915153400972486, + 0.0013120706425979733, + 0.0005245489301159978, + 0.0021774021442979574, + 0.0035789310932159424, + 0.003449345240369439, + 0.002413443522527814, + 0.0023291639517992735, + 0.014692802913486958, + 0.010346745140850544, + 0.001242974540218711, + 0.0032001384533941746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -1.5020678043365479 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210226014256477 + ], + "opposite_logprob_tokens": [ + -4.264210224151611 + ], + "logprob_reward": -0.014210226014256477, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210226014256477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12477973103523254, + 0.01991182006895542, + 0.026320187374949455, + 0.04382871091365814, + 0.09068583697080612, + 0.37186577916145325, + 0.06543945521116257, + 0.07411249727010727, + 0.14248411357402802, + 0.04057186841964722 + ], + "credit_map": [ + 0.01885972172021866, + 0.013421619310975075, + 0.021792646497488022, + 0.011481017805635929, + 0.011644122190773487, + 0.014125381596386433, + 0.008072233758866787, + 0.02538299560546875, + 0.0027082827873528004, + 0.0031662785913795233, + 0.002250370103865862, + 0.0030326892156153917, + 0.0017229458317160606, + 0.0023140760604292154, + 0.0017118268879130483, + 0.003005353733897209, + 0.002390544395893812, + 0.0022968400735408068, + 0.004275064449757338, + 0.005907303653657436, + 0.0025266637094318867, + 0.0028234669007360935, + 0.002617424353957176, + 0.0034828814677894115, + 0.005758942104876041, + 0.009150232188403606, + 0.004302954766899347, + 0.0037182916421443224, + 0.006110291928052902, + 0.0044046551920473576, + 0.00304994429461658, + 0.007333407178521156, + 0.01772811822593212, + 0.003265759674832225, + 0.022388093173503876, + 0.0049562109634280205, + 0.003876121947541833, + 0.031116731464862823, + 0.002800984773784876, + 0.004553826060146093, + 0.1361674666404724, + 0.013436240144073963, + 0.10273270308971405, + 0.011605306528508663, + 0.015473629347980022, + 0.056446101516485214, + 0.00978277251124382, + 0.026221556589007378, + 0.02096378244459629, + 0.0046837665140628815, + 0.014507816173136234, + 0.00325540523044765, + 0.0049647381529212, + 0.009177087806165218, + 0.0026560272090137005, + 0.00523083982989192, + 0.01608709990978241, + 0.004497068468481302, + 0.014292380772531033, + 0.005222462583333254, + 0.006001514848321676, + 0.008649969473481178, + 0.009452383033931255, + 0.00990962889045477, + 0.03641541674733162, + 0.01127841416746378, + 0.01458628848195076, + 0.008187114261090755, + 0.012705995701253414, + 0.023014387115836143, + 0.006429916713386774, + 0.029866589233279228, + 0.005982005503028631, + 0.004184874705970287, + 0.007034458220005035, + 0.003453523153439164, + 0.0040496052242815495, + 0.005533905234187841, + 0.0032629535999149084, + 0.007070546038448811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.014210226014256477 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985887825489044 + ], + "opposite_logprob_tokens": [ + -1.804858922958374 + ], + "logprob_reward": -0.17985887825489044, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985887825489044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023796629160642624, + 0.024978380650281906, + 0.04619494080543518, + 0.048194438219070435, + 0.1569887399673462, + 0.3073388636112213, + 0.09097500145435333, + 0.25004544854164124, + 0.030452681705355644, + 0.021034812554717064 + ], + "credit_map": [ + 0.003933153115212917, + 0.0028464673087000847, + 0.005171481985598803, + 0.002223466755822301, + 0.0013802398461848497, + 0.0032228848431259394, + 0.0012823637807741761, + 0.0037365718744695187, + 0.004609003197401762, + 0.0020175036042928696, + 0.003408905351534486, + 0.0017523672431707382, + 0.001572948764078319, + 0.006332735996693373, + 0.0014302278868854046, + 0.0038546891883015633, + 0.006633690558373928, + 0.0033576563000679016, + 0.010220050811767578, + 0.005920730996876955, + 0.0029341401532292366, + 0.006209659390151501, + 0.0022462881170213223, + 0.008672727271914482, + 0.00676734047010541, + 0.012750719673931599, + 0.0038897343911230564, + 0.003659793408587575, + 0.009014364331960678, + 0.004481385927647352, + 0.0033693008590489626, + 0.004261794500052929, + 0.05684787407517433, + 0.0062493206933140755, + 0.013713023625314236, + 0.004506566096097231, + 0.006566399242728949, + 0.059805091470479965, + 0.0031701582483947277, + 0.006130307912826538, + 0.12208526581525803, + 0.009861722588539124, + 0.09569849073886871, + 0.009215012192726135, + 0.007348666898906231, + 0.04052840918302536, + 0.005903489887714386, + 0.016697825863957405, + 0.023441988974809647, + 0.009539508260786533, + 0.01311785914003849, + 0.005628697574138641, + 0.008953496813774109, + 0.015402290038764477, + 0.005068521481007338, + 0.009822639636695385, + 0.07790784537792206, + 0.012134377844631672, + 0.048594363033771515, + 0.010569389909505844, + 0.012278883717954159, + 0.03737810626626015, + 0.01150822825729847, + 0.03967425972223282, + 0.006912556476891041, + 0.0032754584681242704, + 0.00312828435562551, + 0.002066192450001836, + 0.0028352707158774137, + 0.0036788489669561386, + 0.001467274152673781, + 0.007088794838637114, + 0.003326492151245475, + 0.001609329367056489, + 0.002811858896166086, + 0.0011482464615255594, + 0.002923034131526947, + 0.002959377830848098, + 0.0013282721629366279, + 0.004928202368319035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.17985887825489044 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion shot, shot from an upward tilt, captured in a striking, high-contrast black-and-white aesthetic, captures a strikingly muscular female athlete at the peak of her leap during a beach volleyball match. Her body, tied in a tight ponytail, arches in the air, the tense muscles of her back, shoulders, and arms clearly visible in the strong sunlight. Clad in a dark sports bikini, her facial expression is one of intense concentration, eyes narrowed, jaw set, her complete focus focused on the volleyball hovering directly in front of her. Her right arm is fully extended back, palm open, poised for a thunderous smash. The intense, direct sunlight streaming from above and side sculpts her form into a dynamic composition of deep shadows and dazzling whites, while each drop of sweat dripping from her skin shimmers with a shimmering light. Below her, fine grains of sand, sent flying by her explosive leap, freeze in mid-air, while the upper edge of the volleyball net forms the foreground. The entire scene is imbued with a raw sense of power and frozen dynamic beauty, conveying a tense, focused, and explosive competitive atmosphere.", + "seed": 11933, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_4/index_2141_seed_11933_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.4404188714921474, + "vlm_logprob": -0.4404188714921474, + "vlm_yes_no_margin": 1.6999999761581421, + "vlm_token_credit": -0.4404188714921474, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20072384178638458 + ], + "opposite_logprob_tokens": [ + -2.200723886489868 + ], + "logprob_reward": -0.20072384178638458, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.20072384178638458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24710124731063843, + 0.036466412246227264, + 0.04400932043790817, + 0.06968143582344055, + 0.021784216165542603, + 0.03978170081973076, + 0.054662540555000305, + 0.06624036282300949, + 0.045116376131772995, + 0.37515637278556824 + ], + "credit_map": [ + 0.011707880534231663, + 0.012183930724859238, + 0.009880232624709606, + 0.00730256550014019, + 0.012223017401993275, + 0.09818082302808762, + 0.007766579277813435, + 0.087856225669384, + 0.004108273424208164, + 0.004311863798648119, + 0.004681705962866545, + 0.005909436382353306, + 0.006836595479398966, + 0.0036693622823804617, + 0.003974577412009239, + 0.0029745984356850386, + 0.006853180471807718, + 0.004073094576597214, + 0.007325035985559225, + 0.006395191419869661, + 0.0053103649988770485, + 0.002646469045430422, + 0.005446944385766983, + 0.005959040019661188, + 0.0070986379869282246, + 0.008907251991331577, + 0.011055522598326206, + 0.008328021503984928, + 0.01044414285570383, + 0.007471018936485052, + 0.009606344625353813, + 0.006770492997020483, + 0.004997399169951677, + 0.0033175062853842974, + 0.002291594399139285, + 0.002092457842081785, + 0.002723640063777566, + 0.0021761315874755383, + 0.002297002589330077, + 0.0018884839955717325, + 0.004268505610525608, + 0.005001856945455074, + 0.004627597518265247, + 0.00569460354745388, + 0.004908451810479164, + 0.005863687954843044, + 0.005029128398746252, + 0.004387871827930212, + 0.0074828872457146645, + 0.0037063590716570616, + 0.0070039029233157635, + 0.004240062087774277, + 0.0046775732189416885, + 0.010684330016374588, + 0.004197508096694946, + 0.012669915333390236, + 0.005115436855703592, + 0.01384280901402235, + 0.008463827893137932, + 0.005791929550468922, + 0.008002153597772121, + 0.007170127704739571, + 0.00966467335820198, + 0.008189401589334011, + 0.004450127948075533, + 0.005139404907822609, + 0.006524085532873869, + 0.0043560839258134365, + 0.004760156385600567, + 0.007664578966796398, + 0.004224646370857954, + 0.007997294887900352, + 0.08824551850557327, + 0.03683800622820854, + 0.03718291223049164, + 0.05676696449518204, + 0.03359426185488701, + 0.03154803439974785, + 0.030229076743125916, + 0.06075161322951317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.20072384178638458 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285120368003845 + ], + "opposite_logprob_tokens": [ + -2.0178511142730713 + ], + "logprob_reward": -0.14285120368003845, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14285120368003845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20243115723133087, + 0.05720863491296768, + 0.06822729110717773, + 0.09159162640571594, + 0.04155726358294487, + 0.14231939613819122, + 0.06281708925962448, + 0.07823452353477478, + 0.047372471541166306, + 0.20824046432971954 + ], + "credit_map": [ + 0.01468973234295845, + 0.024953391402959824, + 0.02731599286198616, + 0.014660251326858997, + 0.03341203182935715, + 0.03485186770558357, + 0.019354328513145447, + 0.03319357708096504, + 0.008138279430568218, + 0.005986263044178486, + 0.006201921496540308, + 0.006186530459672213, + 0.011843930929899216, + 0.008037745952606201, + 0.005741525907069445, + 0.005072442349046469, + 0.008204277604818344, + 0.0072483797557652, + 0.009957659058272839, + 0.009683833457529545, + 0.009344197809696198, + 0.0042724874801933765, + 0.011544149369001389, + 0.007972314022481441, + 0.01182868517935276, + 0.015162650495767593, + 0.014927130192518234, + 0.0096053471788764, + 0.009432300925254822, + 0.010113037191331387, + 0.011258391663432121, + 0.00926409661769867, + 0.007782111410051584, + 0.0042701358906924725, + 0.004165660124272108, + 0.004955023527145386, + 0.004333241377025843, + 0.005307923536747694, + 0.005106493830680847, + 0.005636674351990223, + 0.01649564690887928, + 0.020000234246253967, + 0.01925830915570259, + 0.014994429424405098, + 0.014230589382350445, + 0.019412798807024956, + 0.014410363510251045, + 0.02351703681051731, + 0.00757270073518157, + 0.006399913225322962, + 0.00956260971724987, + 0.007827716879546642, + 0.004840180277824402, + 0.008887995965778828, + 0.006858046166598797, + 0.01086792629212141, + 0.011419388465583324, + 0.011447492055594921, + 0.017153333872556686, + 0.007337096147239208, + 0.007688089273869991, + 0.009159992448985577, + 0.004813334438949823, + 0.009215807542204857, + 0.004053968470543623, + 0.0057591283693909645, + 0.006905240006744862, + 0.004276667721569538, + 0.006040405482053757, + 0.009490901604294777, + 0.0037855016998946667, + 0.007060660514980555, + 0.034147799015045166, + 0.0231979638338089, + 0.03318791836500168, + 0.026428837329149246, + 0.016588130965828896, + 0.01891365647315979, + 0.019035952165722847, + 0.03674023225903511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.14285120368003845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7015314102172852 + ], + "opposite_logprob_tokens": [ + -0.20153145492076874 + ], + "logprob_reward": -1.7015314102172852, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015314102172852, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.5389343500137329, + 0.029874760657548904, + 0.10337351262569427, + 0.04218784719705582, + 0.05804308131337166, + 0.01774425245821476, + 0.029799163341522217, + 0.03911420702934265, + 0.05738892778754234, + 0.08353987336158752 + ], + "credit_map": [ + 0.017346933484077454, + 0.011563831008970737, + 0.014954508282244205, + 0.011985626071691513, + 0.023567918688058853, + 0.23151938617229462, + 0.011975984089076519, + 0.21602019667625427, + 0.0035190593916922808, + 0.002620501909404993, + 0.00552030885592103, + 0.0035628352779895067, + 0.005461929365992546, + 0.003290780121460557, + 0.0032434442546218634, + 0.002655899152159691, + 0.009770700708031654, + 0.010747432708740234, + 0.013158558867871761, + 0.01766405813395977, + 0.017471114173531532, + 0.006562499329447746, + 0.01621299982070923, + 0.011786157265305519, + 0.0066532958298921585, + 0.00742013193666935, + 0.009788517840206623, + 0.004509228281676769, + 0.0034036904107779264, + 0.0029790005646646023, + 0.0035086728166788816, + 0.003925307188183069, + 0.007974136620759964, + 0.004845526069402695, + 0.006048774812370539, + 0.006670017261058092, + 0.01154224295169115, + 0.006992397829890251, + 0.007859700359404087, + 0.0061102816835045815, + 0.0018256030743941665, + 0.0020769773982465267, + 0.002198712667450309, + 0.0019007010851055384, + 0.0018783110426738858, + 0.003217219840735197, + 0.0019259786931797862, + 0.002720748772844672, + 0.0033485647290945053, + 0.0027702408842742443, + 0.004268701653927565, + 0.0029415416065603495, + 0.002395706018432975, + 0.0052697001956403255, + 0.0029119662940502167, + 0.005892740096896887, + 0.004505930934101343, + 0.005263338331133127, + 0.008916790597140789, + 0.00427746819332242, + 0.005213197786360979, + 0.004857957828789949, + 0.003233050461858511, + 0.0028464719653129578, + 0.007033184636384249, + 0.003850675653666258, + 0.008829946629703045, + 0.0038247506599873304, + 0.004733853507786989, + 0.011829913593828678, + 0.005937752313911915, + 0.011348851025104523, + 0.018338926136493683, + 0.00729138171300292, + 0.012528602033853531, + 0.01082343515008688, + 0.005931898485869169, + 0.008163160644471645, + 0.00815357081592083, + 0.012308898381888866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -1.7015314102172852 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211636036634445 + ], + "opposite_logprob_tokens": [ + -4.264211654663086 + ], + "logprob_reward": -0.014211636036634445, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211636036634445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15905383229255676, + 0.07928341627120972, + 0.13266412913799286, + 0.06064530462026596, + 0.09071902185678482, + 0.060670241713523865, + 0.08669179677963257, + 0.08681637048721313, + 0.06348960846662521, + 0.17996631562709808 + ], + "credit_map": [ + 0.012744988314807415, + 0.02396954782307148, + 0.026903187856078148, + 0.013187883421778679, + 0.03360503539443016, + 0.019114194437861443, + 0.015004368498921394, + 0.014524624682962894, + 0.009038458578288555, + 0.011958446353673935, + 0.0063248383812606335, + 0.009430780075490475, + 0.01806831918656826, + 0.007637830451130867, + 0.010063888505101204, + 0.006760855205357075, + 0.014418712817132473, + 0.01814114861190319, + 0.01515539363026619, + 0.018190160393714905, + 0.02405657432973385, + 0.010755241848528385, + 0.016841813921928406, + 0.015105079859495163, + 0.007575818803161383, + 0.009527139365673065, + 0.010049881413578987, + 0.007118096575140953, + 0.00892809871584177, + 0.0055614314042031765, + 0.005482328124344349, + 0.006402511615306139, + 0.013789570890367031, + 0.008890165016055107, + 0.009360307827591896, + 0.009668860584497452, + 0.016143713146448135, + 0.01201315876096487, + 0.01047524530440569, + 0.010378003120422363, + 0.004268391989171505, + 0.005154251120984554, + 0.006936530116945505, + 0.00846483837813139, + 0.005868976004421711, + 0.010660860687494278, + 0.009872839786112309, + 0.009443558752536774, + 0.010814898647367954, + 0.008774561807513237, + 0.014497543685138226, + 0.010391471907496452, + 0.007104911375790834, + 0.011872261762619019, + 0.010667145252227783, + 0.012568996287882328, + 0.010447570122778416, + 0.011034002527594566, + 0.011621392332017422, + 0.00883525237441063, + 0.010810277424752712, + 0.011709451675415039, + 0.012998376972973347, + 0.009360045194625854, + 0.00791818555444479, + 0.006356862839311361, + 0.008845795877277851, + 0.005950436927378178, + 0.006235035602003336, + 0.010431552305817604, + 0.006884206552058458, + 0.010867533273994923, + 0.02934972569346428, + 0.01581333577632904, + 0.029574749991297722, + 0.0214906744658947, + 0.013520494103431702, + 0.020901724696159363, + 0.0163260530680418, + 0.03298955783247948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.014211636036634445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle an upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in a black-and-white aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a female athlete shown at the peak of a leap with her right arm extended back for a smash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the volleyball net visible in the foreground below the athlete?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427762657403946 + ], + "opposite_logprob_tokens": [ + -2.0177762508392334 + ], + "logprob_reward": -0.1427762657403946, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1427762657403946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43134668469429016, + 0.024667441844940186, + 0.02263413742184639, + 0.08219154179096222, + 0.04599003866314888, + 0.06402475386857986, + 0.04061634838581085, + 0.01706225797533989, + 0.06658072769641876, + 0.2048860341310501 + ], + "credit_map": [ + 0.016017034649848938, + 0.012277860194444656, + 0.017527854070067406, + 0.011696219444274902, + 0.023401949554681778, + 0.1745273619890213, + 0.010311096906661987, + 0.16558729112148285, + 0.002784992568194866, + 0.003114548744633794, + 0.0033097611740231514, + 0.0034830414224416018, + 0.003910045139491558, + 0.002896040678024292, + 0.0028388698119670153, + 0.002330140443518758, + 0.0025256157387048006, + 0.0019075522432103753, + 0.0031677233055233955, + 0.004394054878503084, + 0.0028335885144770145, + 0.0017704839119687676, + 0.003655055770650506, + 0.0023800628259778023, + 0.010248448699712753, + 0.011635556817054749, + 0.017160898074507713, + 0.009964274242520332, + 0.008414889685809612, + 0.007110394537448883, + 0.008585648611187935, + 0.009071427397429943, + 0.01078824046999216, + 0.005504837725311518, + 0.005407983437180519, + 0.0037775246892124414, + 0.003935774322599173, + 0.004575884435325861, + 0.003794177435338497, + 0.008205615915358067, + 0.007469843607395887, + 0.008466985076665878, + 0.009273247793316841, + 0.006287467200309038, + 0.006528547033667564, + 0.009172999300062656, + 0.0055570704862475395, + 0.011268586851656437, + 0.005265067797154188, + 0.004043301101773977, + 0.006503784563392401, + 0.0029273221734911203, + 0.0030178988818079233, + 0.006936052814126015, + 0.0036209726240485907, + 0.008301948197185993, + 0.0017671337118372321, + 0.0019332998199388385, + 0.0024258990306407213, + 0.0017625904874876142, + 0.0021799779497087, + 0.0024836875963956118, + 0.0022195857018232346, + 0.002290084259584546, + 0.008587753400206566, + 0.004232664126902819, + 0.011782774701714516, + 0.004162722732871771, + 0.004086092580109835, + 0.013188574463129044, + 0.004998460877686739, + 0.01554168201982975, + 0.04722831770777702, + 0.017591742798686028, + 0.030267983675003052, + 0.019616691395640373, + 0.013662002980709076, + 0.017490258440375328, + 0.015132114291191101, + 0.04389692842960358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.1427762657403946 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_6/index_tensor([7174, 7174])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_6/index_tensor([7174, 7174])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1eb55411d96f351e2d7b806db931203234454ee8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_6/index_tensor([7174, 7174])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a medium close-up shot from a slightly upward angle, perfectly recreating Annie Leibovitz's signature dramatic lighting style, a shaved-headed, stern-faced master of Southern Fist martial arts takes center stage. A powerful, sharply defined, cold-toned spotlight slants down from above and sideways, precisely sculpting the rock-like bulges of his bronzed upper body, his latissimus dorsi and deltoids. The veins in his arms, bulging with the force of his intense exertion, are clearly visible, while the other side of his body is completely shrouded in deep, intense shadow, creating a dramatic, high-contrast effect. The background, completely blurred by the extremely shallow depth of field, is reduced to a deep, nearly black blur. The grandmaster's eyes were as sharp as an eagle. In the slow motion captured by the high-speed camera, his body's center of gravity suddenly sank, and he exerted force on his waist, driving his clenched right fist forward with the force of thunder. The moment his fist touched a heavy wooden stake arm with a patina of time, the stake exploded, and countless wood chips and fibers of various sizes exploded in the air. Each fragment was clearly visible under the sharp focus, as if time had frozen. The flying wood chips reflected bits of golden light under the strong light, forming a sharp contrast with the dark background. The whole picture was full of the ultimate beauty of concentration, strength and instantaneous burst.", + "seed": 19676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_6/index_7174_seed_19676_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08801377974450589, + "vlm_logprob": -0.08801377974450589, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.08801377974450589, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06616256386041641 + ], + "opposite_logprob_tokens": [ + -2.816162586212158 + ], + "logprob_reward": -0.06616256386041641, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06616256386041641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22122950851917267, + 0.07465658336877823, + 0.15532052516937256, + 0.05364314094185829, + 0.11593430489301682, + 0.09983885288238525, + 0.08918608725070953, + 0.09342875331640244, + 0.04927730932831764, + 0.04748498275876045 + ], + "credit_map": [ + 0.035875242203474045, + 0.018901314586400986, + 0.020138539373874664, + 0.03473169356584549, + 0.026754681020975113, + 0.01544255018234253, + 0.047770336270332336, + 0.021615128964185715, + 0.01091651152819395, + 0.009058327414095402, + 0.008230834268033504, + 0.010364827699959278, + 0.007504155859351158, + 0.0050027985125780106, + 0.01438106782734394, + 0.009198050945997238, + 0.011775358580052853, + 0.010841047391295433, + 0.03522122651338577, + 0.05429425090551376, + 0.004715401213616133, + 0.007637420669198036, + 0.011975481174886227, + 0.018860332667827606, + 0.00535444263368845, + 0.007017241790890694, + 0.009539320133626461, + 0.009160724468529224, + 0.003640978131443262, + 0.007479678373783827, + 0.005923371762037277, + 0.005527375265955925, + 0.017964042723178864, + 0.010756304487586021, + 0.01997441053390503, + 0.015057743526995182, + 0.007050087209790945, + 0.01695220172405243, + 0.016123255714774132, + 0.012056244537234306, + 0.013099229894578457, + 0.011912001296877861, + 0.01271097082644701, + 0.01172755565494299, + 0.008562997914850712, + 0.013762889429926872, + 0.012494828552007675, + 0.015568368136882782, + 0.009567663073539734, + 0.013925661332905293, + 0.0118250185623765, + 0.01412704773247242, + 0.010111961513757706, + 0.010194805450737476, + 0.011063920333981514, + 0.008370001800358295, + 0.016579609364271164, + 0.01102513913065195, + 0.012150568887591362, + 0.011182209476828575, + 0.011818538419902325, + 0.014733728021383286, + 0.010242330841720104, + 0.00569661520421505, + 0.00557507062330842, + 0.011428030207753181, + 0.0030631041154265404, + 0.0032366327941417694, + 0.007331063039600849, + 0.0073909941129386425, + 0.006039668805897236, + 0.0052127414382994175, + 0.00425177114084363, + 0.006001967471092939, + 0.007398062851279974, + 0.00500475475564599, + 0.006126547697931528, + 0.007097598630934954, + 0.0069910562597215176, + 0.004613219294697046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.06616256386041641 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020595386624336243 + ], + "opposite_logprob_tokens": [ + -3.8955953121185303 + ], + "logprob_reward": -0.020595386624336243, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020595386624336243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054994601756334305, + 0.05909986048936844, + 0.261953204870224, + 0.08252540975809097, + 0.23267747461795807, + 0.024917704984545708, + 0.044999584555625916, + 0.13463035225868225, + 0.05092990770936012, + 0.05327185243368149 + ], + "credit_map": [ + 0.01063547469675541, + 0.0035107783041894436, + 0.006905093789100647, + 0.008845499716699123, + 0.005292064044624567, + 0.004747869446873665, + 0.009534241631627083, + 0.005523586645722389, + 0.009542805142700672, + 0.007646320387721062, + 0.007044890429824591, + 0.007679406087845564, + 0.0065241530537605286, + 0.005273723509162664, + 0.00911684613674879, + 0.006271722726523876, + 0.04039127007126808, + 0.0198838971555233, + 0.026189150288701057, + 0.10240314900875092, + 0.009031517431139946, + 0.009022235870361328, + 0.020417336374521255, + 0.034614671021699905, + 0.009486948139965534, + 0.009091557003557682, + 0.012243373319506645, + 0.016984421759843826, + 0.005511052440851927, + 0.009541351348161697, + 0.008160607889294624, + 0.011506104841828346, + 0.032213613390922546, + 0.012742736376821995, + 0.028430882841348648, + 0.025705881416797638, + 0.016947034746408463, + 0.0704912468791008, + 0.027854127809405327, + 0.01829199120402336, + 0.003908989019691944, + 0.00243726116605103, + 0.0029860269278287888, + 0.0031807562336325645, + 0.0021899405401200056, + 0.002572560915723443, + 0.005249401088804007, + 0.002392770489677787, + 0.005751687567681074, + 0.0066904835402965546, + 0.006604604423046112, + 0.00631125969812274, + 0.003944133408367634, + 0.006581505294889212, + 0.005443618632853031, + 0.003672293620184064, + 0.028896575793623924, + 0.013530360534787178, + 0.019025031477212906, + 0.01867079921066761, + 0.01103209238499403, + 0.020709328353405, + 0.013629418797791004, + 0.009136752225458622, + 0.006811507511883974, + 0.00872297678142786, + 0.003985859453678131, + 0.004294578451663256, + 0.007034100592136383, + 0.008533625863492489, + 0.006085704080760479, + 0.005461555439978838, + 0.005641925614327192, + 0.00864187441766262, + 0.009063001722097397, + 0.006370350252836943, + 0.0050031584687530994, + 0.006691476330161095, + 0.0070149158127605915, + 0.004845153074711561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.020595386624336243 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329378016293049 + ], + "opposite_logprob_tokens": [ + -3.7732937335968018 + ], + "logprob_reward": -0.02329378016293049, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329378016293049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059457529336214066, + 0.03086109831929207, + 0.14389176666736603, + 0.09952730685472488, + 0.14682002365589142, + 0.1342601627111435, + 0.05688238516449928, + 0.2443290650844574, + 0.026899557560682297, + 0.057071078568696976 + ], + "credit_map": [ + 0.010873137041926384, + 0.004116354044526815, + 0.007858242839574814, + 0.012077012099325657, + 0.004642642568796873, + 0.0058329422026872635, + 0.006494319066405296, + 0.007562878541648388, + 0.0049017034471035, + 0.002600210253149271, + 0.0038706688210368156, + 0.005787080619484186, + 0.002759792609140277, + 0.0026986468583345413, + 0.00533673819154501, + 0.0029062596149742603, + 0.019709868356585503, + 0.009551837109029293, + 0.018153442069888115, + 0.05501804128289223, + 0.004699101205915213, + 0.006022836547344923, + 0.012926406227052212, + 0.017810236662626266, + 0.011327422223985195, + 0.009579763747751713, + 0.014709386974573135, + 0.026796428486704826, + 0.0052702827379107475, + 0.011613608337938786, + 0.009120029397308826, + 0.011110383085906506, + 0.020397096872329712, + 0.010105921886861324, + 0.02103753201663494, + 0.01799343340098858, + 0.009211575612425804, + 0.03677718713879585, + 0.02039334923028946, + 0.010903911665081978, + 0.024068113416433334, + 0.012208782136440277, + 0.016069823876023293, + 0.009609049186110497, + 0.00879951473325491, + 0.025567110627889633, + 0.014377615414559841, + 0.023560157045722008, + 0.007605332415550947, + 0.006883573718369007, + 0.0073213051073253155, + 0.005378126632422209, + 0.008056935854256153, + 0.008025528863072395, + 0.00852037500590086, + 0.005091207101941109, + 0.04113565757870674, + 0.027080919593572617, + 0.03142040967941284, + 0.034891944378614426, + 0.02872759848833084, + 0.03830469027161598, + 0.025389857590198517, + 0.017378009855747223, + 0.002852492034435272, + 0.006184902507811785, + 0.0025705823209136724, + 0.0023840356152504683, + 0.003550398861989379, + 0.0040562814101576805, + 0.0024660497438162565, + 0.0028348159976303577, + 0.004796066787093878, + 0.008480159565806389, + 0.007683211006224155, + 0.007595985196530819, + 0.0075288377702236176, + 0.007538145408034325, + 0.007638982497155666, + 0.005809685681015253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.02329378016293049 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866306856274605 + ], + "opposite_logprob_tokens": [ + -3.0486631393432617 + ], + "logprob_reward": -0.04866306856274605, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866306856274605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04573550820350647, + 0.06838855147361755, + 0.17226122319698334, + 0.1231534481048584, + 0.07026491314172745, + 0.08891579508781433, + 0.07471182197332382, + 0.10772591084241867, + 0.12912900745868683, + 0.11971376091241837 + ], + "credit_map": [ + 0.008096081204712391, + 0.003307265928015113, + 0.006859987508505583, + 0.006232092622667551, + 0.004532079212367535, + 0.00516419718042016, + 0.005527150817215443, + 0.006016660947352648, + 0.009954938665032387, + 0.008473701775074005, + 0.008405527099967003, + 0.009463601745665073, + 0.006851370446383953, + 0.006314306985586882, + 0.012141737155616283, + 0.0067833769135177135, + 0.023178059607744217, + 0.010267987847328186, + 0.021957656368613243, + 0.07163407653570175, + 0.005477631464600563, + 0.005952250678092241, + 0.012691708281636238, + 0.021101871505379677, + 0.011687302961945534, + 0.011759220622479916, + 0.021442100405693054, + 0.033738330006599426, + 0.007292683701962233, + 0.011191586032509804, + 0.00901886448264122, + 0.017023369669914246, + 0.009603481739759445, + 0.004801556468009949, + 0.008993614464998245, + 0.00962885096669197, + 0.0052614156156778336, + 0.01684960536658764, + 0.00764796556904912, + 0.007478426676243544, + 0.013384616002440453, + 0.009917961433529854, + 0.011310450732707977, + 0.009324139915406704, + 0.006918164435774088, + 0.013436255045235157, + 0.01417137123644352, + 0.010452844202518463, + 0.008304379880428314, + 0.011236018501222134, + 0.00980447418987751, + 0.016046999022364616, + 0.005579924676567316, + 0.009448939003050327, + 0.008179343305528164, + 0.006111753173172474, + 0.015656378120183945, + 0.01357191801071167, + 0.013761226087808609, + 0.016234716400504112, + 0.013540435582399368, + 0.013869778253138065, + 0.011460217647254467, + 0.009631249122321606, + 0.01258302666246891, + 0.03551037609577179, + 0.0102424630895257, + 0.010066774673759937, + 0.012622361071407795, + 0.01972905732691288, + 0.01457174215465784, + 0.013803225941956043, + 0.012033346109092236, + 0.017316365614533424, + 0.02119927667081356, + 0.013327384367585182, + 0.012064955197274685, + 0.015179803594946861, + 0.016103655099868774, + 0.01248897984623909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.04866306856274605 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813540995121002 + ], + "opposite_logprob_tokens": [ + -1.4063540697097778 + ], + "logprob_reward": -0.2813540995121002, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813540995121002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12247001379728317, + 0.03352389484643936, + 0.12277987599372864, + 0.06387198716402054, + 0.09118709713220596, + 0.04261833429336548, + 0.07129176706075668, + 0.1779678612947464, + 0.14433404803276062, + 0.12995505332946777 + ], + "credit_map": [ + 0.022306907922029495, + 0.010349084623157978, + 0.02150871604681015, + 0.018535394221544266, + 0.014214543625712395, + 0.009696714580059052, + 0.014073161408305168, + 0.011785491369664669, + 0.00462376419454813, + 0.004307572729885578, + 0.006262665148824453, + 0.004358861595392227, + 0.0028265444561839104, + 0.00205454183742404, + 0.005104266572743654, + 0.00398567970842123, + 0.011296641081571579, + 0.008301888592541218, + 0.023963376879692078, + 0.04751965403556824, + 0.0037442694883793592, + 0.005587627645581961, + 0.010586266405880451, + 0.011780155822634697, + 0.0052482872270047665, + 0.007936501875519753, + 0.010095176286995411, + 0.017592836171388626, + 0.0030202765483409166, + 0.005611443892121315, + 0.005829282570630312, + 0.008538185618817806, + 0.012338737025856972, + 0.005965714808553457, + 0.01654510572552681, + 0.010410299524664879, + 0.007606539409607649, + 0.018857482820749283, + 0.013389025814831257, + 0.006074198056012392, + 0.006822954397648573, + 0.0040428051725029945, + 0.005735741928219795, + 0.004481906071305275, + 0.0025455323047935963, + 0.008689546957612038, + 0.0058736191131174564, + 0.004426227882504463, + 0.007596845272928476, + 0.010759104043245316, + 0.009717140346765518, + 0.009548155590891838, + 0.009330147877335548, + 0.010454811155796051, + 0.008551232516765594, + 0.005334328860044479, + 0.024452688172459602, + 0.020202578976750374, + 0.022279109805822372, + 0.026151908561587334, + 0.01792229153215885, + 0.03056589886546135, + 0.020475883036851883, + 0.015917504206299782, + 0.020002467557787895, + 0.031195085495710373, + 0.013892767019569874, + 0.013303501531481743, + 0.010646803304553032, + 0.023329172283411026, + 0.014496702700853348, + 0.017467547208070755, + 0.012366035021841526, + 0.018181312829256058, + 0.020833496004343033, + 0.017732933163642883, + 0.012151026166975498, + 0.017027167603373528, + 0.018401287496089935, + 0.013261808082461357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.2813540995121002 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a medium close-up shot from a slightly upward angle, perfectly recreating Annie Leibovitz's signature dramatic lighting style, a shaved-headed, stern-faced master of Southern Fist martial arts takes center stage. A powerful, sharply defined, cold-toned spotlight slants down from above and sideways, precisely sculpting the rock-like bulges of his bronzed upper body, his latissimus dorsi and deltoids. The veins in his arms, bulging with the force of his intense exertion, are clearly visible, while the other side of his body is completely shrouded in deep, intense shadow, creating a dramatic, high-contrast effect. The background, completely blurred by the extremely shallow depth of field, is reduced to a deep, nearly black blur. The grandmaster's eyes were as sharp as an eagle. In the slow motion captured by the high-speed camera, his body's center of gravity suddenly sank, and he exerted force on his waist, driving his clenched right fist forward with the force of thunder. The moment his fist touched a heavy wooden stake arm with a patina of time, the stake exploded, and countless wood chips and fibers of various sizes exploded in the air. Each fragment was clearly visible under the sharp focus, as if time had frozen. The flying wood chips reflected bits of golden light under the strong light, forming a sharp contrast with the dark background. The whole picture was full of the ultimate beauty of concentration, strength and instantaneous burst.", + "seed": 19677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_6/index_7174_seed_19677_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.19734115526080132, + "vlm_logprob": -0.19734115526080132, + "vlm_yes_no_margin": 1.850000011920929, + "vlm_token_credit": -0.19734115526080132, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08261578530073166 + ], + "opposite_logprob_tokens": [ + -2.582615852355957 + ], + "logprob_reward": -0.08261578530073166, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08261578530073166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23410628736019135, + 0.04975910857319832, + 0.07361477613449097, + 0.0726422443985939, + 0.05651656538248062, + 0.043726056814193726, + 0.09177111089229584, + 0.10513343662023544, + 0.14051373302936554, + 0.1322166621685028 + ], + "credit_map": [ + 0.024594392627477646, + 0.03013770841062069, + 0.017296340316534042, + 0.02704404667019844, + 0.011623590253293514, + 0.05247228220105171, + 0.013858439400792122, + 0.057079482823610306, + 0.003859610529616475, + 0.007601999677717686, + 0.005886307917535305, + 0.007202712818980217, + 0.0050072865560650826, + 0.007726237177848816, + 0.00532021326944232, + 0.007154739927500486, + 0.0039108628407120705, + 0.00755337905138731, + 0.005741972476243973, + 0.017069736495614052, + 0.004286512732505798, + 0.01343206875026226, + 0.009441022761166096, + 0.012179228477180004, + 0.0034129994455724955, + 0.006216396577656269, + 0.005340962670743465, + 0.009739424102008343, + 0.00679752929136157, + 0.01481880247592926, + 0.011708959005773067, + 0.014607170596718788, + 0.0029723949264734983, + 0.005926941987127066, + 0.005394560284912586, + 0.012350324541330338, + 0.003914857283234596, + 0.009598876349627972, + 0.006339594721794128, + 0.010019014589488506, + 0.0024192091077566147, + 0.003782751504331827, + 0.005274442955851555, + 0.006045537069439888, + 0.002517283195629716, + 0.008116704411804676, + 0.006341344676911831, + 0.0092287827283144, + 0.007455692160874605, + 0.008668381720781326, + 0.019674355164170265, + 0.013803455978631973, + 0.004708712920546532, + 0.014925388619303703, + 0.010693875141441822, + 0.011841250583529472, + 0.011158169247210026, + 0.016117749735713005, + 0.017901983112096786, + 0.022235028445720673, + 0.006289420649409294, + 0.009238529950380325, + 0.01028580591082573, + 0.011906743049621582, + 0.018439196050167084, + 0.034392401576042175, + 0.018764376640319824, + 0.018941203132271767, + 0.01699388585984707, + 0.008121544495224953, + 0.013443335890769958, + 0.01141778938472271, + 0.01941586658358574, + 0.019365865737199783, + 0.02265734225511551, + 0.017777753993868828, + 0.02031121961772442, + 0.008817559108138084, + 0.01075960136950016, + 0.013111449778079987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.08261578530073166 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048674535006284714 + ], + "opposite_logprob_tokens": [ + -3.0486745834350586 + ], + "logprob_reward": -0.048674535006284714, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048674535006284714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20693914592266083, + 0.06801530718803406, + 0.10245067626237869, + 0.056217607110738754, + 0.08532025665044785, + 0.07862218469381332, + 0.03201331943273544, + 0.22494162619113922, + 0.09096090495586395, + 0.054518990218639374 + ], + "credit_map": [ + 0.01925879530608654, + 0.01141446828842163, + 0.008706589229404926, + 0.015382708050310612, + 0.007910105399787426, + 0.06683293730020523, + 0.010735559277236462, + 0.06669799238443375, + 0.0066629135981202126, + 0.007252316921949387, + 0.007034800481051207, + 0.010375323705375195, + 0.004866357892751694, + 0.01166750118136406, + 0.006026175804436207, + 0.014129922725260258, + 0.0034582901280373335, + 0.00833454541862011, + 0.007195728365331888, + 0.020793816074728966, + 0.005530193448066711, + 0.02046145871281624, + 0.011036432348191738, + 0.025640219449996948, + 0.004527477081865072, + 0.006141957826912403, + 0.004237279295921326, + 0.009134463965892792, + 0.003898189403116703, + 0.011064560152590275, + 0.006498218048363924, + 0.010715467855334282, + 0.00583155220374465, + 0.00740873534232378, + 0.006818216759711504, + 0.01881355047225952, + 0.004577426239848137, + 0.014566227793693542, + 0.009546513669192791, + 0.017758043482899666, + 0.005212919320911169, + 0.006920219399034977, + 0.006583536043763161, + 0.010118863545358181, + 0.0029857533518224955, + 0.018022796139121056, + 0.005646076984703541, + 0.02313203364610672, + 0.003367172321304679, + 0.0052018254064023495, + 0.004199847113341093, + 0.0049363961443305016, + 0.001852894783951342, + 0.004647268448024988, + 0.0031239709351211786, + 0.004683946259319782, + 0.03998889401555061, + 0.020698370411992073, + 0.017147010192275047, + 0.08957215398550034, + 0.012119225226342678, + 0.009031242690980434, + 0.010516189970076084, + 0.025868553668260574, + 0.012306890450417995, + 0.02046465314924717, + 0.011995268054306507, + 0.013105813413858414, + 0.01159524917602539, + 0.005027292296290398, + 0.008440743200480938, + 0.008025002665817738, + 0.007794756907969713, + 0.006460575386881828, + 0.009996270760893822, + 0.005007236264646053, + 0.00943182222545147, + 0.004038953222334385, + 0.004091287963092327, + 0.007698092609643936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.048674535006284714 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.313493013381958 + ], + "opposite_logprob_tokens": [ + -1.313493013381958 + ], + "logprob_reward": -0.313493013381958, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.313493013381958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13005246222019196, + 0.1286075860261917, + 0.07915281504392624, + 0.06512685120105743, + 0.12758968770503998, + 0.04450840502977371, + 0.04254617169499397, + 0.14785735309123993, + 0.13510790467262268, + 0.09945080429315567 + ], + "credit_map": [ + 0.01616942137479782, + 0.010506000369787216, + 0.006490637082606554, + 0.012834708206355572, + 0.005431165918707848, + 0.03583967313170433, + 0.00553445890545845, + 0.037246402353048325, + 0.007497876882553101, + 0.019732903689146042, + 0.010752108879387379, + 0.02013218402862549, + 0.008653859607875347, + 0.02281365916132927, + 0.014061160385608673, + 0.02496384084224701, + 0.004424028564244509, + 0.010547440499067307, + 0.006910908501595259, + 0.01680997759103775, + 0.004704248625785112, + 0.013655764050781727, + 0.007134404964745045, + 0.014966050162911415, + 0.005366154480725527, + 0.007374466396868229, + 0.0054504829458892345, + 0.007409010082483292, + 0.0057568722404539585, + 0.012459304183721542, + 0.00818681251257658, + 0.013123749755322933, + 0.0050395941361784935, + 0.010247515514492989, + 0.009096186608076096, + 0.029849138110876083, + 0.006158252712339163, + 0.02566560171544552, + 0.013200764544308186, + 0.028332630172371864, + 0.004854693077504635, + 0.0031440691091120243, + 0.00488889217376709, + 0.0048973215743899345, + 0.0015579249011352658, + 0.010734472423791885, + 0.0033030991908162832, + 0.011127933859825134, + 0.007697971537709236, + 0.004501459654420614, + 0.009068373590707779, + 0.005206917878240347, + 0.002188001526519656, + 0.004521000199019909, + 0.0033416824880987406, + 0.006020762491971254, + 0.022743213921785355, + 0.011902907863259315, + 0.0164411012083292, + 0.04674845188856125, + 0.013445565477013588, + 0.010534350760281086, + 0.009970058687031269, + 0.01607169210910797, + 0.01658632420003414, + 0.014533266425132751, + 0.01645559072494507, + 0.014407290145754814, + 0.02542218193411827, + 0.014596935361623764, + 0.01267398800700903, + 0.020432325080037117, + 0.015205360017716885, + 0.01076795719563961, + 0.016307266429066658, + 0.009750720113515854, + 0.014205286279320717, + 0.00975096970796585, + 0.012733052484691143, + 0.01073018554598093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.313493013381958 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295502632856369 + ], + "opposite_logprob_tokens": [ + -2.237955093383789 + ], + "logprob_reward": -0.11295502632856369, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295502632856369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09954517334699631, + 0.09531008452177048, + 0.07774858176708221, + 0.06396332383155823, + 0.11893990635871887, + 0.0347207710146904, + 0.1453799605369568, + 0.14795486629009247, + 0.09086737036705017, + 0.1255699247121811 + ], + "credit_map": [ + 0.01409104373306036, + 0.009030863642692566, + 0.005672751925885677, + 0.013505521230399609, + 0.007030409295111895, + 0.02115454524755478, + 0.005325549282133579, + 0.02373449318110943, + 0.006305264309048653, + 0.013194281607866287, + 0.01626039855182171, + 0.01762872003018856, + 0.007905404083430767, + 0.012075559236109257, + 0.01071176491677761, + 0.01122868712991476, + 0.004759589210152626, + 0.0066951485350728035, + 0.006485237739980221, + 0.017035525292158127, + 0.005396787542849779, + 0.015499565750360489, + 0.008554182946681976, + 0.013322548940777779, + 0.0030835759826004505, + 0.005244783591479063, + 0.004136673174798489, + 0.011299358680844307, + 0.003712050849571824, + 0.012884876690804958, + 0.00840403325855732, + 0.015197972767055035, + 0.007778454106301069, + 0.011148069985210896, + 0.012278260663151741, + 0.024214979261159897, + 0.006527605000883341, + 0.020906522870063782, + 0.01659517176449299, + 0.01949085108935833, + 0.00205541611649096, + 0.0038416204042732716, + 0.004864335060119629, + 0.003907451871782541, + 0.001262050005607307, + 0.007630003150552511, + 0.0027205899823457003, + 0.008439303375780582, + 0.011918697506189346, + 0.01479098480194807, + 0.025353793054819107, + 0.020112115889787674, + 0.008262872695922852, + 0.025045428425073624, + 0.02134104259312153, + 0.01855503022670746, + 0.026083089411258698, + 0.01103194896131754, + 0.011099645867943764, + 0.06013015657663345, + 0.005257189739495516, + 0.007393051404505968, + 0.009475256316363811, + 0.017484532669186592, + 0.01187284104526043, + 0.01067796628922224, + 0.01290072314441204, + 0.015002788044512272, + 0.009668251499533653, + 0.007689510937780142, + 0.007785814348608255, + 0.015269479714334011, + 0.0178851131349802, + 0.011776449158787727, + 0.022413993254303932, + 0.008052373304963112, + 0.022681567817926407, + 0.009096059016883373, + 0.010126338340342045, + 0.02353803440928459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.11295502632856369 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289674162864685 + ], + "opposite_logprob_tokens": [ + -1.0539674758911133 + ], + "logprob_reward": -0.4289674162864685, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.4289674162864685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18967914581298828, + 0.06952684372663498, + 0.05492934584617615, + 0.04708217829465866, + 0.03906608745455742, + 0.042280904948711395, + 0.03661702945828438, + 0.3457561135292053, + 0.09166990965604782, + 0.08339241147041321 + ], + "credit_map": [ + 0.03420177102088928, + 0.015809547156095505, + 0.011141786351799965, + 0.015449579805135727, + 0.0101440055295825, + 0.041565414518117905, + 0.010088457725942135, + 0.05127858370542526, + 0.004493469372391701, + 0.008402084931731224, + 0.008015351369976997, + 0.009390230290591717, + 0.004102567210793495, + 0.013500651344656944, + 0.005775218829512596, + 0.01584727317094803, + 0.0024560019373893738, + 0.00530846556648612, + 0.004921970888972282, + 0.01213742233812809, + 0.00275534694083035, + 0.010244665667414665, + 0.005119980778545141, + 0.011985492892563343, + 0.002874956000596285, + 0.003947453107684851, + 0.004539386834949255, + 0.006823682691901922, + 0.0033689788542687893, + 0.009233247488737106, + 0.007047031540423632, + 0.00924744177609682, + 0.0017010429874062538, + 0.003139386884868145, + 0.005060968454927206, + 0.010421509854495525, + 0.002724047750234604, + 0.006266155280172825, + 0.005050448700785637, + 0.004702524747699499, + 0.0027097256388515234, + 0.004574277438223362, + 0.0037686617579311132, + 0.005782486405223608, + 0.0019139280775561929, + 0.00894266925752163, + 0.00428419467061758, + 0.01030496321618557, + 0.002633911557495594, + 0.0035295053385198116, + 0.005894384812563658, + 0.005032997578382492, + 0.0020587914623320103, + 0.00846849475055933, + 0.0036106412298977375, + 0.00538830179721117, + 0.06502632796764374, + 0.025355923920869827, + 0.02546071819961071, + 0.14799317717552185, + 0.011704396456480026, + 0.011923754587769508, + 0.017870448529720306, + 0.04042137786746025, + 0.013627471402287483, + 0.012584828771650791, + 0.010498030111193657, + 0.009624741040170193, + 0.01151780504733324, + 0.009205753915011883, + 0.008831248618662357, + 0.015780027955770493, + 0.012776270508766174, + 0.007638983894139528, + 0.01065047550946474, + 0.006227367091923952, + 0.013500633649528027, + 0.008299949578940868, + 0.0072061289101839066, + 0.017092613503336906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.4289674162864685 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a medium close-up shot from a slightly upward angle, perfectly recreating Annie Leibovitz's signature dramatic lighting style, a shaved-headed, stern-faced master of Southern Fist martial arts takes center stage. A powerful, sharply defined, cold-toned spotlight slants down from above and sideways, precisely sculpting the rock-like bulges of his bronzed upper body, his latissimus dorsi and deltoids. The veins in his arms, bulging with the force of his intense exertion, are clearly visible, while the other side of his body is completely shrouded in deep, intense shadow, creating a dramatic, high-contrast effect. The background, completely blurred by the extremely shallow depth of field, is reduced to a deep, nearly black blur. The grandmaster's eyes were as sharp as an eagle. In the slow motion captured by the high-speed camera, his body's center of gravity suddenly sank, and he exerted force on his waist, driving his clenched right fist forward with the force of thunder. The moment his fist touched a heavy wooden stake arm with a patina of time, the stake exploded, and countless wood chips and fibers of various sizes exploded in the air. Each fragment was clearly visible under the sharp focus, as if time had frozen. The flying wood chips reflected bits of golden light under the strong light, forming a sharp contrast with the dark background. The whole picture was full of the ultimate beauty of concentration, strength and instantaneous burst.", + "seed": 19678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_6/index_7174_seed_19678_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.048913934454321864, + "vlm_logprob": -0.048913934454321864, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.048913934454321864, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04062395542860031 + ], + "opposite_logprob_tokens": [ + -3.290623903274536 + ], + "logprob_reward": -0.04062395542860031, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04062395542860031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11983062326908112, + 0.03841179236769676, + 0.04591866582632065, + 0.13264510035514832, + 0.20569613575935364, + 0.13692529499530792, + 0.1391250640153885, + 0.061330024152994156, + 0.08492087572813034, + 0.03519637882709503 + ], + "credit_map": [ + 0.0258598942309618, + 0.013984856195747852, + 0.01049233041703701, + 0.020364630967378616, + 0.010578138753771782, + 0.01676180586218834, + 0.013194708153605461, + 0.008594265207648277, + 0.006043971050530672, + 0.006153825204819441, + 0.004902029875665903, + 0.004374863114207983, + 0.0037475985009223223, + 0.005210883915424347, + 0.004961942322552204, + 0.0030166779179126024, + 0.007289815694093704, + 0.004495467524975538, + 0.005547808948904276, + 0.006557483691722155, + 0.004252845887094736, + 0.00707628671079874, + 0.0046614776365458965, + 0.0060374801978468895, + 0.01354509312659502, + 0.0069428193382918835, + 0.0056219277903437614, + 0.008454633876681328, + 0.03465373069047928, + 0.03541279956698418, + 0.008570155128836632, + 0.0194439385086298, + 0.020722754299640656, + 0.03122636489570141, + 0.03350149095058441, + 0.02515247091650963, + 0.024311291053891182, + 0.023290809243917465, + 0.023933401331305504, + 0.023557554930448532, + 0.021380223333835602, + 0.013251852244138718, + 0.019928812980651855, + 0.014977479353547096, + 0.01148109883069992, + 0.01758764311671257, + 0.014380836859345436, + 0.023937344551086426, + 0.01876720041036606, + 0.016810735687613487, + 0.009923645295202732, + 0.0193618331104517, + 0.009761108085513115, + 0.013652108609676361, + 0.033876679837703705, + 0.01697174832224846, + 0.009485808201134205, + 0.010058125481009483, + 0.007327692117542028, + 0.005628598853945732, + 0.005713227670639753, + 0.009511806070804596, + 0.009039010852575302, + 0.004565755371004343, + 0.030516909435391426, + 0.006770663894712925, + 0.008237840607762337, + 0.007721904199570417, + 0.005089497193694115, + 0.013534660451114178, + 0.010057547129690647, + 0.0029918558429926634, + 0.006867787800729275, + 0.0038330298848450184, + 0.0038826847448945045, + 0.0034226563293486834, + 0.002826889744028449, + 0.004929809831082821, + 0.006491522770375013, + 0.0029419988859444857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.04062395542860031 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255360059440136 + ], + "opposite_logprob_tokens": [ + -4.387553691864014 + ], + "logprob_reward": -0.01255360059440136, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255360059440136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050043076276779175, + 0.1233147457242012, + 0.09807891398668289, + 0.25105780363082886, + 0.06966444849967957, + 0.08552679419517517, + 0.08999159187078476, + 0.15059389173984528, + 0.026956871151924133, + 0.05477190390229225 + ], + "credit_map": [ + 0.008801981806755066, + 0.005546207074075937, + 0.0035171820782124996, + 0.008557790890336037, + 0.005006698425859213, + 0.007449273485690355, + 0.007148907519876957, + 0.004015038721263409, + 0.01590212620794773, + 0.022164082154631615, + 0.015100188553333282, + 0.021804267540574074, + 0.008834529668092728, + 0.01603405363857746, + 0.013247239403426647, + 0.010228258557617664, + 0.014810550957918167, + 0.010841773822903633, + 0.009239962324500084, + 0.012308059260249138, + 0.011364453472197056, + 0.016642844304442406, + 0.00982183963060379, + 0.013049435801804066, + 0.03666945919394493, + 0.008573276922106743, + 0.009375223889946938, + 0.0136094456538558, + 0.037859417498111725, + 0.10402552783489227, + 0.010378398932516575, + 0.03056705743074417, + 0.008048143237829208, + 0.011535911820828915, + 0.007591503206640482, + 0.012320203706622124, + 0.007271730341017246, + 0.0077019124291837215, + 0.005620778538286686, + 0.009574261493980885, + 0.014736763201653957, + 0.006189264822751284, + 0.011458635330200195, + 0.011273179203271866, + 0.00752277160063386, + 0.0077546280808746815, + 0.012107373215258121, + 0.014484175480902195, + 0.013354601338505745, + 0.00722505385056138, + 0.00699224229902029, + 0.013156202621757984, + 0.005632887128740549, + 0.00937587209045887, + 0.025939051061868668, + 0.008315681479871273, + 0.02663642168045044, + 0.010441090911626816, + 0.038516417145729065, + 0.008339008316397667, + 0.010062883608043194, + 0.027465255931019783, + 0.01820358633995056, + 0.010929211974143982, + 0.009792642667889595, + 0.002110616769641638, + 0.0028552310541272163, + 0.0027477180119603872, + 0.001282398123294115, + 0.004914750345051289, + 0.0024170675314962864, + 0.0008364492678083479, + 0.011561030521988869, + 0.007136915344744921, + 0.006641363725066185, + 0.0070330421440303326, + 0.005176011007279158, + 0.008120848797261715, + 0.00537694338709116, + 0.0037257506046444178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.01255360059440136 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865330085158348 + ], + "opposite_logprob_tokens": [ + -3.0486533641815186 + ], + "logprob_reward": -0.04865330085158348, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865330085158348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08018331974744797, + 0.07747267186641693, + 0.06734061986207962, + 0.08799663186073303, + 0.0667680948972702, + 0.07007119059562683, + 0.06145298853516579, + 0.3930123448371887, + 0.024098051711916924, + 0.07160410284996033 + ], + "credit_map": [ + 0.019791990518569946, + 0.008664660155773163, + 0.007357362657785416, + 0.00839375238865614, + 0.006094960495829582, + 0.013623512350022793, + 0.010082201100885868, + 0.006174881011247635, + 0.012184311635792255, + 0.010813341476023197, + 0.009839646518230438, + 0.013720675371587276, + 0.005658420734107494, + 0.010424590669572353, + 0.009347834624350071, + 0.005483852233737707, + 0.010483571328222752, + 0.00941802654415369, + 0.005534561350941658, + 0.007971048355102539, + 0.007368205115199089, + 0.012251880019903183, + 0.004835992585867643, + 0.00947733037173748, + 0.014962965622544289, + 0.0026235675904899836, + 0.0032867940608412027, + 0.0059452420100569725, + 0.010556940920650959, + 0.036000773310661316, + 0.003795020282268524, + 0.01082533411681652, + 0.0076155392453074455, + 0.007743685506284237, + 0.0060201771557331085, + 0.008796787820756435, + 0.008576066233217716, + 0.008554604835808277, + 0.009818228892982006, + 0.009643006138503551, + 0.010436264798045158, + 0.009627659805119038, + 0.01030055433511734, + 0.00867916364222765, + 0.006810682825744152, + 0.009277638979256153, + 0.006938820704817772, + 0.008000407367944717, + 0.008709179237484932, + 0.005216770805418491, + 0.007944435812532902, + 0.010506737977266312, + 0.003454343881458044, + 0.005988060496747494, + 0.013202453032135963, + 0.006431007757782936, + 0.06840744614601135, + 0.04144592955708504, + 0.09499820321798325, + 0.0232577882707119, + 0.029182545840740204, + 0.06638776510953903, + 0.03898508474230766, + 0.030347559601068497, + 0.006656081415712833, + 0.0016514199087396264, + 0.002748003927990794, + 0.002041173866018653, + 0.0016394390258938074, + 0.004343491047620773, + 0.004073064308613539, + 0.0009453782113268971, + 0.015805399045348167, + 0.0099727688357234, + 0.0085518229752779, + 0.008034384809434414, + 0.006171809509396553, + 0.010777704417705536, + 0.0074600763618946075, + 0.004830133635550737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.04865330085158348 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980167791247368 + ], + "opposite_logprob_tokens": [ + -3.529801607131958 + ], + "logprob_reward": -0.02980167791247368, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980167791247368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03103330172598362, + 0.02752271108329296, + 0.06669972836971283, + 0.48639804124832153, + 0.12042365968227386, + 0.1583360880613327, + 0.04086139053106308, + 0.01831434667110443, + 0.0324065238237381, + 0.018004247918725014 + ], + "credit_map": [ + 0.007143697701394558, + 0.0034157722257077694, + 0.0018813841743394732, + 0.002884499030187726, + 0.003474004101008177, + 0.004858627449721098, + 0.004087727051228285, + 0.0032875919714570045, + 0.0029595589730888605, + 0.005345339421182871, + 0.002914831042289734, + 0.004880738444626331, + 0.001994046149775386, + 0.0038754078559577465, + 0.0026911753229796886, + 0.002861612942069769, + 0.008798005059361458, + 0.0058760871179401875, + 0.010316089726984501, + 0.008945301175117493, + 0.004862175788730383, + 0.009917146526277065, + 0.005719240289181471, + 0.01226568128913641, + 0.06223487854003906, + 0.014482575468719006, + 0.015374120324850082, + 0.026710495352745056, + 0.08783700317144394, + 0.19654537737369537, + 0.019189294427633286, + 0.06402425467967987, + 0.013069793581962585, + 0.01377534493803978, + 0.01288851909339428, + 0.013918564654886723, + 0.014086906798183918, + 0.019792014732956886, + 0.017167478799819946, + 0.01572503335773945, + 0.021536996588110924, + 0.01609048619866371, + 0.026884863153100014, + 0.021022042259573936, + 0.01636076159775257, + 0.01758662983775139, + 0.018344935029745102, + 0.02050935849547386, + 0.005592585541307926, + 0.004778787959367037, + 0.004990156274288893, + 0.005407824646681547, + 0.002898154081776738, + 0.003942835610359907, + 0.007625377271324396, + 0.0056256684474647045, + 0.0026774315629154444, + 0.0023676210548728704, + 0.0022790804505348206, + 0.0017175079556182027, + 0.0019042458152398467, + 0.0026340694166719913, + 0.0034096939489245415, + 0.0013246976304799318, + 0.010508510284125805, + 0.002588954521343112, + 0.00422701658681035, + 0.0024296478368341923, + 0.002327813534066081, + 0.005630129016935825, + 0.0036406784784048796, + 0.0010537734488025308, + 0.0032334430143237114, + 0.002514396794140339, + 0.0019790686201304197, + 0.0017046118155121803, + 0.0015267106937244534, + 0.0022750524803996086, + 0.003101494163274765, + 0.0016694697551429272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.02980167791247368 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293713748455048 + ], + "opposite_logprob_tokens": [ + -2.2379372119903564 + ], + "logprob_reward": -0.11293713748455048, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293713748455048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051951151341199875, + 0.07923509925603867, + 0.07539556175470352, + 0.21721036732196808, + 0.05587126687169075, + 0.22585581243038177, + 0.06906910240650177, + 0.14194999635219574, + 0.045497022569179535, + 0.0379645936191082 + ], + "credit_map": [ + 0.011803760193288326, + 0.005917717237025499, + 0.003395891049876809, + 0.006883507128804922, + 0.005851726979017258, + 0.007788877934217453, + 0.006954180542379618, + 0.0033554930705577135, + 0.01249968446791172, + 0.011137296445667744, + 0.009341737255454063, + 0.012260010465979576, + 0.007047906052321196, + 0.010017932392656803, + 0.010755005292594433, + 0.006175534334033728, + 0.012205068953335285, + 0.00807976070791483, + 0.007839146070182323, + 0.008313581347465515, + 0.008140754885971546, + 0.012455805204808712, + 0.006055336911231279, + 0.01230611652135849, + 0.01881406456232071, + 0.010650084353983402, + 0.00795164704322815, + 0.013650541193783283, + 0.055689092725515366, + 0.07022896409034729, + 0.011704375967383385, + 0.028521627187728882, + 0.006857309956103563, + 0.00473432382568717, + 0.0055882311426103115, + 0.009944812394678593, + 0.005061520263552666, + 0.006970657035708427, + 0.007895175367593765, + 0.008819244801998138, + 0.03668619319796562, + 0.025129441171884537, + 0.02580721490085125, + 0.025434860959649086, + 0.021937409415841103, + 0.026080120354890823, + 0.033211883157491684, + 0.031568728387355804, + 0.009399103932082653, + 0.007055990397930145, + 0.008185655809938908, + 0.01144376304000616, + 0.005231829825788736, + 0.007074175868183374, + 0.012131459079682827, + 0.008547130972146988, + 0.02510751411318779, + 0.01370745524764061, + 0.03476720675826073, + 0.00826773140579462, + 0.010986385866999626, + 0.02507854998111725, + 0.013576374389231205, + 0.010458793491125107, + 0.01773308962583542, + 0.003839690238237381, + 0.0043992153368890285, + 0.0036047217436134815, + 0.0023199899587780237, + 0.008407654240727425, + 0.0038187901955097914, + 0.0013738753041252494, + 0.008789029903709888, + 0.0041937194764614105, + 0.004537153523415327, + 0.0033999402076005936, + 0.0028305132873356342, + 0.006197249051183462, + 0.004999288823455572, + 0.0030177030712366104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 7 + ], + "selected_reward": -0.11293713748455048 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a medium close-up shot from a slightly upward angle, perfectly recreating Annie Leibovitz's signature dramatic lighting style, a shaved-headed, stern-faced master of Southern Fist martial arts takes center stage. A powerful, sharply defined, cold-toned spotlight slants down from above and sideways, precisely sculpting the rock-like bulges of his bronzed upper body, his latissimus dorsi and deltoids. The veins in his arms, bulging with the force of his intense exertion, are clearly visible, while the other side of his body is completely shrouded in deep, intense shadow, creating a dramatic, high-contrast effect. The background, completely blurred by the extremely shallow depth of field, is reduced to a deep, nearly black blur. The grandmaster's eyes were as sharp as an eagle. In the slow motion captured by the high-speed camera, his body's center of gravity suddenly sank, and he exerted force on his waist, driving his clenched right fist forward with the force of thunder. The moment his fist touched a heavy wooden stake arm with a patina of time, the stake exploded, and countless wood chips and fibers of various sizes exploded in the air. Each fragment was clearly visible under the sharp focus, as if time had frozen. The flying wood chips reflected bits of golden light under the strong light, forming a sharp contrast with the dark background. The whole picture was full of the ultimate beauty of concentration, strength and instantaneous burst.", + "seed": 19679, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_6/index_7174_seed_19679_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06831971965730191, + "vlm_logprob": -0.06831971965730191, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.06831971965730191, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0767851173877716 + ], + "opposite_logprob_tokens": [ + -2.701785087585449 + ], + "logprob_reward": -0.0767851173877716, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0767851173877716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.228543221950531, + 0.06621460616588593, + 0.12842193245887756, + 0.0602937713265419, + 0.07848919183015823, + 0.09904714673757553, + 0.043881915509700775, + 0.06453162431716919, + 0.18105436861515045, + 0.04952216520905495 + ], + "credit_map": [ + 0.026292769238352776, + 0.05731601640582085, + 0.03724002093076706, + 0.017182322219014168, + 0.021265847608447075, + 0.023051206022500992, + 0.02428007312119007, + 0.021914996206760406, + 0.008283739909529686, + 0.005852978676557541, + 0.009085371159017086, + 0.005130174569785595, + 0.012199285440146923, + 0.011192576959729195, + 0.00798709411174059, + 0.006483400706201792, + 0.014765066094696522, + 0.011944790370762348, + 0.010337905958294868, + 0.012090939097106457, + 0.03259316831827164, + 0.017414750531315804, + 0.01708437129855156, + 0.01219096127897501, + 0.0032867707777768373, + 0.003805874614045024, + 0.00312895979732275, + 0.004321711603552103, + 0.012216410599648952, + 0.013166064396500587, + 0.008220860734581947, + 0.012147129513323307, + 0.007340714801102877, + 0.0056718867272138596, + 0.004801553208380938, + 0.008407882414758205, + 0.009981204755604267, + 0.016872476786375046, + 0.010146276094019413, + 0.015267204493284225, + 0.014911362901329994, + 0.008796101436018944, + 0.019748400896787643, + 0.012929558753967285, + 0.011583042331039906, + 0.008357015438377857, + 0.010792148299515247, + 0.011929530650377274, + 0.006069064140319824, + 0.004829986020922661, + 0.004815759137272835, + 0.003911949694156647, + 0.008913557045161724, + 0.005041897762566805, + 0.0049955230206251144, + 0.005304183345288038, + 0.01178834680467844, + 0.007296533323824406, + 0.007867124862968922, + 0.005604881327599287, + 0.008677884936332703, + 0.011107909493148327, + 0.0055065941996872425, + 0.006682356353849173, + 0.03662171959877014, + 0.04881276562809944, + 0.040707916021347046, + 0.008809510618448257, + 0.013291928917169571, + 0.01005032192915678, + 0.009545177221298218, + 0.013215050101280212, + 0.007754657883197069, + 0.005284515675157309, + 0.006784107070416212, + 0.005491809919476509, + 0.006286893505603075, + 0.005550206173211336, + 0.0074734557420015335, + 0.004896525293588638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0767851173877716 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008661558851599693 + ], + "opposite_logprob_tokens": [ + -4.75866174697876 + ], + "logprob_reward": -0.008661558851599693, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008661558851599693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15111401677131653, + 0.035025838762521744, + 0.17942039668560028, + 0.025433728471398354, + 0.031742364168167114, + 0.04354305565357208, + 0.0242290198802948, + 0.022015152499079704, + 0.46578025817871094, + 0.021696187555789948 + ], + "credit_map": [ + 0.015478872694075108, + 0.05093875527381897, + 0.030229052528738976, + 0.009093029424548149, + 0.008870154619216919, + 0.013626046478748322, + 0.012699918821454048, + 0.010178187862038612, + 0.006341095548123121, + 0.004578245338052511, + 0.0048447176814079285, + 0.0030041388235986233, + 0.0034906130749732256, + 0.004761041142046452, + 0.003996261861175299, + 0.0040097287856042385, + 0.01994195580482483, + 0.017734447494149208, + 0.014019826427102089, + 0.012263677082955837, + 0.040374867618083954, + 0.027190404012799263, + 0.029186854138970375, + 0.0187083687633276, + 0.0018584819044917822, + 0.0018378105014562607, + 0.000966873427387327, + 0.001316600013524294, + 0.004542685579508543, + 0.006973244249820709, + 0.001759040867909789, + 0.006178991869091988, + 0.0034662706311792135, + 0.0031009891536086798, + 0.0026463628746569157, + 0.0032259784638881683, + 0.006173757836222649, + 0.0041795228607952595, + 0.0047971392050385475, + 0.004152345005422831, + 0.008075340650975704, + 0.004280131310224533, + 0.012367773801088333, + 0.004372192546725273, + 0.0028540065977722406, + 0.00311541766859591, + 0.004278522450476885, + 0.0041996692307293415, + 0.004469096660614014, + 0.0023671232629567385, + 0.0035932937171310186, + 0.0023371304851025343, + 0.002679188270121813, + 0.0020863276440650225, + 0.0024838470853865147, + 0.004213014151901007, + 0.004475543741136789, + 0.0034508982207626104, + 0.0021848364267498255, + 0.001851778943091631, + 0.0024346678983420134, + 0.003295845352113247, + 0.0017848527058959007, + 0.0025367303751409054, + 0.09685711562633514, + 0.15648967027664185, + 0.12855012714862823, + 0.01730288378894329, + 0.02081945538520813, + 0.01141133438795805, + 0.015317999757826328, + 0.019031692296266556, + 0.0022559273056685925, + 0.003046745201572776, + 0.0033085704781115055, + 0.002107527107000351, + 0.0033434382639825344, + 0.003070228034630418, + 0.00252095190808177, + 0.0020428027492016554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.008661558851599693 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033706627786159515 + ], + "opposite_logprob_tokens": [ + -3.4087066650390625 + ], + "logprob_reward": -0.033706627786159515, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033706627786159515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14252838492393494, + 0.13032321631908417, + 0.3062581717967987, + 0.03463253751397133, + 0.03194697946310043, + 0.06543348729610443, + 0.05807908624410629, + 0.07719756662845612, + 0.09139382839202881, + 0.06220676749944687 + ], + "credit_map": [ + 0.013366748578846455, + 0.03594528138637543, + 0.0233499426394701, + 0.012534585781395435, + 0.00939737819135189, + 0.01877649687230587, + 0.01158105581998825, + 0.017576903104782104, + 0.014371175318956375, + 0.01678840070962906, + 0.016968639567494392, + 0.01628880947828293, + 0.019593404605984688, + 0.012197819538414478, + 0.016345189884305, + 0.01776977814733982, + 0.039812348783016205, + 0.02821701020002365, + 0.022746285423636436, + 0.020592086017131805, + 0.06638152152299881, + 0.04577145725488663, + 0.05197500064969063, + 0.030762454494833946, + 0.0031748851761221886, + 0.003918450325727463, + 0.0021673503797501326, + 0.003261502366513014, + 0.007105662953108549, + 0.00602885615080595, + 0.003511338494718075, + 0.005464487709105015, + 0.005292670335620642, + 0.002096153562888503, + 0.0014850797597318888, + 0.0021711736917495728, + 0.00497962674126029, + 0.006785389501601458, + 0.003486540401354432, + 0.005650344770401716, + 0.010147634893655777, + 0.006583931855857372, + 0.015719396993517876, + 0.006533748935908079, + 0.005346828140318394, + 0.005400574300438166, + 0.007479439489543438, + 0.00822193268686533, + 0.0081322081387043, + 0.005930510815232992, + 0.007518533617258072, + 0.005778537131845951, + 0.011524670757353306, + 0.0039045189041644335, + 0.008252881467342377, + 0.00703722657635808, + 0.010880563408136368, + 0.01025495957583189, + 0.007576082833111286, + 0.010450812056660652, + 0.00896886084228754, + 0.008864575065672398, + 0.008832336403429508, + 0.011369382962584496, + 0.010884563438594341, + 0.024447452276945114, + 0.023966269567608833, + 0.005991149693727493, + 0.007361246272921562, + 0.005773607641458511, + 0.006079790182411671, + 0.006889745127409697, + 0.007703961804509163, + 0.007507137022912502, + 0.00797750148922205, + 0.007411088794469833, + 0.008087639696896076, + 0.009925315156579018, + 0.008144168183207512, + 0.005449953023344278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.033706627786159515 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205545365810394 + ], + "opposite_logprob_tokens": [ + -2.8120553493499756 + ], + "logprob_reward": -0.06205545365810394, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205545365810394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08884521573781967, + 0.07994601130485535, + 0.04751133546233177, + 0.04135210067033768, + 0.09367381781339645, + 0.10520046204328537, + 0.1530884951353073, + 0.07888688892126083, + 0.2715199589729309, + 0.03997570276260376 + ], + "credit_map": [ + 0.009785136207938194, + 0.019495941698551178, + 0.013833718374371529, + 0.007603667676448822, + 0.009713499806821346, + 0.009208881296217442, + 0.011282297782599926, + 0.007922076620161533, + 0.007295165676623583, + 0.005640089977532625, + 0.010040614753961563, + 0.005242558196187019, + 0.015429513528943062, + 0.009833279065787792, + 0.013284705579280853, + 0.013180081732571125, + 0.007282141130417585, + 0.0042033689096570015, + 0.004010637756437063, + 0.004819109104573727, + 0.008297418244183064, + 0.006363909691572189, + 0.00785908941179514, + 0.004675659351050854, + 0.007207821123301983, + 0.0031993170268833637, + 0.0023969169706106186, + 0.004042256623506546, + 0.007561095990240574, + 0.00617808336392045, + 0.005025192629545927, + 0.005741416942328215, + 0.005260856356471777, + 0.009221268817782402, + 0.0076652769930660725, + 0.010374670848250389, + 0.017706524580717087, + 0.01336889062076807, + 0.016934705898165703, + 0.013141625560820103, + 0.01914156787097454, + 0.007744888309389353, + 0.027526570484042168, + 0.008912625722587109, + 0.009146465919911861, + 0.008629979565739632, + 0.011137531138956547, + 0.012960834428668022, + 0.025870904326438904, + 0.015994908288121223, + 0.019260192289948463, + 0.01509535126388073, + 0.027786284685134888, + 0.011181791312992573, + 0.016949445009231567, + 0.020949624478816986, + 0.010882063768804073, + 0.007981456816196442, + 0.0073285093531012535, + 0.007899492047727108, + 0.010036561638116837, + 0.009590517729520798, + 0.01650180108845234, + 0.008666492067277431, + 0.06026637554168701, + 0.07380201667547226, + 0.057209257036447525, + 0.012168454006314278, + 0.02124566026031971, + 0.014189054258167744, + 0.013710329309105873, + 0.018928825855255127, + 0.004868881776928902, + 0.0043641566298902035, + 0.006175747141242027, + 0.0036481497809290886, + 0.005008581094443798, + 0.006263731513172388, + 0.004744683392345905, + 0.004901769105345011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.06205545365810394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a shaved-headed, stern-faced master of Southern Fist martial arts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lighting style characterized by a cold-toned spotlight creating high contrast with deep shadows on one side of the body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background completely blurred into a deep, nearly black blur due to extremely shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the action show the character driving a clenched right fist forward to explode a heavy wooden stake arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wood chips and fibers from the exploding stake clearly visible in the air with sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038984060287476 + ], + "opposite_logprob_tokens": [ + -1.9103899002075195 + ], + "logprob_reward": -0.16038984060287476, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038984060287476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16495323181152344, + 0.06960378587245941, + 0.057226620614528656, + 0.04710474982857704, + 0.05362546071410179, + 0.26007264852523804, + 0.14340007305145264, + 0.12268467247486115, + 0.050916887819767, + 0.030411839485168457 + ], + "credit_map": [ + 0.018708448857069016, + 0.040842656046152115, + 0.026050597429275513, + 0.013367906212806702, + 0.013298357836902142, + 0.018344076350331306, + 0.016826068982481956, + 0.01751513034105301, + 0.009083067066967487, + 0.006672058254480362, + 0.009999359026551247, + 0.006962762214243412, + 0.01100192777812481, + 0.0074303122237324715, + 0.008051209151744843, + 0.01040309015661478, + 0.010181700810790062, + 0.004998891614377499, + 0.0046335309743881226, + 0.004332963842898607, + 0.009078425355255604, + 0.00797260832041502, + 0.010259550996124744, + 0.005768949631601572, + 0.006894123740494251, + 0.004230185877531767, + 0.002950438763946295, + 0.004137362353503704, + 0.009377985261380672, + 0.008000317960977554, + 0.004778766073286533, + 0.006735569331794977, + 0.005064116790890694, + 0.005012071691453457, + 0.0041480595245957375, + 0.00557229109108448, + 0.009726925753057003, + 0.007617270108312368, + 0.009893273934721947, + 0.006591449491679668, + 0.06258239597082138, + 0.01385111827403307, + 0.09108507633209229, + 0.01940573751926422, + 0.014462770894169807, + 0.014595950953662395, + 0.02526329830288887, + 0.018826274201273918, + 0.020425070077180862, + 0.016933485865592957, + 0.018279902637004852, + 0.0167415589094162, + 0.02139688841998577, + 0.014411979354918003, + 0.016255678609013557, + 0.01895551197230816, + 0.016579531133174896, + 0.013912467285990715, + 0.011474939063191414, + 0.01679498516023159, + 0.017985278740525246, + 0.015108124352991581, + 0.015785183757543564, + 0.01504416111856699, + 0.013252293691039085, + 0.013865957967936993, + 0.01066684816032648, + 0.002500745467841625, + 0.0032694630790501833, + 0.0021486852783709764, + 0.0020795674063265324, + 0.00313332537189126, + 0.003600767580792308, + 0.0036150345113128424, + 0.0034326936583966017, + 0.003954779822379351, + 0.0038993177004158497, + 0.004644486121833324, + 0.004124518483877182, + 0.0031402388121932745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.16038984060287476 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_7/index_tensor([4340, 4340])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_7/index_tensor([4340, 4340])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..26d25459548e896f48f5781e576f4a1e36738788 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_7/index_tensor([4340, 4340])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme low-angle shot, almost touching the ground, slowly advances as a bizarre plant grows, capturing every detail as it perseveres through a deep crack in the dusty concrete ruins. The flat stems of this metallic plant, the \"Prickly Pear\" (Prickly Pear), are not fleshy but composed of sharp-edged, gunmetal-colored metal. Their surface gleams with an iridescent, oily sheen under the cool, diffused daylight. As the camera pans forward at an almost imperceptible pace, new segments slowly extend and unfurl from the tips of the previous ones in an unnatural, mechanical fashion. Each growth is accompanied by a tiny tremor, pushing aside the surrounding rubble and dust. At the joints where the segments meet, a faint, warning crimson light pulses rhythmically, tinting the surrounding rubble and dust with an ominous hue: this is its silent warning cry. The entire scene is painted in a cold, charcoal grey and rusty brown tones, with the only bright colour being the pulsating red light. Together, they create a heavy, depressing dystopian sci-fi atmosphere where a bizarre and tenacious life form is quietly rising after the apocalypse.", + "seed": 28843, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_7/index_4340_seed_28843_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04119396535679698, + "vlm_logprob": -0.04119396535679698, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.04119396535679698, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15892216563224792 + ], + "opposite_logprob_tokens": [ + -2.1589221954345703 + ], + "logprob_reward": -0.15892216563224792, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.15892216563224792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2235044538974762, + 0.23042449355125427, + 0.0705079436302185, + 0.044000010937452316, + 0.10219073295593262, + 0.0937163308262825, + 0.08184067159891129, + 0.07975735515356064, + 0.045089755207300186, + 0.02896820195019245 + ], + "credit_map": [ + 0.02416267991065979, + 0.02886784076690674, + 0.0268531646579504, + 0.047174207866191864, + 0.026838047429919243, + 0.018275702372193336, + 0.032083652913570404, + 0.01924915798008442, + 0.012854890897870064, + 0.018595904111862183, + 0.028770683333277702, + 0.048938121646642685, + 0.028402969241142273, + 0.01658955030143261, + 0.024017296731472015, + 0.05225507542490959, + 0.008768252097070217, + 0.007160556968301535, + 0.009112145751714706, + 0.008576144464313984, + 0.011993993073701859, + 0.0070039634592831135, + 0.007624889723956585, + 0.010267996229231358, + 0.005300958175212145, + 0.004968326073139906, + 0.005305489059537649, + 0.0069839321076869965, + 0.006954686716198921, + 0.0037765002343803644, + 0.005406620912253857, + 0.005303498823195696, + 0.010059022344648838, + 0.015604357235133648, + 0.013336123898625374, + 0.016069181263446808, + 0.0126694580540061, + 0.0064949761144816875, + 0.014944657683372498, + 0.013012961484491825, + 0.010767413303256035, + 0.009649028070271015, + 0.012868726626038551, + 0.012972232885658741, + 0.016932226717472076, + 0.007275558076798916, + 0.012509187683463097, + 0.010741954669356346, + 0.011427133344113827, + 0.008131816983222961, + 0.014887561090290546, + 0.011710717342793941, + 0.00972684845328331, + 0.0050930301658809185, + 0.011321994476020336, + 0.009541565552353859, + 0.009066791273653507, + 0.009513757191598415, + 0.012574137188494205, + 0.01403325516730547, + 0.009818040765821934, + 0.0057556345127522945, + 0.009578821249306202, + 0.009416918270289898, + 0.004060639068484306, + 0.003714952850714326, + 0.0049947029910981655, + 0.004856092855334282, + 0.012494864873588085, + 0.004616828169673681, + 0.005857507232576609, + 0.004494167398661375, + 0.002070350106805563, + 0.0031994434539228678, + 0.001601964351721108, + 0.0021722232922911644, + 0.007092481944710016, + 0.0036141052842140198, + 0.005935888271778822, + 0.003281746990978718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.15892216563224792 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020597022026777267 + ], + "opposite_logprob_tokens": [ + -3.895596981048584 + ], + "logprob_reward": -0.020597022026777267, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020597022026777267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21238909661769867, + 0.04822152853012085, + 0.04370245337486267, + 0.04220673069357872, + 0.04953523725271225, + 0.2168155014514923, + 0.07024027407169342, + 0.04111718013882637, + 0.18476398289203644, + 0.09100806713104248 + ], + "credit_map": [ + 0.012712565250694752, + 0.02108275145292282, + 0.026933792978525162, + 0.04929652437567711, + 0.020057983696460724, + 0.01474013365805149, + 0.033725183457136154, + 0.03384013846516609, + 0.005205072462558746, + 0.005817595403641462, + 0.006148932967334986, + 0.0076616988517344, + 0.004602566361427307, + 0.004556665197014809, + 0.005516625475138426, + 0.00871237087994814, + 0.005660956725478172, + 0.0044388435781002045, + 0.005558053031563759, + 0.00485434802249074, + 0.006892332807183266, + 0.0034467454534024, + 0.0053154765628278255, + 0.007535694632679224, + 0.005168763920664787, + 0.004444508347660303, + 0.005574149079620838, + 0.005283697508275509, + 0.006560214329510927, + 0.004515347070991993, + 0.005690376274287701, + 0.004969669505953789, + 0.005164699163287878, + 0.004976351745426655, + 0.005481400061398745, + 0.00793997012078762, + 0.009134300984442234, + 0.004340927582234144, + 0.0059349313378334045, + 0.006562655791640282, + 0.022852135822176933, + 0.024957716464996338, + 0.02843453921377659, + 0.031865108758211136, + 0.02590431645512581, + 0.0215383879840374, + 0.03194812312722206, + 0.02931516058743, + 0.006112958304584026, + 0.005471975542604923, + 0.013246627524495125, + 0.015181812457740307, + 0.0061423820443451405, + 0.005796774290502071, + 0.012799272313714027, + 0.005488471128046513, + 0.004399445839226246, + 0.003967348486185074, + 0.0058405278250575066, + 0.008985553868114948, + 0.0037711297627538443, + 0.0026474674232304096, + 0.005118741653859615, + 0.006386963184922934, + 0.016603514552116394, + 0.01809649169445038, + 0.02450116164982319, + 0.016151290386915207, + 0.03118898905813694, + 0.029648609459400177, + 0.02656046859920025, + 0.02201344072818756, + 0.010367271490395069, + 0.013752471655607224, + 0.00448086392134428, + 0.006566984113305807, + 0.015290956944227219, + 0.012620404362678528, + 0.01701977849006653, + 0.010909324511885643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.020597022026777267 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864903163164854 + ], + "opposite_logprob_tokens": [ + -4.758648872375488 + ], + "logprob_reward": -0.00864903163164854, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864903163164854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19201914966106415, + 0.14907316863536835, + 0.05590103566646576, + 0.023366710171103477, + 0.06932607293128967, + 0.16834543645381927, + 0.07329470664262772, + 0.09282878041267395, + 0.10234183818101883, + 0.07350314408540726 + ], + "credit_map": [ + 0.013697881251573563, + 0.018000101670622826, + 0.026100818067789078, + 0.04679317772388458, + 0.009294339455664158, + 0.013573112897574902, + 0.027409493923187256, + 0.037150219082832336, + 0.018871819600462914, + 0.018031936138868332, + 0.0201911311596632, + 0.02400972507894039, + 0.016618331894278526, + 0.015101958997547626, + 0.012382131069898605, + 0.023866118863224983, + 0.006287348922342062, + 0.006115093361586332, + 0.009522540494799614, + 0.008401505649089813, + 0.007239294238388538, + 0.005552862770855427, + 0.006603660527616739, + 0.006178728770464659, + 0.0031342129223048687, + 0.002382618607953191, + 0.002859905594959855, + 0.004165037535130978, + 0.003253174014389515, + 0.0018108610529452562, + 0.0028131615836173296, + 0.002947739092633128, + 0.0062189470045268536, + 0.009109447710216045, + 0.009365883655846119, + 0.01162286102771759, + 0.010279537178575993, + 0.004530254751443863, + 0.010272189974784851, + 0.007926950231194496, + 0.02095530368387699, + 0.019566521048545837, + 0.022691577672958374, + 0.02698868326842785, + 0.016812825575470924, + 0.017178744077682495, + 0.021900858730077744, + 0.022250913083553314, + 0.007906388491392136, + 0.007378756068646908, + 0.010796913877129555, + 0.013429132290184498, + 0.01104342844337225, + 0.0049240910448133945, + 0.011312498711049557, + 0.0065034981817007065, + 0.008672128431499004, + 0.008351464755833149, + 0.01701604761183262, + 0.009415158070623875, + 0.009372498840093613, + 0.008198308758437634, + 0.01063449215143919, + 0.021168682724237442, + 0.011054685339331627, + 0.010590468533337116, + 0.012000344693660736, + 0.009870180860161781, + 0.017041603103280067, + 0.01359946932643652, + 0.017439665272831917, + 0.010745411738753319, + 0.010508467443287373, + 0.010561150498688221, + 0.006182175595313311, + 0.005235620774328709, + 0.010425377637147903, + 0.012318180873990059, + 0.010320733301341534, + 0.007951432839035988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.00864903163164854 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734176073223352 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734176073223352, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734176073223352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11298484355211258, + 0.13310469686985016, + 0.06400445848703384, + 0.05731201171875, + 0.0360586903989315, + 0.14031435549259186, + 0.1560545563697815, + 0.05657421424984932, + 0.13843417167663574, + 0.10515792667865753 + ], + "credit_map": [ + 0.008562993258237839, + 0.016407111659646034, + 0.011882969178259373, + 0.021950671449303627, + 0.01529608853161335, + 0.008897113613784313, + 0.016035238280892372, + 0.013952679000794888, + 0.016449714079499245, + 0.014236440882086754, + 0.01766471564769745, + 0.023179618641734123, + 0.011545510031282902, + 0.01404221635311842, + 0.010145481675863266, + 0.025841018185019493, + 0.0070039574056863785, + 0.006514598615467548, + 0.009899276308715343, + 0.008592093363404274, + 0.008294515311717987, + 0.005661736708134413, + 0.00966339185833931, + 0.008374897763133049, + 0.005718889646232128, + 0.003938508685678244, + 0.006886470131576061, + 0.013306002132594585, + 0.0059431022964417934, + 0.005375092849135399, + 0.007545342203229666, + 0.008598613552749157, + 0.004225472453981638, + 0.003648069454357028, + 0.0043100169859826565, + 0.004614993464201689, + 0.004906808026134968, + 0.0027765100821852684, + 0.0061296382918953896, + 0.005447187926620245, + 0.015425330027937889, + 0.012352397665381432, + 0.017816705629229546, + 0.022713135927915573, + 0.01845690608024597, + 0.011528625153005123, + 0.021343894302845, + 0.020677383989095688, + 0.012829531915485859, + 0.012405375018715858, + 0.023996762931346893, + 0.018856193870306015, + 0.034108106046915054, + 0.014516700990498066, + 0.02471492625772953, + 0.014626987278461456, + 0.004886464681476355, + 0.004640344530344009, + 0.0096993837505579, + 0.006144585087895393, + 0.008736314252018929, + 0.004787413403391838, + 0.007298456970602274, + 0.010381262749433517, + 0.012676372192800045, + 0.013978828676044941, + 0.016618652269244194, + 0.02258387580513954, + 0.02485526166856289, + 0.01419353298842907, + 0.019581565633416176, + 0.013946104794740677, + 0.01574353128671646, + 0.01937292516231537, + 0.007698763627558947, + 0.005584165453910828, + 0.014599811285734177, + 0.01677863858640194, + 0.012222517281770706, + 0.013157589361071587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.006734176073223352 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011067431420087814 + ], + "opposite_logprob_tokens": [ + -4.5110673904418945 + ], + "logprob_reward": -0.011067431420087814, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011067431420087814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13181909918785095, + 0.10074387490749359, + 0.06174413114786148, + 0.022870659828186035, + 0.05295853689312935, + 0.11032366007566452, + 0.05540812760591507, + 0.02757212519645691, + 0.22735102474689484, + 0.20920872688293457 + ], + "credit_map": [ + 0.008909464813768864, + 0.009181175380945206, + 0.01662258617579937, + 0.03202638775110245, + 0.008728795684874058, + 0.009879142045974731, + 0.01882559061050415, + 0.027645964175462723, + 0.008977116085588932, + 0.01108606532216072, + 0.012792379595339298, + 0.01770714670419693, + 0.01391628384590149, + 0.007315961644053459, + 0.01102830097079277, + 0.01792062073945999, + 0.005401477217674255, + 0.0059048207476735115, + 0.008951377123594284, + 0.010544060729444027, + 0.007146967574954033, + 0.005706239957362413, + 0.011144320480525494, + 0.006944869179278612, + 0.0022958896588534117, + 0.0024425117298960686, + 0.0032235383987426758, + 0.003665790194645524, + 0.0027271006256341934, + 0.0023155389353632927, + 0.0029789479449391365, + 0.0032213402446359396, + 0.005291243549436331, + 0.008103994652628899, + 0.0062845987267792225, + 0.008050852455198765, + 0.007959457114338875, + 0.0034837566781789064, + 0.007455925922840834, + 0.006328709423542023, + 0.01547719445079565, + 0.014609190635383129, + 0.01302897185087204, + 0.013286645524203777, + 0.017923325300216675, + 0.008987033739686012, + 0.015541240572929382, + 0.011470058932900429, + 0.00487224105745554, + 0.004032122902572155, + 0.008923709392547607, + 0.00848294235765934, + 0.00687275780364871, + 0.0053181215189397335, + 0.011742180213332176, + 0.005164054222404957, + 0.0035816836170852184, + 0.002589993644505739, + 0.00349157745949924, + 0.004354771226644516, + 0.003267013467848301, + 0.0022249952889978886, + 0.003866068320348859, + 0.004196022637188435, + 0.02060520276427269, + 0.018149374052882195, + 0.018764836713671684, + 0.022142862901091576, + 0.057080287486314774, + 0.034990936517715454, + 0.03579894080758095, + 0.019818579778075218, + 0.020507505163550377, + 0.027602022513747215, + 0.014244097284972668, + 0.012336920015513897, + 0.04268378019332886, + 0.031874995678663254, + 0.03599025309085846, + 0.02396915853023529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.011067431420087814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme low-angle shot, almost touching the ground, slowly advances as a bizarre plant grows, capturing every detail as it perseveres through a deep crack in the dusty concrete ruins. The flat stems of this metallic plant, the \"Prickly Pear\" (Prickly Pear), are not fleshy but composed of sharp-edged, gunmetal-colored metal. Their surface gleams with an iridescent, oily sheen under the cool, diffused daylight. As the camera pans forward at an almost imperceptible pace, new segments slowly extend and unfurl from the tips of the previous ones in an unnatural, mechanical fashion. Each growth is accompanied by a tiny tremor, pushing aside the surrounding rubble and dust. At the joints where the segments meet, a faint, warning crimson light pulses rhythmically, tinting the surrounding rubble and dust with an ominous hue: this is its silent warning cry. The entire scene is painted in a cold, charcoal grey and rusty brown tones, with the only bright colour being the pulsating red light. Together, they create a heavy, depressing dystopian sci-fi atmosphere where a bizarre and tenacious life form is quietly rising after the apocalypse.", + "seed": 28844, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_7/index_4340_seed_28844_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07326316069811582, + "vlm_logprob": -0.07326316069811582, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.07326316069811582, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2229136824607849 + ], + "opposite_logprob_tokens": [ + -1.7229137420654297 + ], + "logprob_reward": -0.2229136824607849, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2229136824607849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04125015810132027, + 0.8821372389793396, + 0.009478776715695858, + 0.027805296704173088, + 0.011037218384444714, + 0.006674634292721748, + 0.0032565155997872353, + 0.004903138615190983, + 0.00712353503331542, + 0.006333473604172468 + ], + "credit_map": [ + 0.005679337773472071, + 0.005598760209977627, + 0.0034101021010428667, + 0.006764659192413092, + 0.0072498866356909275, + 0.0042877197265625, + 0.004242769908159971, + 0.004016929306089878, + 0.0515124686062336, + 0.024599697440862656, + 0.04527278617024422, + 0.03962280601263046, + 0.04513221234083176, + 0.3667986989021301, + 0.021712597459554672, + 0.2874860465526581, + 0.0008776283939369023, + 0.002538861008360982, + 0.0007749504293315113, + 0.0008875732310116291, + 0.001145488815382123, + 0.0015471152728423476, + 0.0005738609470427036, + 0.0011333002476021647, + 0.003852189751341939, + 0.003143659560009837, + 0.003178003244102001, + 0.002466778736561537, + 0.0035498703364282846, + 0.005495727993547916, + 0.0033450082410126925, + 0.002774062566459179, + 0.0009101679897867143, + 0.0009792097844183445, + 0.0033981092274188995, + 0.0009699718793854117, + 0.0010090996511280537, + 0.0007762531167827547, + 0.0008590276120230556, + 0.002135380171239376, + 0.0006609185948036611, + 0.0010839770548045635, + 0.0010139934020116925, + 0.0007205608417280018, + 0.0009487915085628629, + 0.0009413975058123469, + 0.0007284117746166885, + 0.0005765847163274884, + 0.00033761252416297793, + 0.0004899369669146836, + 0.00039439089596271515, + 0.00036599652958102524, + 0.00038182339631021023, + 0.00042644215864129364, + 0.0004667430475819856, + 0.00039357063360512257, + 0.0004440282646100968, + 0.0005391029408201575, + 0.0005629335646517575, + 0.0005362958763726056, + 0.0009356411173939705, + 0.0007621971890330315, + 0.0005001791287213564, + 0.0006227613193914294, + 0.0005980664864182472, + 0.0008848835132084787, + 0.0010067395633086562, + 0.001260238466784358, + 0.0009622957441024482, + 0.0009169213590212166, + 0.0007935194298624992, + 0.0007008713437244296, + 0.0006221277872100472, + 0.0006986501393839717, + 0.0004814056446775794, + 0.0006824994343332946, + 0.0009681197698228061, + 0.0007952582091093063, + 0.0010993379401043057, + 0.000986075378023088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.2229136824607849 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811715170741081 + ], + "opposite_logprob_tokens": [ + -3.2881171703338623 + ], + "logprob_reward": -0.03811715170741081, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811715170741081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.00998806394636631, + 0.8444159030914307, + 0.005949990823864937, + 0.007389784324914217, + 0.07076438516378403, + 0.02972700446844101, + 0.003290712833404541, + 0.001519752899184823, + 0.014161976985633373, + 0.012792433612048626 + ], + "credit_map": [ + 0.0010878213215619326, + 0.0010467497631907463, + 0.0008954136865213513, + 0.001038630143739283, + 0.0023346978705376387, + 0.0008832890307530761, + 0.001571544911712408, + 0.0011299167526885867, + 0.049276888370513916, + 0.02358105592429638, + 0.04306842386722565, + 0.03808344155550003, + 0.04329254850745201, + 0.35142916440963745, + 0.020742252469062805, + 0.2749421298503876, + 0.0005484840949065983, + 0.001236519543454051, + 0.0006048697978258133, + 0.0007054061279632151, + 0.0008741292986087501, + 0.0007085767574608326, + 0.0005691749975085258, + 0.0007028303807601333, + 0.0009230095893144608, + 0.0012562392512336373, + 0.0007141914684325457, + 0.0010720615973696113, + 0.0010952773736789823, + 0.0008879920351319015, + 0.0007213205681182444, + 0.0007196925580501556, + 0.006184801924973726, + 0.010248687118291855, + 0.01627298630774021, + 0.007539769634604454, + 0.007761616259813309, + 0.00570253748446703, + 0.006344050168991089, + 0.010709935799241066, + 0.0026494842022657394, + 0.007580408826470375, + 0.0038786474615335464, + 0.003234605072066188, + 0.0040658628568053246, + 0.0023368867114186287, + 0.0032055163756012917, + 0.0027755936607718468, + 0.00039324138197116554, + 0.000475705397548154, + 0.0005136097897775471, + 0.00039832922630012035, + 0.00045252626296132803, + 0.0003388977493159473, + 0.0004534781619440764, + 0.00026492495089769363, + 0.00015942803293000907, + 0.0002791591687127948, + 0.0001567768631502986, + 0.00013091049913782626, + 0.000196291264728643, + 0.00022511277347803116, + 0.0001489656133344397, + 0.00022310877102427185, + 0.0012097671860828996, + 0.0014395120088011026, + 0.001989128766581416, + 0.002358925063163042, + 0.0025965163949877024, + 0.0015430160565301776, + 0.0016635730862617493, + 0.0013615385396406054, + 0.001042742864228785, + 0.0019483076175674796, + 0.001345947035588324, + 0.0021218068432062864, + 0.001995507162064314, + 0.0012040098663419485, + 0.0018194927833974361, + 0.0013146190904080868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.03811715170741081 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647494949400425 + ], + "opposite_logprob_tokens": [ + -4.758647441864014 + ], + "logprob_reward": -0.008647494949400425, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647494949400425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1516789346933365, + 0.5950296521186829, + 0.020496051758527756, + 0.045611992478370667, + 0.0495925210416317, + 0.027648823335766792, + 0.019958902150392532, + 0.01619836688041687, + 0.014640268869698048, + 0.05914446711540222 + ], + "credit_map": [ + 0.02483827993273735, + 0.024652691558003426, + 0.013062913902103901, + 0.02090059220790863, + 0.021099699661135674, + 0.021049391478300095, + 0.016386285424232483, + 0.00968907680362463, + 0.03376507759094238, + 0.016101481392979622, + 0.030611993744969368, + 0.02601640671491623, + 0.03192147612571716, + 0.24809648096561432, + 0.014292648993432522, + 0.19422408938407898, + 0.0017959262477234006, + 0.00293543073348701, + 0.002164857927709818, + 0.002362690167501569, + 0.003251823130995035, + 0.0018224293598905206, + 0.0038405731320381165, + 0.0023223210591822863, + 0.004976415075361729, + 0.009612832218408585, + 0.005363676697015762, + 0.005354511085897684, + 0.005785009823739529, + 0.005724998190999031, + 0.004603943787515163, + 0.004190606530755758, + 0.003817987395450473, + 0.006288859993219376, + 0.011996282264590263, + 0.005455194041132927, + 0.004572824575006962, + 0.00374557520262897, + 0.004655202850699425, + 0.009060593321919441, + 0.0022625746205449104, + 0.0045030564069747925, + 0.002363019622862339, + 0.004683166276663542, + 0.0037621522787958384, + 0.003501314902678132, + 0.003908824175596237, + 0.0026647162158042192, + 0.0020372713916003704, + 0.003538417164236307, + 0.0016349715879186988, + 0.002448244020342827, + 0.003756855381652713, + 0.002111331094056368, + 0.0029649219941347837, + 0.0014668890507891774, + 0.0018675134051591158, + 0.0012456728145480156, + 0.002169750165194273, + 0.001511524198576808, + 0.002452766289934516, + 0.0020276643335819244, + 0.002306172391399741, + 0.002617303980514407, + 0.001237010583281517, + 0.0016616771463304758, + 0.001721740234643221, + 0.0024760793894529343, + 0.002354803029447794, + 0.001977838110178709, + 0.001692866557277739, + 0.001518252887763083, + 0.004586345516145229, + 0.008339607156813145, + 0.0052309283055365086, + 0.009873404167592525, + 0.007797484751790762, + 0.006445505656301975, + 0.009811735711991787, + 0.007059464231133461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.008647494949400425 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076280031353235245 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.0076280031353235245, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076280031353235245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05377737805247307, + 0.8020358681678772, + 0.010108781978487968, + 0.022508127614855766, + 0.035281840711832047, + 0.023619920015335083, + 0.019922085106372833, + 0.0044562541879713535, + 0.004810920916497707, + 0.02347884699702263 + ], + "credit_map": [ + 0.005102286580950022, + 0.009409784339368343, + 0.003179841674864292, + 0.005754838697612286, + 0.013199550099670887, + 0.0071693784557282925, + 0.005510540213435888, + 0.004451157990843058, + 0.04775741323828697, + 0.02300119586288929, + 0.04074490815401077, + 0.03711838647723198, + 0.041195347905159, + 0.3328174650669098, + 0.020207976922392845, + 0.25919318199157715, + 0.0011494988575577736, + 0.0024014427326619625, + 0.0009440078865736723, + 0.0012871192302554846, + 0.0012710168957710266, + 0.0011038135271519423, + 0.0011421905364841223, + 0.0008096923120319843, + 0.002785550430417061, + 0.0035806538071483374, + 0.0028584792744368315, + 0.003424775553867221, + 0.0024814344942569733, + 0.002594448858872056, + 0.0024764363188296556, + 0.0023063505068421364, + 0.002961396472528577, + 0.004599881824105978, + 0.009104126133024693, + 0.003613328328356147, + 0.003918267320841551, + 0.0026851051952689886, + 0.0031268030870705843, + 0.005272929556667805, + 0.002299371175467968, + 0.005452041979879141, + 0.0033117705024778843, + 0.0031966811511665583, + 0.0027630492113530636, + 0.0018655745079740882, + 0.002724015386775136, + 0.002007415983825922, + 0.0018266630358994007, + 0.0032655869144946337, + 0.0025173360481858253, + 0.002154299523681402, + 0.0037064889911562204, + 0.0018083531176671386, + 0.002559482352808118, + 0.0020838745404034853, + 0.00045389027218334377, + 0.0010108196875080466, + 0.00045525451423600316, + 0.0005654646665789187, + 0.0006168252439238131, + 0.0004446953535079956, + 0.00041093051549978554, + 0.0004983735852874815, + 0.00045290947309695184, + 0.0010143633699044585, + 0.000580919673666358, + 0.0005791644216515124, + 0.00083472381811589, + 0.0003779235412366688, + 0.0005418860819190741, + 0.0004290307406336069, + 0.0018053764943033457, + 0.005314414389431477, + 0.001765721826814115, + 0.0031928245443850756, + 0.003332242602482438, + 0.0018595610745251179, + 0.0039021160919219255, + 0.0023065877612680197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0076280031353235245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900947123765945 + ], + "opposite_logprob_tokens": [ + -2.4640095233917236 + ], + "logprob_reward": -0.08900947123765945, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900947123765945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018825361505150795, + 0.7802766561508179, + 0.010550830513238907, + 0.009309958666563034, + 0.009612871333956718, + 0.018899934366345406, + 0.01081231702119112, + 0.007603065576404333, + 0.020454183220863342, + 0.11365489661693573 + ], + "credit_map": [ + 0.002502727322280407, + 0.0024012986104935408, + 0.0019658608362078667, + 0.003387796226888895, + 0.002149083185940981, + 0.002523649949580431, + 0.0020598049741238356, + 0.001835138420574367, + 0.04574872553348541, + 0.021854190155863762, + 0.04005488008260727, + 0.035073745995759964, + 0.040275756269693375, + 0.3244403600692749, + 0.019050683826208115, + 0.25377821922302246, + 0.0013635127106681466, + 0.002083500847220421, + 0.0010511837899684906, + 0.0012967942748218775, + 0.0014191354857757688, + 0.0010976815829053521, + 0.0009297500946559012, + 0.0013092701556161046, + 0.00106942153070122, + 0.0016939804190769792, + 0.000865462701767683, + 0.0010397584410384297, + 0.0013864928623661399, + 0.0011542681604623795, + 0.0010856346925720572, + 0.0010149383451789618, + 0.0011703775962814689, + 0.002074482850730419, + 0.0008309129625558853, + 0.0010907754767686129, + 0.001431098091416061, + 0.0011004811385646462, + 0.0010584534611552954, + 0.0008562887669540942, + 0.0014084870927035809, + 0.003680092515423894, + 0.0018735902849584818, + 0.0018325840355828404, + 0.0023732155095785856, + 0.0012093429686501622, + 0.0038110222667455673, + 0.002711598528549075, + 0.0010062684305012226, + 0.002306914422661066, + 0.001182975247502327, + 0.0016906778328120708, + 0.0012571676634252071, + 0.0010646235896274447, + 0.0014407074777409434, + 0.0008629816002212465, + 0.000949151290114969, + 0.0015151825500652194, + 0.0010289741912856698, + 0.0009057391434907913, + 0.0009504666086286306, + 0.0008370117866434157, + 0.000651978887617588, + 0.0007645603618584573, + 0.001770712318830192, + 0.004261444788426161, + 0.0022647676523774862, + 0.003051314502954483, + 0.002876190934330225, + 0.0024870673660188913, + 0.001989792101085186, + 0.0017528906464576721, + 0.008731917478144169, + 0.025383811444044113, + 0.008345823734998703, + 0.01653684675693512, + 0.016401382163167, + 0.01045916322618723, + 0.0170198492705822, + 0.010776087641716003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.08900947123765945 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme low-angle shot, almost touching the ground, slowly advances as a bizarre plant grows, capturing every detail as it perseveres through a deep crack in the dusty concrete ruins. The flat stems of this metallic plant, the \"Prickly Pear\" (Prickly Pear), are not fleshy but composed of sharp-edged, gunmetal-colored metal. Their surface gleams with an iridescent, oily sheen under the cool, diffused daylight. As the camera pans forward at an almost imperceptible pace, new segments slowly extend and unfurl from the tips of the previous ones in an unnatural, mechanical fashion. Each growth is accompanied by a tiny tremor, pushing aside the surrounding rubble and dust. At the joints where the segments meet, a faint, warning crimson light pulses rhythmically, tinting the surrounding rubble and dust with an ominous hue: this is its silent warning cry. The entire scene is painted in a cold, charcoal grey and rusty brown tones, with the only bright colour being the pulsating red light. Together, they create a heavy, depressing dystopian sci-fi atmosphere where a bizarre and tenacious life form is quietly rising after the apocalypse.", + "seed": 28845, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_7/index_4340_seed_28845_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06768296957015991, + "vlm_logprob": -0.06768296957015991, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.06768296957015991, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18207043409347534 + ], + "opposite_logprob_tokens": [ + -1.9320704936981201 + ], + "logprob_reward": -0.18207043409347534, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.18207043409347534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1669038087129593, + 0.09831216931343079, + 0.14835749566555023, + 0.35200801491737366, + 0.03823680430650711, + 0.07707386463880539, + 0.013319713063538074, + 0.029124002903699875, + 0.037642624229192734, + 0.03902152553200722 + ], + "credit_map": [ + 0.02041802741587162, + 0.014471596106886864, + 0.021546928212046623, + 0.03251627832651138, + 0.014486746862530708, + 0.014172139577567577, + 0.03385404869914055, + 0.015438033267855644, + 0.010985294356942177, + 0.011649733409285545, + 0.013501872308552265, + 0.008944678120315075, + 0.009055403992533684, + 0.012930384837090969, + 0.015441404655575752, + 0.015803387388586998, + 0.01779783144593239, + 0.010318799875676632, + 0.025221463292837143, + 0.029074620455503464, + 0.010547466576099396, + 0.008977874182164669, + 0.026125898584723473, + 0.02029353752732277, + 0.021753408014774323, + 0.014235308393836021, + 0.039298802614212036, + 0.1286144107580185, + 0.01362453680485487, + 0.014496481977403164, + 0.022289501503109932, + 0.09769555926322937, + 0.004237012472003698, + 0.0038566787261515856, + 0.006372947711497545, + 0.0057183741591870785, + 0.0030731018632650375, + 0.0032248017378151417, + 0.006405986379832029, + 0.005347898695617914, + 0.009045432321727276, + 0.01040378026664257, + 0.011922712437808514, + 0.014310574159026146, + 0.00900245551019907, + 0.007322782184928656, + 0.008266475982964039, + 0.0067996433936059475, + 0.0016860658070072532, + 0.0014118639519438148, + 0.0016760921571403742, + 0.001226266147568822, + 0.0012840776471421123, + 0.001593579538166523, + 0.002609758172184229, + 0.0018320092931389809, + 0.003693736158311367, + 0.00351170776411891, + 0.005096597131341696, + 0.0028631321620196104, + 0.003318458329886198, + 0.002613464603200555, + 0.004742725752294064, + 0.003284180536866188, + 0.006197433918714523, + 0.003039414994418621, + 0.0048813531175255775, + 0.007890592329204082, + 0.003521527862176299, + 0.0039804307743906975, + 0.005085894837975502, + 0.003045976161956787, + 0.005374525673687458, + 0.004604768473654985, + 0.005747432354837656, + 0.005121626891195774, + 0.0025912150740623474, + 0.0035095494240522385, + 0.005590115208178759, + 0.006482288707047701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.18207043409347534 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372357040643692 + ], + "opposite_logprob_tokens": [ + -3.4087235927581787 + ], + "logprob_reward": -0.03372357040643692, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372357040643692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08599492162466049, + 0.032963164150714874, + 0.15355198085308075, + 0.2974017262458801, + 0.06105079501867294, + 0.13571903109550476, + 0.04285473749041557, + 0.08247478306293488, + 0.03602160885930061, + 0.07196719944477081 + ], + "credit_map": [ + 0.0083193089812994, + 0.010492921806871891, + 0.010924534872174263, + 0.01207907497882843, + 0.006974065210670233, + 0.011660347692668438, + 0.013188842684030533, + 0.012355822138488293, + 0.003921397961676121, + 0.0043113697320222855, + 0.003309383522719145, + 0.005605210550129414, + 0.004235986620187759, + 0.003692741272971034, + 0.003914342727512121, + 0.003972731065005064, + 0.018689410760998726, + 0.012919174507260323, + 0.029002061113715172, + 0.026652542874217033, + 0.012183256447315216, + 0.013845679350197315, + 0.025151418522000313, + 0.015108433552086353, + 0.01643363945186138, + 0.012616349384188652, + 0.029943406581878662, + 0.07560168951749802, + 0.01624785177409649, + 0.01265745796263218, + 0.04235091805458069, + 0.09155042469501495, + 0.006765188183635473, + 0.005746308248490095, + 0.016424408182501793, + 0.0038974075578153133, + 0.004683979321271181, + 0.004605995491147041, + 0.006714714225381613, + 0.01221279613673687, + 0.021461373195052147, + 0.01160494890064001, + 0.017513861879706383, + 0.01727687567472458, + 0.012838948518037796, + 0.018024859949946404, + 0.01916876994073391, + 0.017829399555921555, + 0.005818687379360199, + 0.0038546782452613115, + 0.008196470327675343, + 0.005245517939329147, + 0.005868878681212664, + 0.004842102061957121, + 0.004677878227084875, + 0.004350524395704269, + 0.009577359072864056, + 0.007563854567706585, + 0.017360571771860123, + 0.010124108754098415, + 0.004801444709300995, + 0.008354329504072666, + 0.012852529995143414, + 0.01184059027582407, + 0.004483272787183523, + 0.0019035243894904852, + 0.010169797576963902, + 0.0037800290156155825, + 0.0022336111869663, + 0.0023821883369237185, + 0.004062609281390905, + 0.007006574887782335, + 0.014191746711730957, + 0.006078535225242376, + 0.00912078469991684, + 0.008434575982391834, + 0.004009018186479807, + 0.006713793613016605, + 0.014036388136446476, + 0.009382359683513641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.03372357040643692 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549126520752907 + ], + "opposite_logprob_tokens": [ + -4.387548923492432 + ], + "logprob_reward": -0.012549126520752907, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549126520752907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13144221901893616, + 0.07037243992090225, + 0.04969460517168045, + 0.25757500529289246, + 0.04672887176275253, + 0.08824527263641357, + 0.11482390761375427, + 0.04710177704691887, + 0.04274106025695801, + 0.15127478539943695 + ], + "credit_map": [ + 0.01735125482082367, + 0.012459224089980125, + 0.02292158640921116, + 0.021438410505652428, + 0.011435436084866524, + 0.013795762322843075, + 0.016165608540177345, + 0.015874939039349556, + 0.00806687492877245, + 0.00582733191549778, + 0.0074567534029483795, + 0.007700024638324976, + 0.007274481002241373, + 0.008297418244183064, + 0.013215895742177963, + 0.012533670291304588, + 0.006730150431394577, + 0.004091569688171148, + 0.00873480923473835, + 0.005353898275643587, + 0.004089026246219873, + 0.005684548523277044, + 0.009263332933187485, + 0.005747274029999971, + 0.01438179425895214, + 0.010359948500990868, + 0.02624005638062954, + 0.07333731651306152, + 0.011892019771039486, + 0.01171258557587862, + 0.03019917942583561, + 0.07945210486650467, + 0.007192209828644991, + 0.00346221961081028, + 0.00985508132725954, + 0.004759326111525297, + 0.003701790003105998, + 0.0062631224282085896, + 0.006844644900411367, + 0.004650478716939688, + 0.010868685320019722, + 0.01106166746467352, + 0.01171496044844389, + 0.01109373103827238, + 0.012009049765765667, + 0.009191769175231457, + 0.011183460243046284, + 0.011121951974928379, + 0.013392679393291473, + 0.010851549915969372, + 0.016159195452928543, + 0.016477124765515327, + 0.013469084165990353, + 0.012085581198334694, + 0.018045969307422638, + 0.014342732727527618, + 0.006715148221701384, + 0.003999126143753529, + 0.006836270913481712, + 0.005415802821516991, + 0.004129141103476286, + 0.00489293597638607, + 0.007728896103799343, + 0.0073844571597874165, + 0.006550063844770193, + 0.003295394591987133, + 0.007501246407628059, + 0.009952256456017494, + 0.0034883986227214336, + 0.00275812647305429, + 0.005822007544338703, + 0.003373571904376149, + 0.02004914917051792, + 0.014963093213737011, + 0.01757092773914337, + 0.01880894973874092, + 0.012881900183856487, + 0.013872809708118439, + 0.02866150252521038, + 0.024466456845402718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.012549126520752907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781880304217339 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781880304217339, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781880304217339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08782481402158737, + 0.04946888983249664, + 0.04709671065211296, + 0.27883899211883545, + 0.029491117224097252, + 0.07650800049304962, + 0.10205039381980896, + 0.10041378438472748, + 0.120659239590168, + 0.10764802247285843 + ], + "credit_map": [ + 0.009412513114511967, + 0.013117699883878231, + 0.015407136641442776, + 0.01493802759796381, + 0.009246762841939926, + 0.007174641825258732, + 0.011625418439507484, + 0.006902608554810286, + 0.004683640785515308, + 0.0038909900467842817, + 0.006254508160054684, + 0.006015817169100046, + 0.005231628660112619, + 0.006764490623027086, + 0.012602259404957294, + 0.004025557078421116, + 0.005086909513920546, + 0.002626541769132018, + 0.010566790588200092, + 0.007466509472578764, + 0.0026958556845784187, + 0.0030950799118727446, + 0.009971391409635544, + 0.005587636027485132, + 0.015192605555057526, + 0.011579868383705616, + 0.027088332921266556, + 0.06176881119608879, + 0.01474025472998619, + 0.011482485570013523, + 0.04988131299614906, + 0.08710532635450363, + 0.0051416982896625996, + 0.002776637440547347, + 0.004258851055055857, + 0.004734030459076166, + 0.0028311503119766712, + 0.0031744835432618856, + 0.004212735686451197, + 0.002361530438065529, + 0.013572119176387787, + 0.006905590184032917, + 0.011142414063215256, + 0.009943068027496338, + 0.007174327503889799, + 0.008371507748961449, + 0.010103905573487282, + 0.009295071475207806, + 0.013695298694074154, + 0.012039112858474255, + 0.013772230595350266, + 0.013555602170526981, + 0.011127527803182602, + 0.012141347862780094, + 0.014675561338663101, + 0.011043716222047806, + 0.011609718203544617, + 0.01137276366353035, + 0.02056034468114376, + 0.009650956839323044, + 0.007206715177744627, + 0.007642412092536688, + 0.01659269630908966, + 0.015778182074427605, + 0.011027492582798004, + 0.008282145485281944, + 0.04251610115170479, + 0.009308950044214725, + 0.0066304984502494335, + 0.0054852706380188465, + 0.008345306850969791, + 0.029063472524285316, + 0.013102225959300995, + 0.011728005483746529, + 0.014284320175647736, + 0.013567475602030754, + 0.008648079819977283, + 0.01025259867310524, + 0.018060429021716118, + 0.0180048905313015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.009781880304217339 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028983652591705 + ], + "opposite_logprob_tokens": [ + -2.350289821624756 + ], + "logprob_reward": -0.10028983652591705, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028983652591705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12245483696460724, + 0.05254598334431648, + 0.048241447657346725, + 0.09932039678096771, + 0.025808660313487053, + 0.0678342878818512, + 0.044934239238500595, + 0.049348585307598114, + 0.07759714871644974, + 0.4119144082069397 + ], + "credit_map": [ + 0.017637014389038086, + 0.018443269655108452, + 0.018339458853006363, + 0.020717142149806023, + 0.012912541627883911, + 0.009702695533633232, + 0.014462586492300034, + 0.010240127332508564, + 0.004322371445596218, + 0.005340959411114454, + 0.007575548253953457, + 0.005464928224682808, + 0.0038818474858999252, + 0.006319922860711813, + 0.011393554508686066, + 0.008246851153671741, + 0.004860267508774996, + 0.00272793578915298, + 0.008852265775203705, + 0.005188800860196352, + 0.0023955204524099827, + 0.004071408417075872, + 0.012493829242885113, + 0.007651418447494507, + 0.00391240231692791, + 0.0023110785987228155, + 0.009120908565819263, + 0.01839718036353588, + 0.0031424821354448795, + 0.003534013405442238, + 0.04046643525362015, + 0.018435897305607796, + 0.0032113133929669857, + 0.0022980128414928913, + 0.0056493403390049934, + 0.003019465832039714, + 0.001697490457445383, + 0.0030830539762973785, + 0.0038764674682170153, + 0.0029735176358371973, + 0.0100727379322052, + 0.0073677245527505875, + 0.008915889076888561, + 0.00937370304018259, + 0.005737286061048508, + 0.007751301396638155, + 0.00997859425842762, + 0.00863705389201641, + 0.005572163965553045, + 0.006032431032508612, + 0.0061831227503716946, + 0.006555675994604826, + 0.007359532173722982, + 0.004169450141489506, + 0.004957960918545723, + 0.004103900399059057, + 0.0065810526721179485, + 0.003976148087531328, + 0.007993916049599648, + 0.0076278382912278175, + 0.003575184615328908, + 0.005304763559252024, + 0.006797309499233961, + 0.007492371369153261, + 0.013014193624258041, + 0.006126380991190672, + 0.009401208721101284, + 0.017301853746175766, + 0.005900532007217407, + 0.0063367136754095554, + 0.011842990294098854, + 0.0076732696034014225, + 0.05308888107538223, + 0.0416518859565258, + 0.05242180451750755, + 0.049274422228336334, + 0.03594421595335007, + 0.037182390689849854, + 0.07037000358104706, + 0.0719807967543602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.10028983652591705 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme low-angle shot, almost touching the ground, slowly advances as a bizarre plant grows, capturing every detail as it perseveres through a deep crack in the dusty concrete ruins. The flat stems of this metallic plant, the \"Prickly Pear\" (Prickly Pear), are not fleshy but composed of sharp-edged, gunmetal-colored metal. Their surface gleams with an iridescent, oily sheen under the cool, diffused daylight. As the camera pans forward at an almost imperceptible pace, new segments slowly extend and unfurl from the tips of the previous ones in an unnatural, mechanical fashion. Each growth is accompanied by a tiny tremor, pushing aside the surrounding rubble and dust. At the joints where the segments meet, a faint, warning crimson light pulses rhythmically, tinting the surrounding rubble and dust with an ominous hue: this is its silent warning cry. The entire scene is painted in a cold, charcoal grey and rusty brown tones, with the only bright colour being the pulsating red light. Together, they create a heavy, depressing dystopian sci-fi atmosphere where a bizarre and tenacious life form is quietly rising after the apocalypse.", + "seed": 28846, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_7/index_4340_seed_28846_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02707055788487196, + "vlm_logprob": -0.02707055788487196, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.02707055788487196, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05846845358610153 + ], + "opposite_logprob_tokens": [ + -3.3084683418273926 + ], + "logprob_reward": -0.05846845358610153, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05846845358610153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12501820921897888, + 0.0759415253996849, + 0.039590559899806976, + 0.07414796948432922, + 0.04354725405573845, + 0.054965540766716, + 0.04787583276629448, + 0.380084365606308, + 0.07561994343996048, + 0.0832088366150856 + ], + "credit_map": [ + 0.017730310559272766, + 0.013175689615309238, + 0.019238144159317017, + 0.00891365297138691, + 0.015559975057840347, + 0.015421830117702484, + 0.017706822603940964, + 0.017271779477596283, + 0.008904129266738892, + 0.005227925721555948, + 0.010832926258444786, + 0.009833083488047123, + 0.006600179709494114, + 0.007906814105808735, + 0.01256039272993803, + 0.014076062478125095, + 0.004975354764610529, + 0.0035050506703555584, + 0.004741574637591839, + 0.0037238963413983583, + 0.00379172177053988, + 0.00525937182828784, + 0.005316300317645073, + 0.008277283981442451, + 0.011159704998135567, + 0.0076850056648254395, + 0.009373968467116356, + 0.005517098121345043, + 0.00812915526330471, + 0.008487178944051266, + 0.012345846742391586, + 0.011450008489191532, + 0.003317889291793108, + 0.006194014102220535, + 0.007488340139389038, + 0.00452074920758605, + 0.0067863184958696365, + 0.004582587629556656, + 0.004979727324098349, + 0.005677621345967054, + 0.004679814912378788, + 0.00989906769245863, + 0.007599625736474991, + 0.004849674645811319, + 0.004046977497637272, + 0.01097304280847311, + 0.003923280630260706, + 0.00899404939264059, + 0.004177490249276161, + 0.006046523340046406, + 0.006693089846521616, + 0.007833058014512062, + 0.004328905139118433, + 0.0073900967836380005, + 0.005455381236970425, + 0.005951284430921078, + 0.039212435483932495, + 0.027795730158686638, + 0.047822341322898865, + 0.05408301576972008, + 0.04158846661448479, + 0.04578477144241333, + 0.03638545423746109, + 0.08741212636232376, + 0.00842892937362194, + 0.008585216477513313, + 0.007195066194981337, + 0.008657200261950493, + 0.009245150722563267, + 0.007320731412619352, + 0.008587825112044811, + 0.017599808052182198, + 0.006927716080099344, + 0.009356293827295303, + 0.009913026355206966, + 0.010818498209118843, + 0.014798173680901527, + 0.013118259608745575, + 0.008752299472689629, + 0.009524565190076828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.05846845358610153 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981533110141754 + ], + "opposite_logprob_tokens": [ + -3.529815435409546 + ], + "logprob_reward": -0.02981533110141754, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981533110141754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08842213451862335, + 0.05074901878833771, + 0.017717240378260612, + 0.02717553824186325, + 0.01930328831076622, + 0.016473595052957535, + 0.05111085623502731, + 0.11372438073158264, + 0.21674947440624237, + 0.3985744118690491 + ], + "credit_map": [ + 0.011158077977597713, + 0.009910623542964458, + 0.012552645057439804, + 0.00891138706356287, + 0.010636166669428349, + 0.009400417096912861, + 0.012940549291670322, + 0.012912278063595295, + 0.005631302949041128, + 0.005550769157707691, + 0.006091948598623276, + 0.0046621509827673435, + 0.006428659427911043, + 0.006127132568508387, + 0.007851320318877697, + 0.008405744098126888, + 0.001972149359062314, + 0.0014253502013161778, + 0.0030342668760567904, + 0.0028144933748990297, + 0.0014465036801993847, + 0.0019479668699204922, + 0.002357650548219681, + 0.0027188609819859266, + 0.003247871994972229, + 0.003200069535523653, + 0.0037025022320449352, + 0.0034740164410322905, + 0.0025782643351703882, + 0.0032352949492633343, + 0.004070310387760401, + 0.003667210228741169, + 0.0017810160061344504, + 0.0038829611148685217, + 0.0020170328207314014, + 0.0020305344369262457, + 0.003585004946216941, + 0.002262767171487212, + 0.0016415416030213237, + 0.0021024304442107677, + 0.0021029538474977016, + 0.0016535052563995123, + 0.0022432166151702404, + 0.0017663928447291255, + 0.001363088726066053, + 0.003575082402676344, + 0.0014198701828718185, + 0.002349487505853176, + 0.005708153359591961, + 0.007130997721105814, + 0.005451455246657133, + 0.0065321666188538074, + 0.006865160074084997, + 0.006904955953359604, + 0.004817709792405367, + 0.007700261659920216, + 0.015731988474726677, + 0.007105682510882616, + 0.01664828322827816, + 0.01889892853796482, + 0.007278247270733118, + 0.014474645256996155, + 0.010660790838301182, + 0.02292582578957081, + 0.024135539308190346, + 0.022182287648320198, + 0.01804221235215664, + 0.028580818325281143, + 0.020575441420078278, + 0.025672895833849907, + 0.02442598156630993, + 0.05313432961702347, + 0.032052185386419296, + 0.05652080103754997, + 0.023815959692001343, + 0.04621614143252373, + 0.09425390511751175, + 0.04813334345817566, + 0.05098303407430649, + 0.04659904167056084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.02981533110141754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646313101053238 + ], + "opposite_logprob_tokens": [ + -4.758646488189697 + ], + "logprob_reward": -0.008646313101053238, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646313101053238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14430691301822662, + 0.08614746481180191, + 0.042081136256456375, + 0.06568323075771332, + 0.03709655627608299, + 0.03638850152492523, + 0.04665860906243324, + 0.302460253238678, + 0.17205694317817688, + 0.06712038069963455 + ], + "credit_map": [ + 0.022701194509863853, + 0.016361407935619354, + 0.01699262484908104, + 0.011284871958196163, + 0.022231338545680046, + 0.021230408921837807, + 0.018075255677103996, + 0.015429810620844364, + 0.011667515151202679, + 0.007875386625528336, + 0.009626982733607292, + 0.009655064903199673, + 0.0062132361344993114, + 0.01320905052125454, + 0.014282379299402237, + 0.013617840595543385, + 0.0043417406268417835, + 0.0033598891459405422, + 0.006292963400483131, + 0.006758310832083225, + 0.0038484542164951563, + 0.004947423469275236, + 0.00594670232385397, + 0.006585652474313974, + 0.00574848847463727, + 0.010363446548581123, + 0.009555946104228497, + 0.005697042215615511, + 0.008859803900122643, + 0.008459789678454399, + 0.008217360824346542, + 0.008781353011727333, + 0.003079144051298499, + 0.005183064378798008, + 0.00641535734757781, + 0.0038765922654420137, + 0.005067726597189903, + 0.004949293565005064, + 0.004112627822905779, + 0.004412748385220766, + 0.0030497550033032894, + 0.003704572794958949, + 0.004407096188515425, + 0.003771018236875534, + 0.0025082651991397142, + 0.008018814958631992, + 0.003386948024854064, + 0.007542029023170471, + 0.0052029965445399284, + 0.0060177999548614025, + 0.005205738823860884, + 0.005983078386634588, + 0.004755606409162283, + 0.008259441703557968, + 0.0046118865720927715, + 0.006622057408094406, + 0.03264567255973816, + 0.02530273236334324, + 0.0419347882270813, + 0.05193362757563591, + 0.02147849090397358, + 0.035715650767087936, + 0.028579464182257652, + 0.06486982852220535, + 0.01639893278479576, + 0.018600264564156532, + 0.01588798314332962, + 0.026018379256129265, + 0.02511180192232132, + 0.01530988235026598, + 0.02486170083284378, + 0.02986799366772175, + 0.005878659430891275, + 0.007212899625301361, + 0.005884813144803047, + 0.008361968211829662, + 0.012912848964333534, + 0.008514422923326492, + 0.007823081687092781, + 0.010531687177717686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.008646313101053238 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863721314817667 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.00863721314817667, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863721314817667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12169349193572998, + 0.08402739465236664, + 0.05109717324376106, + 0.0367317758500576, + 0.025718018412590027, + 0.039698339998722076, + 0.047030940651893616, + 0.27987295389175415, + 0.25072959065437317, + 0.06340035051107407 + ], + "credit_map": [ + 0.01563258282840252, + 0.014489816501736641, + 0.01776975393295288, + 0.01314691174775362, + 0.014416408725082874, + 0.013994899578392506, + 0.015785017982125282, + 0.01645810715854168, + 0.009745890274643898, + 0.009389912709593773, + 0.010275096632540226, + 0.012428323738276958, + 0.009062068536877632, + 0.013239649124443531, + 0.011705809272825718, + 0.008180651813745499, + 0.006681244820356369, + 0.005827692337334156, + 0.00592007115483284, + 0.005581323057413101, + 0.005882406607270241, + 0.005725018214434385, + 0.007351656910032034, + 0.00812776293605566, + 0.004769330844283104, + 0.0029796697199344635, + 0.0063405935652554035, + 0.0029779328033328056, + 0.0034930112306028605, + 0.005547336768358946, + 0.005996315274387598, + 0.004627587739378214, + 0.0025654989294707775, + 0.0029071851167827845, + 0.0038775813300162554, + 0.002839399268850684, + 0.003962354268878698, + 0.004083668813109398, + 0.002635022159665823, + 0.0028473089914768934, + 0.003266711486503482, + 0.008778193034231663, + 0.004936675541102886, + 0.003875972470268607, + 0.0035828466061502695, + 0.005674382671713829, + 0.003811779897660017, + 0.005771777592599392, + 0.004627390298992395, + 0.008005448617041111, + 0.004372358322143555, + 0.005016452632844448, + 0.006671826355159283, + 0.006593451369553804, + 0.004044444765895605, + 0.00769956735894084, + 0.033908095210790634, + 0.019050976261496544, + 0.037791866809129715, + 0.03785012289881706, + 0.02306145615875721, + 0.034341879189014435, + 0.02512100711464882, + 0.06874753534793854, + 0.021781740710139275, + 0.02491110749542713, + 0.02157440036535263, + 0.05988282337784767, + 0.01877661794424057, + 0.027901077643036842, + 0.030245665460824966, + 0.04565616697072983, + 0.0061315009370446205, + 0.011525945737957954, + 0.0064383335411548615, + 0.006606479175388813, + 0.01286828238517046, + 0.007691640872508287, + 0.0057633123360574245, + 0.006374854128807783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.00863721314817667 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low angle, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a metallic plant with flat, sharp-edged stems resembling a 'Prickly Pear'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the plant have a gunmetal color with an iridescent, oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do new segments of the plant extend and unfurl from the tips in a mechanical fashion accompanied by tiny tremors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a faint crimson light pulsing rhythmically at the joints of the plant segments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029785478487610817 + ], + "opposite_logprob_tokens": [ + -3.529785394668579 + ], + "logprob_reward": -0.029785478487610817, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029785478487610817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02884736657142639, + 0.021477529779076576, + 0.019342981278896332, + 0.014230064116418362, + 0.013705057092010975, + 0.008394717238843441, + 0.02983427420258522, + 0.2027432769536972, + 0.5405316352844238, + 0.12089310586452484 + ], + "credit_map": [ + 0.004050210118293762, + 0.0028946578968316317, + 0.004337514750659466, + 0.002410353161394596, + 0.0035332213155925274, + 0.003810419701039791, + 0.003826866624876857, + 0.003984122071415186, + 0.0025781341828405857, + 0.002601388841867447, + 0.0020496640354394913, + 0.0012281470699235797, + 0.004538462497293949, + 0.004300873260945082, + 0.0019881329499185085, + 0.0021927268244326115, + 0.0019131535664200783, + 0.002045466098934412, + 0.002796510700136423, + 0.002578183775767684, + 0.0018980310996994376, + 0.0022318512201309204, + 0.002469591097906232, + 0.0034101936034858227, + 0.0017588702030479908, + 0.0014552666107192636, + 0.002259656321257353, + 0.0016841001342982054, + 0.0017931563779711723, + 0.0018409407930448651, + 0.0018234968883916736, + 0.0016145770205184817, + 0.001502867671661079, + 0.0016079972265288234, + 0.0024574310518801212, + 0.0014944816939532757, + 0.0019085841486230493, + 0.0014665852067992091, + 0.0015465569449588656, + 0.0017205524491146207, + 0.0010647187009453773, + 0.001239463104866445, + 0.0009611517889425159, + 0.0007953635649755597, + 0.0007370932144112885, + 0.0015020790742710233, + 0.0008701126789674163, + 0.001224735053256154, + 0.0026473584584891796, + 0.0027115403208881617, + 0.0037708645686507225, + 0.004673958756029606, + 0.0023683346807956696, + 0.0051155248656868935, + 0.0036481120623648167, + 0.00489857979118824, + 0.01767156459391117, + 0.01592090353369713, + 0.026687493547797203, + 0.041011299937963486, + 0.0166787039488554, + 0.022284401580691338, + 0.02209731936454773, + 0.04039158672094345, + 0.033760230988264084, + 0.06518616527318954, + 0.03784439340233803, + 0.15650787949562073, + 0.05718949809670448, + 0.05273151770234108, + 0.08110768347978592, + 0.05620425567030907, + 0.009543794207274914, + 0.01897428184747696, + 0.007879414595663548, + 0.012958199717104435, + 0.027631673961877823, + 0.01587417721748352, + 0.014908664859831333, + 0.013122904114425182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.029785478487610817 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_9/index_tensor([8807, 8807])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_9/index_tensor([8807, 8807])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9480fc50866ce92b91258657177bb9786c627c75 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_9/index_tensor([8807, 8807])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angle macro shot meticulously captures the moment of suspended action in surreal slow motion. A clear glass bottle, partially filled with dark liquid, rests on its side on a polished black granite countertop dotted with subtle gray mineral patterns. From the bottle's unsealed white plastic flip-top lid, a glob of thick, dark syrup oozes relentlessly. The shimmering, viscous strands, stretched by the deliberate pull of gravity, extend into translucent, deep amber threads that catch the bright, clean kitchen light, revealing their dense, syrupy texture. As it contacts the cool granite surface, a small, perfectly round puddle begins to form, its surface tension causing it to bulge slightly before slowly expanding. The shimmering, intact thread of liquid sugar remains connected to its source, creating a captivating visual bridge between the bottle and the growing puddle. Bright, clinical overhead lighting casts sharp, mirrored highlights across the syrupy smooth surface and reflects off the mirror-like countertop, emphasizing every minute ripple and transforming a simple overflow into a captivating scene of focused, sculptural beauty.", + "seed": 11087, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_9/index_8807_seed_11087_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02911565584363416, + "vlm_logprob": -0.02911565584363416, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.02911565584363416, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02340266853570938 + ], + "opposite_logprob_tokens": [ + -5.898402690887451 + ], + "logprob_reward": -0.02340266853570938, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.02340266853570938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18651427328586578, + 0.0664321705698967, + 0.07827258855104446, + 0.04727836325764656, + 0.09558220207691193, + 0.07226502895355225, + 0.06122148036956787, + 0.26915934681892395, + 0.048796750605106354, + 0.07447782158851624 + ], + "credit_map": [ + 0.03267350792884827, + 0.020911019295454025, + 0.028656180948019028, + 0.02043268457055092, + 0.022312816232442856, + 0.020538590848445892, + 0.026097184047102928, + 0.014892272651195526, + 0.009349562227725983, + 0.008023452945053577, + 0.00783571694046259, + 0.0056150839664042, + 0.0065410262905061245, + 0.006633493583649397, + 0.017305582761764526, + 0.005128250923007727, + 0.013461556285619736, + 0.005901519674807787, + 0.014803610742092133, + 0.008961042389273643, + 0.007987352088093758, + 0.009603165090084076, + 0.008166124112904072, + 0.009388218633830547, + 0.010270529426634312, + 0.006054204888641834, + 0.00811274815350771, + 0.005132170859724283, + 0.003511987393721938, + 0.004646114073693752, + 0.003596322610974312, + 0.005954287014901638, + 0.01569139026105404, + 0.010420606471598148, + 0.010070479474961758, + 0.00573043804615736, + 0.016583936288952827, + 0.017642762511968613, + 0.008758989162743092, + 0.010683596134185791, + 0.015457508154213428, + 0.009507991373538971, + 0.01046249084174633, + 0.005927540827542543, + 0.006956494878977537, + 0.00785780604928732, + 0.005291577894240618, + 0.010803617537021637, + 0.008715352974832058, + 0.008646531961858273, + 0.00912552885711193, + 0.0040335324592888355, + 0.00396143039688468, + 0.007733514532446861, + 0.004672104027122259, + 0.014333481900393963, + 0.053255464881658554, + 0.03070867247879505, + 0.05152617022395134, + 0.026927772909402847, + 0.015038985759019852, + 0.026036370545625687, + 0.024072978645563126, + 0.041592907160520554, + 0.010432666167616844, + 0.004384064581245184, + 0.007126535754650831, + 0.003805972635746002, + 0.0030831010080873966, + 0.005263404455035925, + 0.0029982554260641336, + 0.011702749878168106, + 0.013047689571976662, + 0.013299793936312199, + 0.013450528495013714, + 0.0055881068110466, + 0.005146869458258152, + 0.007232508156448603, + 0.004795128479599953, + 0.011917195282876492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.02340266853570938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01257255394011736 + ], + "opposite_logprob_tokens": [ + -4.387572765350342 + ], + "logprob_reward": -0.01257255394011736, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01257255394011736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21752333641052246, + 0.05367922782897949, + 0.0316002182662487, + 0.06473109126091003, + 0.0747889056801796, + 0.048721324652433395, + 0.12838484346866608, + 0.030869141221046448, + 0.05392720177769661, + 0.2957746088504791 + ], + "credit_map": [ + 0.03234483674168587, + 0.023676183074712753, + 0.04524747654795647, + 0.03161628544330597, + 0.027367589995265007, + 0.017069904133677483, + 0.023454569280147552, + 0.016746491193771362, + 0.010802888311445713, + 0.0054450565949082375, + 0.008593648672103882, + 0.0048697637394070625, + 0.004950169939547777, + 0.005534933414310217, + 0.010044949129223824, + 0.0034378175623714924, + 0.006006635259836912, + 0.0025229223538190126, + 0.004597308114171028, + 0.003931324928998947, + 0.0027845394797623158, + 0.004505096934735775, + 0.003782993648201227, + 0.0034693977795541286, + 0.013218282721936703, + 0.008992493152618408, + 0.010712655261158943, + 0.007426728494465351, + 0.00518018938601017, + 0.005094315391033888, + 0.003050620900467038, + 0.01105580199509859, + 0.011016877368092537, + 0.009197989478707314, + 0.013382267206907272, + 0.0061923908069729805, + 0.008370375260710716, + 0.012554651126265526, + 0.006903867237269878, + 0.007170491386204958, + 0.010163677856326103, + 0.005977333057671785, + 0.007594827562570572, + 0.0030226295348256826, + 0.003931292798370123, + 0.0053572580218315125, + 0.0032852396834641695, + 0.009389061480760574, + 0.02144203707575798, + 0.026513976976275444, + 0.014163519255816936, + 0.009662597440183163, + 0.010622547008097172, + 0.016723548993468285, + 0.0080562187358737, + 0.021200399845838547, + 0.005198270548135042, + 0.003232415532693267, + 0.003934851847589016, + 0.0034008794464170933, + 0.0028250603936612606, + 0.0026968426536768675, + 0.003998689353466034, + 0.005582132376730442, + 0.010389106348156929, + 0.005602711346000433, + 0.007565504405647516, + 0.004862392786890268, + 0.003673822619020939, + 0.005360232666134834, + 0.003458100138232112, + 0.01301533356308937, + 0.05978292599320412, + 0.026920415461063385, + 0.051712796092033386, + 0.023470519110560417, + 0.01842515729367733, + 0.02975427359342575, + 0.016829974949359894, + 0.06887856125831604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.01257255394011736 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980699948966503 + ], + "opposite_logprob_tokens": [ + -3.5298070907592773 + ], + "logprob_reward": -0.02980699948966503, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980699948966503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1315135955810547, + 0.0895484983921051, + 0.06382016092538834, + 0.07142821699380875, + 0.19578321278095245, + 0.07459308952093124, + 0.1630849689245224, + 0.05956772714853287, + 0.042276620864868164, + 0.10838394612073898 + ], + "credit_map": [ + 0.01686529815196991, + 0.014827238395810127, + 0.023409651592373848, + 0.0216208565980196, + 0.0174526609480381, + 0.014079824090003967, + 0.01321836281567812, + 0.010039696469902992, + 0.018008679151535034, + 0.012753942050039768, + 0.010260271839797497, + 0.006467233877629042, + 0.009683262556791306, + 0.008386618457734585, + 0.015597639605402946, + 0.008390850387513638, + 0.011127949692308903, + 0.004102337174117565, + 0.010597152635455132, + 0.006772174034267664, + 0.007395041175186634, + 0.008338747546076775, + 0.007761868182569742, + 0.007724890019744635, + 0.014642691239714622, + 0.005564850755035877, + 0.011358147487044334, + 0.005506839603185654, + 0.003836222691461444, + 0.007151383440941572, + 0.004139911849051714, + 0.019228171557188034, + 0.025227617472410202, + 0.025621699169278145, + 0.030152151361107826, + 0.016070498153567314, + 0.025654371827840805, + 0.03120446391403675, + 0.01979818008840084, + 0.022054223343729973, + 0.013694933615624905, + 0.010581650771200657, + 0.00899052806198597, + 0.006162411067634821, + 0.011679055169224739, + 0.008584119379520416, + 0.00537359993904829, + 0.009526792913675308, + 0.0276030320674181, + 0.027209242805838585, + 0.019407913088798523, + 0.013868951238691807, + 0.013870099559426308, + 0.020297367125749588, + 0.013033699244260788, + 0.027794677764177322, + 0.009009934030473232, + 0.006395329721271992, + 0.008398751728236675, + 0.006050386931747198, + 0.004366703797131777, + 0.006578927394002676, + 0.006098923273384571, + 0.012668771669268608, + 0.008685333654284477, + 0.005113292019814253, + 0.006798777263611555, + 0.002966348547488451, + 0.0025382775347679853, + 0.00468917703256011, + 0.003059790004044771, + 0.008425623178482056, + 0.019443662837147713, + 0.02694127894937992, + 0.015222209505736828, + 0.007939502596855164, + 0.008466590195894241, + 0.00975782424211502, + 0.006166271865367889, + 0.014446603134274483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.02980699948966503 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008141063735820353 + ], + "opposite_logprob_tokens": [ + -7.125813961029053 + ], + "logprob_reward": -0.0008141063735820353, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008141063735820353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060991305857896805, + 0.038221027702093124, + 0.0344405323266983, + 0.04228372499346733, + 0.027157751843333244, + 0.05335318669676781, + 0.1297578364610672, + 0.5396572351455688, + 0.019088158383965492, + 0.055049311369657516 + ], + "credit_map": [ + 0.008140556514263153, + 0.008809409104287624, + 0.009967823512852192, + 0.007588601671159267, + 0.011169842444360256, + 0.004847205244004726, + 0.007017382886260748, + 0.0034504851792007685, + 0.0066052889451384544, + 0.0037614735774695873, + 0.005277356132864952, + 0.0038962604012340307, + 0.003784444183111191, + 0.0044170250184834, + 0.007538701873272657, + 0.0029404759407043457, + 0.005022951867431402, + 0.002299389336258173, + 0.006207347847521305, + 0.004030905663967133, + 0.0034223489928990602, + 0.0049147154204547405, + 0.004846810828894377, + 0.0036960612051188946, + 0.0097767673432827, + 0.00519161019474268, + 0.005192001815885305, + 0.004018514882773161, + 0.0029419760685414076, + 0.0029387096874415874, + 0.0025527672842144966, + 0.009671376086771488, + 0.004470431245863438, + 0.003006490645930171, + 0.003846358973532915, + 0.002540351590141654, + 0.0034814467653632164, + 0.0034761535935103893, + 0.002632850082591176, + 0.0037036670837551355, + 0.009354780428111553, + 0.005058106034994125, + 0.009651804342865944, + 0.0028224969282746315, + 0.007242633029818535, + 0.00582200288772583, + 0.0034504395443946123, + 0.00995092187076807, + 0.019659629091620445, + 0.030724264681339264, + 0.015471355058252811, + 0.012986838817596436, + 0.011144492775201797, + 0.017517073079943657, + 0.008512421511113644, + 0.013741753064095974, + 0.10297537595033646, + 0.05884462967514992, + 0.10044786334037781, + 0.0542641319334507, + 0.03197646886110306, + 0.051058173179626465, + 0.0504295714199543, + 0.08966100960969925, + 0.0036785670090466738, + 0.0018253398593515158, + 0.0027713372837752104, + 0.0017108098836615682, + 0.0009723835974000394, + 0.0018591679399833083, + 0.0017015625489875674, + 0.004568989388644695, + 0.008254696615040302, + 0.01533884834498167, + 0.006728567183017731, + 0.004624682944267988, + 0.004999828990548849, + 0.0056430259719491005, + 0.0028276608791202307, + 0.006631998810917139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0008141063735820353 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898195087909698 + ], + "opposite_logprob_tokens": [ + -2.578981876373291 + ], + "logprob_reward": -0.07898195087909698, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898195087909698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15690606832504272, + 0.0413949228823185, + 0.03905395418405533, + 0.03979739174246788, + 0.07589249312877655, + 0.2234952747821808, + 0.09003569185733795, + 0.15813609957695007, + 0.08711686730384827, + 0.08817118406295776 + ], + "credit_map": [ + 0.025227632373571396, + 0.018667900934815407, + 0.0351463258266449, + 0.025826940312981606, + 0.015235137194395065, + 0.012919752858579159, + 0.015639860183000565, + 0.008242510259151459, + 0.007425462827086449, + 0.003766183042898774, + 0.013461205177009106, + 0.003563079284504056, + 0.0032546967267990112, + 0.00383086153306067, + 0.003171679563820362, + 0.0029217591509222984, + 0.0076080248691141605, + 0.0028717617969959974, + 0.00553089939057827, + 0.004689408931881189, + 0.0038117014337331057, + 0.004637503530830145, + 0.004872578661888838, + 0.0050320751033723354, + 0.008605286478996277, + 0.004569672048091888, + 0.006622150540351868, + 0.00461184699088335, + 0.003258298384025693, + 0.0039583491161465645, + 0.0027534018736332655, + 0.005418389104306698, + 0.011961840093135834, + 0.010677372105419636, + 0.012356845661997795, + 0.006791328079998493, + 0.0079979058355093, + 0.0116514191031456, + 0.007931542582809925, + 0.006524236872792244, + 0.04184740036725998, + 0.02798021398484707, + 0.03369687870144844, + 0.016030004248023033, + 0.02148161083459854, + 0.028426026925444603, + 0.0166513379663229, + 0.03738180547952652, + 0.016788702458143234, + 0.014916719868779182, + 0.010778757743537426, + 0.006303159520030022, + 0.007273783441632986, + 0.010876812972128391, + 0.006104466039687395, + 0.016993289813399315, + 0.029415011405944824, + 0.016384601593017578, + 0.0280611515045166, + 0.015818625688552856, + 0.009196501225233078, + 0.015436225570738316, + 0.014061834663152695, + 0.029762156307697296, + 0.02021075040102005, + 0.011901424266397953, + 0.015207468532025814, + 0.006570674013346434, + 0.005578344222158194, + 0.008393540047109127, + 0.0061615766026079655, + 0.013093092478811741, + 0.016738440841436386, + 0.010158803313970566, + 0.015376156195998192, + 0.0068281651474535465, + 0.007536825258284807, + 0.008817629888653755, + 0.0050341300666332245, + 0.017681028693914413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.07898195087909698 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle macro shot meticulously captures the moment of suspended action in surreal slow motion. A clear glass bottle, partially filled with dark liquid, rests on its side on a polished black granite countertop dotted with subtle gray mineral patterns. From the bottle's unsealed white plastic flip-top lid, a glob of thick, dark syrup oozes relentlessly. The shimmering, viscous strands, stretched by the deliberate pull of gravity, extend into translucent, deep amber threads that catch the bright, clean kitchen light, revealing their dense, syrupy texture. As it contacts the cool granite surface, a small, perfectly round puddle begins to form, its surface tension causing it to bulge slightly before slowly expanding. The shimmering, intact thread of liquid sugar remains connected to its source, creating a captivating visual bridge between the bottle and the growing puddle. Bright, clinical overhead lighting casts sharp, mirrored highlights across the syrupy smooth surface and reflects off the mirror-like countertop, emphasizing every minute ripple and transforming a simple overflow into a captivating scene of focused, sculptural beauty.", + "seed": 11088, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_9/index_8807_seed_11088_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07473337342962623, + "vlm_logprob": -0.07473337342962623, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.07473337342962623, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024687355384230614 + ], + "opposite_logprob_tokens": [ + -5.52468729019165 + ], + "logprob_reward": -0.024687355384230614, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.024687355384230614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19669196009635925, + 0.051945120096206665, + 0.11745712906122208, + 0.11218532174825668, + 0.07920295000076294, + 0.059136420488357544, + 0.09355754405260086, + 0.03721710667014122, + 0.1907702386379242, + 0.06183619424700737 + ], + "credit_map": [ + 0.008341691456735134, + 0.007084525190293789, + 0.02379888854920864, + 0.01891934871673584, + 0.006918244529515505, + 0.05099952965974808, + 0.009869224391877651, + 0.07076052576303482, + 0.004994240123778582, + 0.0034265348222106695, + 0.007339089643210173, + 0.006475953850895166, + 0.003515571355819702, + 0.010349098592996597, + 0.005800389684736729, + 0.010044245049357414, + 0.013936840929090977, + 0.01485616248100996, + 0.01399554405361414, + 0.012621263973414898, + 0.014098835177719593, + 0.01880534365773201, + 0.0134897381067276, + 0.015653396025300026, + 0.008013350889086723, + 0.01056982483714819, + 0.012010338716208935, + 0.012441332451999187, + 0.011650252155959606, + 0.02453434094786644, + 0.011839084327220917, + 0.021126801148056984, + 0.006915195379406214, + 0.013232370838522911, + 0.008561297319829464, + 0.006560944020748138, + 0.006606766022741795, + 0.021427098661661148, + 0.00454399548470974, + 0.011355281807482243, + 0.006422846578061581, + 0.005692712962627411, + 0.008980568498373032, + 0.00526253180578351, + 0.008010799996554852, + 0.011019626632332802, + 0.005863201804459095, + 0.007884133607149124, + 0.010043851099908352, + 0.010743866674602032, + 0.012076737359166145, + 0.00753530440852046, + 0.015044966712594032, + 0.015993893146514893, + 0.00715189054608345, + 0.014967028982937336, + 0.0031966185197234154, + 0.005239368882030249, + 0.0044121937826275826, + 0.003089051228016615, + 0.007167147938162088, + 0.005453199613839388, + 0.0043235220946371555, + 0.004336007405072451, + 0.017506856471300125, + 0.02137855440378189, + 0.02648232690989971, + 0.013987909071147442, + 0.027204595506191254, + 0.03557272627949715, + 0.021972335875034332, + 0.026664935052394867, + 0.0060514952056109905, + 0.008552979677915573, + 0.0076766968704760075, + 0.005092462059110403, + 0.008484193123877048, + 0.011325903236865997, + 0.006864640396088362, + 0.007787822280079126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.024687355384230614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008660495281219482 + ], + "opposite_logprob_tokens": [ + -4.758660316467285 + ], + "logprob_reward": -0.008660495281219482, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008660495281219482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12245499342679977, + 0.06815343350172043, + 0.12834787368774414, + 0.0945989117026329, + 0.19574232399463654, + 0.08177592605352402, + 0.07565486431121826, + 0.047059088945388794, + 0.06506186723709106, + 0.12115071713924408 + ], + "credit_map": [ + 0.010445074178278446, + 0.005977565888315439, + 0.017108779400587082, + 0.013356129638850689, + 0.008930398151278496, + 0.027367230504751205, + 0.012023276649415493, + 0.027246542274951935, + 0.006656398065388203, + 0.0044349064119160175, + 0.009202108718454838, + 0.007302009034901857, + 0.006918775383383036, + 0.014569523744285107, + 0.007339559029787779, + 0.011730154976248741, + 0.012542317621409893, + 0.014407627284526825, + 0.01407199539244175, + 0.015380519442260265, + 0.0200547743588686, + 0.014257087372243404, + 0.022100575268268585, + 0.015532971359789371, + 0.011064406484365463, + 0.008388637565076351, + 0.010170456022024155, + 0.013314113020896912, + 0.01228929404169321, + 0.013989229686558247, + 0.010559727437794209, + 0.01482304371893406, + 0.01680459827184677, + 0.020937351509928703, + 0.02333059161901474, + 0.021769631654024124, + 0.029182013124227524, + 0.034800201654434204, + 0.019295522943139076, + 0.029622407630085945, + 0.00924157164990902, + 0.007882409729063511, + 0.012608763761818409, + 0.005794709548354149, + 0.01090984232723713, + 0.01306796632707119, + 0.009253137744963169, + 0.01301752869039774, + 0.006400591228157282, + 0.011839546263217926, + 0.008471227250993252, + 0.005474179517477751, + 0.012936489656567574, + 0.011441217735409737, + 0.006875995080918074, + 0.012215610593557358, + 0.00595179945230484, + 0.005389038939028978, + 0.008287390694022179, + 0.003438261803239584, + 0.007108340505510569, + 0.006010374519973993, + 0.003999576438218355, + 0.006874303333461285, + 0.006566338241100311, + 0.007450011093169451, + 0.011455144733190536, + 0.005946303252130747, + 0.007583387661725283, + 0.01128790620714426, + 0.006262385752052069, + 0.008510383777320385, + 0.010251111350953579, + 0.013860131613910198, + 0.015653425827622414, + 0.010078541934490204, + 0.02059103362262249, + 0.020936833694577217, + 0.01287164818495512, + 0.016907991841435432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.008660495281219482 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635420672595501 + ], + "opposite_logprob_tokens": [ + -3.6513543128967285 + ], + "logprob_reward": -0.02635420672595501, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635420672595501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17961207032203674, + 0.0975189208984375, + 0.1297287493944168, + 0.16010229289531708, + 0.06386948376893997, + 0.07452382147312164, + 0.15578168630599976, + 0.05976531282067299, + 0.02567426487803459, + 0.053423453122377396 + ], + "credit_map": [ + 0.025456001982092857, + 0.013360826298594475, + 0.029642514884471893, + 0.03069956973195076, + 0.013013611547648907, + 0.023683220148086548, + 0.01721346564590931, + 0.02654286101460457, + 0.008545596152544022, + 0.005730901844799519, + 0.012180818244814873, + 0.01124695036560297, + 0.006972509436309338, + 0.018663126975297928, + 0.011452188715338707, + 0.02272683009505272, + 0.009935080073773861, + 0.016306180506944656, + 0.015013129450380802, + 0.015960700809955597, + 0.023506617173552513, + 0.014862027950584888, + 0.018044164404273033, + 0.016100848093628883, + 0.015602593310177326, + 0.013697821646928787, + 0.016806110739707947, + 0.018534256145358086, + 0.024431759491562843, + 0.032622408121824265, + 0.016284216195344925, + 0.022123120725154877, + 0.004829417914152145, + 0.010411659255623817, + 0.007064784411340952, + 0.0066502755507826805, + 0.00594226224347949, + 0.01164289005100727, + 0.007390172686427832, + 0.009938024915754795, + 0.008106070570647717, + 0.005800177808851004, + 0.012691402807831764, + 0.006105479318648577, + 0.006512907799333334, + 0.009372907690703869, + 0.008078313432633877, + 0.017856566235423088, + 0.011706434190273285, + 0.025629797950387, + 0.01449747197329998, + 0.010937312617897987, + 0.036671385169029236, + 0.019136572256684303, + 0.012350191362202168, + 0.024852532893419266, + 0.005139450542628765, + 0.009101671166718006, + 0.006158101372420788, + 0.004145877901464701, + 0.014316328801214695, + 0.006817568093538284, + 0.006558880675584078, + 0.0075274319387972355, + 0.002404821803793311, + 0.003043993841856718, + 0.0034998345654457808, + 0.0022325627505779266, + 0.004286215640604496, + 0.004019697662442923, + 0.003558613359928131, + 0.0026285266503691673, + 0.0037755914963781834, + 0.006549295969307423, + 0.006785419303923845, + 0.003955427557229996, + 0.009394539520144463, + 0.010626795701682568, + 0.005914831068366766, + 0.0064215511083602905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.02635420672595501 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004967409186065197 + ], + "opposite_logprob_tokens": [ + -7.625496864318848 + ], + "logprob_reward": -0.0004967409186065197, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0004967409186065197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21099285781383514, + 0.06857150793075562, + 0.1483399122953415, + 0.10644044727087021, + 0.05192642658948898, + 0.031509071588516235, + 0.16001099348068237, + 0.04266131669282913, + 0.045753806829452515, + 0.1337936669588089 + ], + "credit_map": [ + 0.014210754074156284, + 0.009445714764297009, + 0.03231978043913841, + 0.02802698314189911, + 0.010189478285610676, + 0.04168698191642761, + 0.020428970456123352, + 0.05468419939279556, + 0.00908378791064024, + 0.004186143632978201, + 0.0084213986992836, + 0.005608781706541777, + 0.00505636353045702, + 0.012999627739191055, + 0.006630102172493935, + 0.016585301607847214, + 0.01599133387207985, + 0.018715988844633102, + 0.02282894402742386, + 0.015692707151174545, + 0.012206868268549442, + 0.018392618745565414, + 0.02293945476412773, + 0.021571990102529526, + 0.009661256335675716, + 0.010720054619014263, + 0.01200934313237667, + 0.012216894887387753, + 0.01349350530654192, + 0.02207295410335064, + 0.009742751717567444, + 0.016523685306310654, + 0.0049870735965669155, + 0.007471086457371712, + 0.006327357143163681, + 0.005308019462972879, + 0.006159819662570953, + 0.008729202672839165, + 0.004230195190757513, + 0.008713672868907452, + 0.0032620886340737343, + 0.00346455373801291, + 0.004095415584743023, + 0.0023087298031896353, + 0.004358906298875809, + 0.005760187283158302, + 0.0029830641578882933, + 0.005276128184050322, + 0.011762554757297039, + 0.030238216742873192, + 0.013142603449523449, + 0.012906362302601337, + 0.04341801255941391, + 0.01673230342566967, + 0.01584673672914505, + 0.015964196994900703, + 0.0033826283179223537, + 0.0046023414470255375, + 0.0055124107748270035, + 0.0040955268777906895, + 0.0067073507234454155, + 0.005737380124628544, + 0.004992657341063023, + 0.007631019223481417, + 0.004274604842066765, + 0.006992768496274948, + 0.005750290583819151, + 0.0023307253140956163, + 0.008589240722358227, + 0.009663311764597893, + 0.0026655560359358788, + 0.005487305577844381, + 0.012502731755375862, + 0.018459158018231392, + 0.016171354800462723, + 0.011540993116796017, + 0.021438349038362503, + 0.027420135214924812, + 0.012109900824725628, + 0.014151048846542835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.0004967409186065197 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134680688381195 + ], + "opposite_logprob_tokens": [ + -1.313468098640442 + ], + "logprob_reward": -0.3134680688381195, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134680688381195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11876913905143738, + 0.09079436957836151, + 0.07267170399427414, + 0.033263493329286575, + 0.09482087939977646, + 0.029470928013324738, + 0.13398650288581848, + 0.12347663193941116, + 0.1353917121887207, + 0.16735464334487915 + ], + "credit_map": [ + 0.010153030045330524, + 0.008873891085386276, + 0.016928289085626602, + 0.015223016031086445, + 0.00940651074051857, + 0.023476116359233856, + 0.012143599800765514, + 0.022564677521586418, + 0.009928896091878414, + 0.006316260434687138, + 0.010030465200543404, + 0.00820014625787735, + 0.008063402026891708, + 0.02334320731461048, + 0.009966747835278511, + 0.01494523137807846, + 0.007219595368951559, + 0.008019796572625637, + 0.010060601867735386, + 0.011209619231522083, + 0.006276736967265606, + 0.010863490402698517, + 0.009516053833067417, + 0.009505807422101498, + 0.0032468989957123995, + 0.003607369726523757, + 0.0029220827855169773, + 0.003340835450217128, + 0.006181060336530209, + 0.006452648900449276, + 0.0032107841689139605, + 0.004301811568439007, + 0.013648162595927715, + 0.008408469147980213, + 0.011377778835594654, + 0.007532746531069279, + 0.008705208078026772, + 0.018435023725032806, + 0.008357727900147438, + 0.018355755135416985, + 0.003978729713708162, + 0.0021819178946316242, + 0.00425839563831687, + 0.002551988000050187, + 0.0028355340473353863, + 0.005058347247540951, + 0.0022639709059149027, + 0.006342040374875069, + 0.01341303437948227, + 0.01595565676689148, + 0.015405756421387196, + 0.00831768848001957, + 0.02290448360145092, + 0.019785262644290924, + 0.010800836607813835, + 0.02740376815199852, + 0.01116663496941328, + 0.015440091490745544, + 0.015583619475364685, + 0.007926320657134056, + 0.01604219153523445, + 0.023226046934723854, + 0.01128025446087122, + 0.02281145565211773, + 0.012542542070150375, + 0.01647966168820858, + 0.018697664141654968, + 0.008168460801243782, + 0.01956276409327984, + 0.026981694623827934, + 0.013831915333867073, + 0.01912701316177845, + 0.02004092000424862, + 0.02102593705058098, + 0.028013646602630615, + 0.010669758543372154, + 0.020490964874625206, + 0.03689827024936676, + 0.011827835813164711, + 0.01838729903101921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.3134680688381195 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle macro shot meticulously captures the moment of suspended action in surreal slow motion. A clear glass bottle, partially filled with dark liquid, rests on its side on a polished black granite countertop dotted with subtle gray mineral patterns. From the bottle's unsealed white plastic flip-top lid, a glob of thick, dark syrup oozes relentlessly. The shimmering, viscous strands, stretched by the deliberate pull of gravity, extend into translucent, deep amber threads that catch the bright, clean kitchen light, revealing their dense, syrupy texture. As it contacts the cool granite surface, a small, perfectly round puddle begins to form, its surface tension causing it to bulge slightly before slowly expanding. The shimmering, intact thread of liquid sugar remains connected to its source, creating a captivating visual bridge between the bottle and the growing puddle. Bright, clinical overhead lighting casts sharp, mirrored highlights across the syrupy smooth surface and reflects off the mirror-like countertop, emphasizing every minute ripple and transforming a simple overflow into a captivating scene of focused, sculptural beauty.", + "seed": 11089, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_9/index_8807_seed_11089_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.015506765502505005, + "vlm_logprob": -0.015506765502505005, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.015506765502505005, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018898429349064827 + ], + "opposite_logprob_tokens": [ + -5.8938984870910645 + ], + "logprob_reward": -0.018898429349064827, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.018898429349064827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2321440875530243, + 0.13709518313407898, + 0.1590532660484314, + 0.03612775728106499, + 0.09987124055624008, + 0.10440901666879654, + 0.057395339012145996, + 0.03868661820888519, + 0.06061159446835518, + 0.07460585981607437 + ], + "credit_map": [ + 0.026392580941319466, + 0.011411932297050953, + 0.02455577254295349, + 0.05176646634936333, + 0.011060100980103016, + 0.011472360230982304, + 0.0373217798769474, + 0.0581631064414978, + 0.021568018943071365, + 0.011428764089941978, + 0.016686348244547844, + 0.049154337495565414, + 0.008001656271517277, + 0.005042698699980974, + 0.008505113422870636, + 0.016708245500922203, + 0.033034395426511765, + 0.016520606353878975, + 0.024748334661126137, + 0.01645229198038578, + 0.01872062310576439, + 0.012961520813405514, + 0.017739741131663322, + 0.01887575164437294, + 0.004241123329848051, + 0.0036439779214560986, + 0.005524564068764448, + 0.0036482245195657015, + 0.005331274121999741, + 0.0032285607885569334, + 0.005500758998095989, + 0.005009273998439312, + 0.010567964054644108, + 0.008840980939567089, + 0.013560999184846878, + 0.0151594877243042, + 0.010593467392027378, + 0.009349400177598, + 0.01966196671128273, + 0.012136981822550297, + 0.012850632891058922, + 0.008196583017706871, + 0.01629340834915638, + 0.01396357361227274, + 0.01111058983951807, + 0.010700871236622334, + 0.01831052079796791, + 0.012982840649783611, + 0.006694397889077663, + 0.004068402573466301, + 0.007638398092240095, + 0.005883791483938694, + 0.010488166473805904, + 0.005976331885904074, + 0.00806516408920288, + 0.00858068373054266, + 0.004249195568263531, + 0.0027170893736183643, + 0.005330221261829138, + 0.003909391351044178, + 0.006004509050399065, + 0.00551050528883934, + 0.004743667319417, + 0.006222040392458439, + 0.0074233198538422585, + 0.004421831574290991, + 0.006820390932261944, + 0.005701567977666855, + 0.007680777460336685, + 0.010137375444173813, + 0.009291132912039757, + 0.009135197848081589, + 0.008712981827557087, + 0.005847237538546324, + 0.012673694640398026, + 0.006631416734308004, + 0.007970345206558704, + 0.009996525943279266, + 0.009742643684148788, + 0.013031012378633022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.018898429349064827 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007659352850168943 + ], + "opposite_logprob_tokens": [ + -4.882659435272217 + ], + "logprob_reward": -0.007659352850168943, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007659352850168943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18510474264621735, + 0.18455970287322998, + 0.06642954796552658, + 0.03908145800232887, + 0.07288814336061478, + 0.07362917810678482, + 0.06799935549497604, + 0.12869609892368317, + 0.06489575654268265, + 0.11671607941389084 + ], + "credit_map": [ + 0.02283339947462082, + 0.016571996733546257, + 0.02446576952934265, + 0.03316927328705788, + 0.025045180693268776, + 0.010256626643240452, + 0.030944906175136566, + 0.021817568689584732, + 0.0174670722335577, + 0.008677261881530285, + 0.017697857692837715, + 0.06802026182413101, + 0.0057965777814388275, + 0.007672064006328583, + 0.01413970347493887, + 0.04508887603878975, + 0.0104758245870471, + 0.005998517852276564, + 0.008382885716855526, + 0.006145318504422903, + 0.008134393952786922, + 0.005064820870757103, + 0.009126575663685799, + 0.013101199641823769, + 0.005707540549337864, + 0.0030594405252486467, + 0.008082881569862366, + 0.004336563870310783, + 0.002796405926346779, + 0.006280416622757912, + 0.005282143130898476, + 0.0035360599867999554, + 0.011539940722286701, + 0.007067785132676363, + 0.011069010011851788, + 0.008508504368364811, + 0.005873826798051596, + 0.01134911086410284, + 0.008907007053494453, + 0.008572947233915329, + 0.008069091476500034, + 0.005842183250933886, + 0.010543911717832088, + 0.009477704763412476, + 0.00942017138004303, + 0.008851062506437302, + 0.010974748060107231, + 0.010450288653373718, + 0.006580868270248175, + 0.003795526921749115, + 0.01210753247141838, + 0.007695466745644808, + 0.0077495151199400425, + 0.006026639603078365, + 0.01377262081950903, + 0.010271178558468819, + 0.01606013998389244, + 0.008043121546506882, + 0.025215761736035347, + 0.015780026093125343, + 0.011380949057638645, + 0.011155632324516773, + 0.017053181305527687, + 0.024007271975278854, + 0.009607984684407711, + 0.004886220674961805, + 0.006297833751887083, + 0.005096693057566881, + 0.012033126316964626, + 0.01118330005556345, + 0.00812109187245369, + 0.007669496349990368, + 0.015927370637655258, + 0.006013995502144098, + 0.013868098147213459, + 0.007361273281276226, + 0.008593674749135971, + 0.025290029123425484, + 0.014255826361477375, + 0.025405799970030785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.007659352850168943 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329564467072487 + ], + "opposite_logprob_tokens": [ + -3.7732956409454346 + ], + "logprob_reward": -0.02329564467072487, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329564467072487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17514468729496002, + 0.11795615404844284, + 0.06629342585802078, + 0.08184157311916351, + 0.09421124309301376, + 0.14909259974956512, + 0.05948827415704727, + 0.11528845876455307, + 0.06233464553952217, + 0.07834889739751816 + ], + "credit_map": [ + 0.016759756952524185, + 0.010222573764622211, + 0.01629018783569336, + 0.036972396075725555, + 0.015670295804739, + 0.009635399095714092, + 0.030380727723240852, + 0.03921336680650711, + 0.01761581189930439, + 0.006911937613040209, + 0.01640322059392929, + 0.037955623120069504, + 0.00956598948687315, + 0.004830448422580957, + 0.009149904362857342, + 0.015523235313594341, + 0.006362380459904671, + 0.004140877164900303, + 0.006017518695443869, + 0.014583604410290718, + 0.007832614704966545, + 0.004402756225317717, + 0.010808100923895836, + 0.012145585380494595, + 0.011605308391153812, + 0.005903779994696379, + 0.013210214674472809, + 0.005716350395232439, + 0.008316143415868282, + 0.006947624031454325, + 0.013188941404223442, + 0.01695321314036846, + 0.013357910327613354, + 0.007285780273377895, + 0.012879252433776855, + 0.009863356128334999, + 0.0078999362885952, + 0.008267948403954506, + 0.018527144566178322, + 0.016129925847053528, + 0.011951117776334286, + 0.01134516205638647, + 0.016850164160132408, + 0.020523102954030037, + 0.021190095692873, + 0.015351230278611183, + 0.032141830772161484, + 0.019739905372262, + 0.005333743989467621, + 0.005610433407127857, + 0.006487447768449783, + 0.008294419385492802, + 0.009778345003724098, + 0.004452204797416925, + 0.012667889706790447, + 0.006863798480480909, + 0.012576509267091751, + 0.007447071373462677, + 0.017977140843868256, + 0.00871187075972557, + 0.02348506450653076, + 0.012650474905967712, + 0.016697734594345093, + 0.015742609277367592, + 0.009792822413146496, + 0.004090054426342249, + 0.006890850607305765, + 0.0051001207903027534, + 0.011363707482814789, + 0.010341986082494259, + 0.007043328136205673, + 0.0077117858454585075, + 0.012745232321321964, + 0.004963288549333811, + 0.010220480151474476, + 0.0059709628112614155, + 0.005585859064012766, + 0.018392406404018402, + 0.009553615003824234, + 0.010917069390416145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.02329564467072487 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013369916705414653 + ], + "opposite_logprob_tokens": [ + -6.626337051391602 + ], + "logprob_reward": -0.0013369916705414653, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013369916705414653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2644670009613037, + 0.08093351125717163, + 0.04023735970258713, + 0.07772043347358704, + 0.13962775468826294, + 0.09445422142744064, + 0.08256969600915909, + 0.1252104640007019, + 0.03150181844830513, + 0.06327775120735168 + ], + "credit_map": [ + 0.025174252688884735, + 0.01402715127915144, + 0.022522568702697754, + 0.06922600418329239, + 0.01592937298119068, + 0.013963037170469761, + 0.027149422094225883, + 0.07647516578435898, + 0.009998143650591373, + 0.01093362458050251, + 0.01696983352303505, + 0.009552223607897758, + 0.008139614015817642, + 0.004513207823038101, + 0.009544427506625652, + 0.011282439343631268, + 0.00661276513710618, + 0.0032049210276454687, + 0.003390248864889145, + 0.004367449786514044, + 0.005308999214321375, + 0.0021836101077497005, + 0.008868050761520863, + 0.006301310379058123, + 0.014979427680373192, + 0.004221499431878328, + 0.009145507588982582, + 0.005757947452366352, + 0.006393043790012598, + 0.017523502930998802, + 0.009376524947583675, + 0.010322974994778633, + 0.013930901885032654, + 0.010551889427006245, + 0.01743166334927082, + 0.020458044484257698, + 0.014300237409770489, + 0.012882492505013943, + 0.030672939494252205, + 0.019399575889110565, + 0.008288655430078506, + 0.008541288785636425, + 0.010129168629646301, + 0.01164800114929676, + 0.01567692868411541, + 0.008819151669740677, + 0.020280953496694565, + 0.011070072650909424, + 0.00695051671937108, + 0.007613129913806915, + 0.010649144649505615, + 0.010184341110289097, + 0.013584612868726254, + 0.007250261027365923, + 0.01850839890539646, + 0.007829295471310616, + 0.01496143452823162, + 0.008112533017992973, + 0.02031613700091839, + 0.01626676134765148, + 0.017776163294911385, + 0.012161350809037685, + 0.017311565577983856, + 0.01830451376736164, + 0.0041463556699454784, + 0.002187321661040187, + 0.004358465317636728, + 0.0025817356072366238, + 0.0052087814547121525, + 0.005267825908958912, + 0.003314581699669361, + 0.004436749964952469, + 0.010419858619570732, + 0.003997062798589468, + 0.007502008695155382, + 0.0036963645834475756, + 0.0068437871523201466, + 0.016189802438020706, + 0.0070286407135427, + 0.0076002259738743305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.0013369916705414653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343408972024918 + ], + "opposite_logprob_tokens": [ + -3.65134334564209 + ], + "logprob_reward": -0.026343408972024918, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343408972024918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.124955914914608, + 0.24167500436306, + 0.09045884758234024, + 0.20368117094039917, + 0.06643936038017273, + 0.029776735231280327, + 0.022305572405457497, + 0.03894830867648125, + 0.05418732762336731, + 0.1275717318058014 + ], + "credit_map": [ + 0.014122487045824528, + 0.006908590905368328, + 0.012028895318508148, + 0.02796575240790844, + 0.008234415203332901, + 0.0071011013351380825, + 0.018286503851413727, + 0.03030816651880741, + 0.02158752642571926, + 0.013284604996442795, + 0.03069751337170601, + 0.08010900765657425, + 0.01448359340429306, + 0.009748413227498531, + 0.019297737628221512, + 0.05246659368276596, + 0.012449812144041061, + 0.008844551630318165, + 0.010418281890451908, + 0.02010318450629711, + 0.006464078556746244, + 0.006700119469314814, + 0.009249174036085606, + 0.01622963696718216, + 0.028246261179447174, + 0.01395045593380928, + 0.055821314454078674, + 0.01759316213428974, + 0.013835166580975056, + 0.024464670568704605, + 0.021917350590229034, + 0.02785278856754303, + 0.009633701294660568, + 0.004714468959718943, + 0.010698569007217884, + 0.006562837399542332, + 0.004744427744299173, + 0.0077840001322329044, + 0.010760005563497543, + 0.01154134701937437, + 0.002972456393763423, + 0.0024539511650800705, + 0.0039018148090690374, + 0.004209593404084444, + 0.0032629678025841713, + 0.0030565077904611826, + 0.0053714667446911335, + 0.0045479764230549335, + 0.002119110431522131, + 0.0016203522682189941, + 0.0027347379364073277, + 0.002685260260477662, + 0.002927931025624275, + 0.0027945730835199356, + 0.00420883996412158, + 0.0032147669699043036, + 0.004298673011362553, + 0.0022089676931500435, + 0.0068702069111168385, + 0.004910110030323267, + 0.004800822585821152, + 0.003316748421639204, + 0.005263800732791424, + 0.007278977893292904, + 0.008761097677052021, + 0.003418527077883482, + 0.009545731358230114, + 0.003722074441611767, + 0.007190548814833164, + 0.004142207093536854, + 0.006406098138540983, + 0.01100104209035635, + 0.018442947417497635, + 0.006162078119814396, + 0.022729896008968353, + 0.006959877908229828, + 0.006024399306625128, + 0.022094817832112312, + 0.015154877677559853, + 0.030002832412719727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 9 + ], + "selected_reward": -0.026343408972024918 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle macro shot meticulously captures the moment of suspended action in surreal slow motion. A clear glass bottle, partially filled with dark liquid, rests on its side on a polished black granite countertop dotted with subtle gray mineral patterns. From the bottle's unsealed white plastic flip-top lid, a glob of thick, dark syrup oozes relentlessly. The shimmering, viscous strands, stretched by the deliberate pull of gravity, extend into translucent, deep amber threads that catch the bright, clean kitchen light, revealing their dense, syrupy texture. As it contacts the cool granite surface, a small, perfectly round puddle begins to form, its surface tension causing it to bulge slightly before slowly expanding. The shimmering, intact thread of liquid sugar remains connected to its source, creating a captivating visual bridge between the bottle and the growing puddle. Bright, clinical overhead lighting casts sharp, mirrored highlights across the syrupy smooth surface and reflects off the mirror-like countertop, emphasizing every minute ripple and transforming a simple overflow into a captivating scene of focused, sculptural beauty.", + "seed": 11090, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000164/group_9/index_8807_seed_11090_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02898523861076683, + "vlm_logprob": -0.02898523861076683, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.02898523861076683, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025784416124224663 + ], + "opposite_logprob_tokens": [ + -6.025784492492676 + ], + "logprob_reward": -0.025784416124224663, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.025784416124224663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13974304497241974, + 0.16219359636306763, + 0.07608062773942947, + 0.05118956044316292, + 0.04600691422820091, + 0.08769601583480835, + 0.1320057064294815, + 0.19118531048297882, + 0.08942774683237076, + 0.024471454322338104 + ], + "credit_map": [ + 0.01302500069141388, + 0.026374924927949905, + 0.030929800122976303, + 0.027195654809474945, + 0.008772367611527443, + 0.012261548079550266, + 0.011712382547557354, + 0.009471355006098747, + 0.01504510547965765, + 0.009499936364591122, + 0.05444351211190224, + 0.020410126075148582, + 0.00499288085848093, + 0.04791473224759102, + 0.0036769260186702013, + 0.006210354156792164, + 0.010348259471356869, + 0.010023093782365322, + 0.01223696582019329, + 0.008582154288887978, + 0.006816314999014139, + 0.008560682646930218, + 0.009873569011688232, + 0.009639590047299862, + 0.00749647943302989, + 0.008942537941038609, + 0.007131706457585096, + 0.007803841028362513, + 0.0051401760429143906, + 0.00463832775130868, + 0.004470868036150932, + 0.005565619561821222, + 0.0044093974865973, + 0.005891456734389067, + 0.007536860182881355, + 0.006287741474807262, + 0.006105406675487757, + 0.0051584928296506405, + 0.005501235369592905, + 0.005116317421197891, + 0.013080048374831676, + 0.007748250383883715, + 0.014193139038980007, + 0.009925457648932934, + 0.008575871586799622, + 0.013715660199522972, + 0.00871919933706522, + 0.011738385073840618, + 0.018042802810668945, + 0.009873255155980587, + 0.03251650556921959, + 0.011374174617230892, + 0.008515001274645329, + 0.023239390924572945, + 0.007584202103316784, + 0.020860372111201286, + 0.036335598677396774, + 0.01768852025270462, + 0.03162817284464836, + 0.021628065034747124, + 0.019070489332079887, + 0.02049112133681774, + 0.02010205201804638, + 0.024241292849183083, + 0.016492201015353203, + 0.009244213812053204, + 0.012774357572197914, + 0.009660479612648487, + 0.012206361629068851, + 0.006203534547239542, + 0.010935058817267418, + 0.01191153284162283, + 0.004087490029633045, + 0.00239211181178689, + 0.003355064895004034, + 0.003791451221331954, + 0.003246620297431946, + 0.0014439552323892713, + 0.0033913790248334408, + 0.002763381227850914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.025784416124224663 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014228442683815956 + ], + "opposite_logprob_tokens": [ + -4.264228343963623 + ], + "logprob_reward": -0.014228442683815956, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014228442683815956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1524866670370102, + 0.24303798377513885, + 0.0408581867814064, + 0.054635610431432724, + 0.04318467155098915, + 0.052379265427589417, + 0.1345147341489792, + 0.17348302900791168, + 0.07979092746973038, + 0.025628870353102684 + ], + "credit_map": [ + 0.01854596845805645, + 0.022546814754605293, + 0.024477621540427208, + 0.02967658080160618, + 0.012864581309258938, + 0.016144689172506332, + 0.013857861049473286, + 0.014372562058269978, + 0.02219562791287899, + 0.012532473541796207, + 0.08115120232105255, + 0.03148297592997551, + 0.008140655234456062, + 0.06942294538021088, + 0.005556041374802589, + 0.012556089088320732, + 0.005483878310769796, + 0.004719083197414875, + 0.009651909582316875, + 0.004455224610865116, + 0.0028828196227550507, + 0.0038560787215828896, + 0.0044114370830357075, + 0.005397760309278965, + 0.006334527395665646, + 0.007182698231190443, + 0.010195723734796047, + 0.005430732388049364, + 0.007849221117794514, + 0.005599096417427063, + 0.005655452609062195, + 0.006388166919350624, + 0.005567506421357393, + 0.005255855154246092, + 0.008376527577638626, + 0.005976473446935415, + 0.0031466768123209476, + 0.006059022154659033, + 0.00390784814953804, + 0.004894765093922615, + 0.008381987921893597, + 0.005297166761010885, + 0.009003746323287487, + 0.0073618278838694096, + 0.00548773305490613, + 0.0059685190208256245, + 0.005459635518491268, + 0.005418657325208187, + 0.013272474519908428, + 0.007971887476742268, + 0.031514205038547516, + 0.01157323457300663, + 0.009615302085876465, + 0.02558215893805027, + 0.008329262025654316, + 0.026656214147806168, + 0.030708611011505127, + 0.015415015630424023, + 0.02936600148677826, + 0.021589666604995728, + 0.021109215915203094, + 0.015158168040215969, + 0.020673926919698715, + 0.01946244388818741, + 0.01908048242330551, + 0.008196321316063404, + 0.012450378388166428, + 0.009047319181263447, + 0.007628901395946741, + 0.005776425823569298, + 0.007119819987565279, + 0.010491285473108292, + 0.003312724642455578, + 0.002370490226894617, + 0.005226636305451393, + 0.0031282673589885235, + 0.0038448164705187082, + 0.0017238297732546926, + 0.0033481919672340155, + 0.0026739162858575583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.014228442683815956 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033709049224853516 + ], + "opposite_logprob_tokens": [ + -3.4087090492248535 + ], + "logprob_reward": -0.033709049224853516, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033709049224853516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10603223741054535, + 0.5126250386238098, + 0.09002458304166794, + 0.06754130125045776, + 0.022740811109542847, + 0.03548889234662056, + 0.055933427065610886, + 0.060609120875597, + 0.0237084049731493, + 0.025296030566096306 + ], + "credit_map": [ + 0.010282875038683414, + 0.013541911728680134, + 0.018663156777620316, + 0.016624387353658676, + 0.010080866515636444, + 0.012496745213866234, + 0.009492657147347927, + 0.0148496488109231, + 0.051599808037281036, + 0.026432769373059273, + 0.1685042679309845, + 0.059718310832977295, + 0.01626528613269329, + 0.15935486555099487, + 0.010664686560630798, + 0.02008507028222084, + 0.012640181928873062, + 0.011490533128380775, + 0.018989356234669685, + 0.009228043258190155, + 0.006421240046620369, + 0.011239609681069851, + 0.008702661842107773, + 0.01131297368556261, + 0.00968667957931757, + 0.00835559330880642, + 0.009998415596783161, + 0.006832808721810579, + 0.011317959055304527, + 0.006926598493009806, + 0.007239537313580513, + 0.00718371756374836, + 0.004140461329370737, + 0.0024289367720484734, + 0.0031595765613019466, + 0.003088358324021101, + 0.002638304838910699, + 0.002564535243436694, + 0.002575658494606614, + 0.0021449816413223743, + 0.0032640849240124226, + 0.002902006031945348, + 0.00458016712218523, + 0.0044572907499969006, + 0.004015705548226833, + 0.006239252630621195, + 0.004167943261563778, + 0.0058624460361897945, + 0.007332479581236839, + 0.004759175702929497, + 0.01309356838464737, + 0.005630286410450935, + 0.0037043816410005093, + 0.010418637655675411, + 0.0031727049499750137, + 0.007822197861969471, + 0.008824883960187435, + 0.00556293036788702, + 0.01069642137736082, + 0.00741615891456604, + 0.009744321927428246, + 0.0048931255005300045, + 0.006112587638199329, + 0.007358699105679989, + 0.005177922546863556, + 0.002867478644475341, + 0.0043650739826262, + 0.002394184935837984, + 0.002229711040854454, + 0.0017736443551257253, + 0.0021617107558995485, + 0.0027386818546801805, + 0.0025504117365926504, + 0.0028840897139161825, + 0.004939504899084568, + 0.003726080758497119, + 0.0029355830047279596, + 0.0019414749694988132, + 0.0031834037508815527, + 0.003135484643280506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.033709049224853516 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011810240102931857 + ], + "opposite_logprob_tokens": [ + -6.751181125640869 + ], + "logprob_reward": -0.0011810240102931857, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011810240102931857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14504337310791016, + 0.3918360769748688, + 0.0767262801527977, + 0.05810302868485451, + 0.030131809413433075, + 0.08324824273586273, + 0.08725626021623611, + 0.08230394124984741, + 0.03132924437522888, + 0.014021704904735088 + ], + "credit_map": [ + 0.020150016993284225, + 0.017727747559547424, + 0.028833236545324326, + 0.023692499846220016, + 0.011871987022459507, + 0.016395024955272675, + 0.011937025934457779, + 0.014435848221182823, + 0.03545873984694481, + 0.025533467531204224, + 0.1341802328824997, + 0.04668080434203148, + 0.01477563101798296, + 0.10670794546604156, + 0.01020787376910448, + 0.01829138770699501, + 0.010344773530960083, + 0.009673850610852242, + 0.016122842207551003, + 0.007884437218308449, + 0.005736569873988628, + 0.009999357163906097, + 0.007532815914601088, + 0.009431639686226845, + 0.006279174238443375, + 0.005759691819548607, + 0.008071070536971092, + 0.007646354380995035, + 0.011105699464678764, + 0.007167538162320852, + 0.005746136419475079, + 0.006327368319034576, + 0.004349295515567064, + 0.0037017976865172386, + 0.004112111870199442, + 0.003740716027095914, + 0.003987621981650591, + 0.004490946885198355, + 0.0033922921866178513, + 0.0023570286575704813, + 0.006946447771042585, + 0.00787795428186655, + 0.010873038321733475, + 0.010241485200822353, + 0.019443044438958168, + 0.008727381937205791, + 0.012170197442173958, + 0.006968702655285597, + 0.010212923400104046, + 0.0053313360549509525, + 0.02294114977121353, + 0.007779437582939863, + 0.007033448200672865, + 0.015291333198547363, + 0.004706136882305145, + 0.013960503041744232, + 0.014198378659784794, + 0.008106311783194542, + 0.012649721466004848, + 0.010888398624956608, + 0.008379272185266018, + 0.00854452420026064, + 0.00930682010948658, + 0.01023052353411913, + 0.006281619891524315, + 0.0031339689157903194, + 0.005035842303186655, + 0.003961862064898014, + 0.0042967889457941055, + 0.0025525593664497137, + 0.002495883498340845, + 0.0035707245115190744, + 0.001661180634982884, + 0.0011572082294151187, + 0.0023860547225922346, + 0.0019135294714942575, + 0.001747007598169148, + 0.0012403256259858608, + 0.0016280246200039983, + 0.0022883755154907703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0011810240102931857 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, slightly high-angle macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a clear glass bottle partially filled with dark liquid resting on its side on a polished black granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a glob of thick dark syrup ooze from the unsealed white plastic flip-top lid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do translucent deep amber syrupy threads extend from the bottle to a small round puddle forming on the granite surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright clinical overhead lighting that casts sharp mirrored highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002326101064682 + ], + "opposite_logprob_tokens": [ + -2.69502329826355 + ], + "logprob_reward": -0.07002326101064682, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002326101064682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14401377737522125, + 0.3702143430709839, + 0.08257210999727249, + 0.07939206808805466, + 0.024092666804790497, + 0.033830754458904266, + 0.06593520939350128, + 0.08163528144359589, + 0.041445352137088776, + 0.07686839252710342 + ], + "credit_map": [ + 0.01836148463189602, + 0.02025376260280609, + 0.027752535417675972, + 0.02067454904317856, + 0.010542118921875954, + 0.01369598787277937, + 0.011214557103812695, + 0.021518778055906296, + 0.03860132768750191, + 0.01870488002896309, + 0.11941304057836533, + 0.04347742721438408, + 0.010258876718580723, + 0.11594577878713608, + 0.008863325230777264, + 0.014949684962630272, + 0.01046386081725359, + 0.00842937733978033, + 0.017010651528835297, + 0.009234849363565445, + 0.004356624558568001, + 0.008427967317402363, + 0.007919540628790855, + 0.016729235649108887, + 0.01276811957359314, + 0.012842737138271332, + 0.015857506543397903, + 0.00831612478941679, + 0.00713029969483614, + 0.007086262106895447, + 0.0062503996305167675, + 0.009140624664723873, + 0.003395468695089221, + 0.0025471362750977278, + 0.0036520524881780148, + 0.0028795094694942236, + 0.0026381744537502527, + 0.0027831620536744595, + 0.0037594600580632687, + 0.0024377035442739725, + 0.0032814175356179476, + 0.004494759254157543, + 0.00440771272405982, + 0.004452371504157782, + 0.005352771375328302, + 0.0035876792389899492, + 0.00532760052010417, + 0.0029264420736581087, + 0.010416604578495026, + 0.005858590826392174, + 0.014914564788341522, + 0.007072730455547571, + 0.0060669491067528725, + 0.00913684070110321, + 0.005621552933007479, + 0.006847378332167864, + 0.012858127243816853, + 0.008894659578800201, + 0.013464421965181828, + 0.008425911888480186, + 0.006257406901568174, + 0.0116456039249897, + 0.008438612334430218, + 0.011650542728602886, + 0.009188060648739338, + 0.0035935547202825546, + 0.007941504940390587, + 0.004362734500318766, + 0.005744076333940029, + 0.0029800243210047483, + 0.0044174278154969215, + 0.0032179676927626133, + 0.012754622846841812, + 0.0066877733916044235, + 0.01130038034170866, + 0.012970726005733013, + 0.007796372286975384, + 0.006248170975595713, + 0.007059717085212469, + 0.012050633318722248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.07002326101064682 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file